engineering today - · pdf filearshia azam1*, ch. d.v. paradesi rao2, j. amarnath3 finite...

Engineering TodayQuarterly Journal

Pi and pd type fuzzy controllers with reduced rules And membership functions 3-10ARSHIA AZAM1*, CH. D.V. PARADESI RAO2, J. AMARNATH3

Finite Length Effect On Cordic Based Oscillator 11-16N.P.NARKHEDE1, S.S. LIMAYE2

Advananced Single-Phase Boost Converter System For Wide Range Of Load 17-24VariationsG.MURUGAN , J.MURUGANANTHAM

Calculation of static transmission error for nylon 6/6 spurs gear byhouser’s 25-28analytical method. - a case studyPROF V.G. ARAJPURE, PROF . P. M. PADOLE, PROF.D.B.PAWAR

Video Indexing Using Radial Basis Function Network 29- 361S.PRASANNA, 2DR.A.JOTHI AND 3DR.S.PURUSHOTHAMAN

Review Paper On Analysis And Synthesis Of Antenna Array Patterns Using Convex Optimization 37- 42 AMANPREET KAUR, JASPREET KAUR , GURDEEP MOHAL

Time Reduction For Wireless Communicationusing Ofdm 43- 461BALRAJ B, 2SIVAKUMAR D, 3THAMARAI SELVI D

Effect Of Sic Content On Mechanical And Tribological Properties Of Al Alloy Sic Composites 47- 54SANJAY SONI, S. DAS, G. DIXIT

“F-test approach for evaluation of process parameter effect in manufacturing for l25 55 - 64 orthogonal array based experimentation”PROF.V.R. NAIK, PROF.DR.S.D. MADNAIK,

Recent advances in knowledge discovery database and fuzzy logic. 65- 68 M.SURESH BABU, S.SURESH BABU, K.N.SUDEEPTHI, MCA, MISTE

Effect of lpg content on the performance and emission characteristics of a diesel-lpg dual-fuel engine 69-74

G.A.RAO, A.V.S.RAJU, C.V.MOHAN RAO, K.GOVINDA RAJULUUnsteady mhd pulsatile generalized couette flow of a couple stress fluid through a porous medium under 75-84the influence of periodic body acceleration S.V. SUNEETHA, Dr. M. VEERA KRISHNA, Dr. S. VENKATESWARLU andProf. R. SIVA PRASAD

ENGINEERING TODAY QUARTERLY JOURNAL 3 DECEMBER 2010 VOLUME II ISSUE 4 ISSN: 2180-0995

Content Page No.

Control of Three Phase Cascaded Multilevel Inverter Using Various Noval Pulse Width Modulation Techniques 85-100P.PALANIVEL, SUBHRANSU SEKHAR DASH

comparison of cell adhesion characteristics of anodized and hydroxyapatite coated ti6al4v 101-112implant material based on wet ability and in-vitro studiesK.K.SAJU

1, S.VIDYANAND

3, JAYADAS .N.H

1, M.K.JAYARAJ

2,

JACKSON JAMES3

optimization of energy for an air conditioning system using plc based dcv approach 113-120S.P. VENDAN, P.RAVIKUMAR

Rdf - a technology for describing web resources 121-124C.SUBHASHINI 1 , S.PREMALATHA2 , A.NAGARATHINAM3

impact strength and workabilty behaviour Of Glass Fibre Concrete 125-130 CHANDRAMOULI ,DR. SESHADRI SEKHAR. T, DR.SRINIVASA RAO P, DR.P.SRAVANA

Integration of agile software development with the existing organizational practices and thereby 131-136 improving the software process

N.GANESH1, S.THANGASAMY2

Wavelet transforms based artificial intelligent approach to the detection of brain 137-144 injuries and epilepsiesPROF. H. N. SURESH1, SHARANABASAVESHWAR G HIREMATH DR .ARULARASU 2,

Pitch detection algorithm using harmonic pattern matching in Fourier of Fourier transform domain 145-154AKANT K. [1], PANDE R. [2], LIMAYE S.S [3]

Development of mathematical model for group replacement of electronic system using markov chains 155-1621NAVEEN KILARI, 2DR.C. NADHAMUNI REDDY, 3DR.B.BALU NAIK

Comparative Study of Face Representation Methods for Efficient Face Recognition- 163-180Survey PaperT.SYED AKHEEL1,S.A.K JILANI 2, V.VENKAT RAMI REDDY 3

Analysis of phisically informed parametric synthesis of sound effects 181-188M.KAMAL BASHA., DR.R.RANI HEMA MALINI., K.ANUJA BERKATH BANU.,


Content Page No.

PI AND PD TYPE FUZZY CONTROLLERS WITH REDUCED RULES AND MEMBERSHIP FUNCTIONS

ARSHIA AZAM1*, CH. D.V. PARADESI RAO2, J. AMARNATH3

1IT Department, Muffakham Jah College of Engineering & Technology, Hyderabad, A.P., India

2 ECE Department, JNTUHCE, JNTUH, Kukatpally, Hyderabad, A.P., India

3 EEE Department, JNTUHCE, JNTUH, Kukatpally, Hyderabad, A.P., India

[email protected], [email protected], [email protected]

ABSTRACT: Fuzzy logic controllers (FLC’s) have many advantages over conventional control methods involving PI controllers. FLC is not only cheap but also covers a wide range of operating conditions and are more readily customizable in natural language terms. Fuzzy controller utilizes a rule base for describing the relationship between input and output variables. The relationship between the input and output variables is defined by using rules that are dependent on number of inputs and the linguistic variables. To improve the performance of any fuzzy controller it is very much required that the number of inputs or number of linguistic variables or both have to be increased. Due to this, the rules present in the rule base of fuzzy controller are increased thereby slowing down the process. In this paper a novel method utilizing the concept of clustering is proposed for a multi input single output (MISO) which reduces the rules present in the rule base at the same time keeping the performance of the process at the same level. Control of two different processes is considered and simulations are performed to validate the proposed method.

Keywords: Fuzzy logic controller, rule base, reduction of rules, MISO.

1.INTRODUCTIONThe research and industrial applications mainly concentrate their efforts to provide simple and easy control methods to cope with the increased complexity of the process or system to be controlled. The practical systems mostly used are complex, time varying with delays and non-linearities and are not defined properly with unknown plant variations and disturbances. Classical control techniques require linear and accurate models of the system under consideration and therefore come across major difficulties if the system is non-linear or if the accurate model of the system is not known. Therefore, new controllers that are effective even when accurate model is unknown are very much necessitated. Intelligent controllers based on fuzzy logic or neural network find a lot of attention as they do not require linear and accurate mathematical model of the plant to be controlled. Fuzzy logic controllers (FLC) are being used for many applications involving non-linear systems, time varying systems, inverted pendulum, aircraft wind regulation, speed control of motors and systems with external disturbances [1-4]. Fuzzy logic controllers exhibit excellent performance in number of applications that include industrial processes,

flexible arm control, insurance, robotics [5-8]. Fuzzy logic controllers are based on the theory of fuzzy logic employing a mode of approximate reasoning resembling the process of human decision making. Mamdani was the first to introduce this control [9]

that was later pioneered by Zadeh with his work in fuzzy sets [10]. Usually the relationship between the input and output of a process in FLC is expressed by “if-then-rules”. In recent research, clustering techniques is being utilized for extracting fuzzy rules which require user to identify structure of the knowledge or rule base. Clustering is nothing but unsupervised method of classification of patterns or data item or observations into clusters or groups and is helpful in constructing fuzzy rules from data. The clustering algorithm requires the user to specify the initial location of the cluster and the initial locations of the cluster center. The cluster centers are used in such a way that the cluster center represents a set of typical data points covering the range of data behavior. The know examples of clustering algorithms are Fuzzy c-Means (FCM) [11] and Kohonen’s self organizing Maps [12]. Kusaik and Chow in 1987 proposed an efficient algorithm using the idea of clustering which has low computational


time and complexity [13]. In 1996 Runkler and Palm [14] proposed a regular fuzzy c-elliptotype clustering algorithm for the extraction of fuzzy system from measure data. Tari et. al [15] presented clustering method in which the structure is captured through rough constructs such as rough prototypes themselves. In [16] Valls et. al presented a method in which background knowledge was used for developing the cluster models. Pedrycz et. al. proposed a method of fuzzy clustering using view points [17].

It is presented that the view points are available to the clustering problem and are helpful in communicating the real meaning of the problem in which the results of the clustering are directly utilized. In [18] a robust fuzzy local information c-means algorithm was proposed which can detect the clusters of an image overcoming the disadvantage of the Fuzzy c-means and their variants. Chiu [19] presented an effective method for extracting fuzzy rules using a cluster estimation method known as subtractive clustering that is helpful in extracting rules from high dimensional characteristic area. The mountain clustering method utilized the auto generation ability for determining the number and initial location of the cluster center through search method known as mountain function at each grid point. This method is a type of subtractive clustering with improved computational attempts in which the data points are considered as candidates for cluster center rather than grid point. The main issue with the clustering based partition is that the corresponding membership functions in each input variable are not clear mainly during the high input cases. This is against the spirit of fuzzy logic which requires the easy understanding of the system. Also the clustering based partitioning creates large number of membership functions when compared to a grid type partitioning [20]. In this paper grid type partitioning with subtractive clustering is proposed for partitioning and then identical membership functions that are present are eliminated by using the logic of similarity measures. A comparison is presented between the proposed clustering based fuzzy logic controller and the conventional fuzzy logic controller. It is found that with the proposed fuzzy logic controller the number of rules are reduced thereby reducing the computational time with the same level of accuracy and performance.

2. CLUSTER ANALYSIS

Usually we say that there is a structure in nature, in the same way there is a structure in almost all the phenomenon we generally try to understand. This is structure is found using the data obtained by observations such as visual, audio or any other optical signal. The main essence of classification is to find structure in the observed data. This classification is nothing but known as clustering which requires the knowledge on what basis the classification or clustering has to be performed. In fuzzy systems the most widely used clustering is performed by using equivalent relations [21] and by using fuzzy c-means [11].

Clustering refers to finding out the number of subclasses of c clusters in a data universe consisting of n data samples. The usual clustering methods defines optimum partitions through a global criterion function that measures the extent to which candidate partitions optimize a weighted sum of squared errors between data points and cluster centers in feature space. In most of the cases the number c of clusters are known initially and some cases it may be reasonable to expect cluster substructure at more than one value of c. In this situation it is very much required to find out the vale of c which gives the most possible number of clusters in the data for the analysis available. If there exist, cluster tendency assessment technique signal of good substructure of data then it will be easy to find an optimal number of rules. However, it is possible to partition the data into number of subsets even if the input output data has cluster substructure or not. Consider a sample set of n data points A = {a1, a2,…an} in M dimensional space where each data sample, ai is defined by m features in the universe A i.e. ai = {ai1, ai2,…aim}. Now we are required to generalize the data points in each dimension so that they are bounded by a unit hypercube or hyper-spherical cluster. Each data point is considered as a possible cluster center and the measure of potential of data point a i is defined by (1)

∑=

−−=

n

j

jaiaeiP1

2α

(1)

where 2/4 ac=α (2)

is a Euclidean norm or distance, ca is a positive

constant. Hence, it can be stated that the measure of the potential for any data point is function of its distance to all other data points. Also a data point


will have high potential value if it has many neighboring data points and will have low potential if it has less number of neighboring data points. The value of constant ca is effectively radius defining the neighborhood. For a particular data point there will be some data points within this radius having affect on the data point and some outside this radius having very low or no affect. In this way the potential of each and every data point is computed and then the data point with highest potential is

taken as the cluster center. Now let *1a be the

location of the first cluster center and *iP be its

potential value. The potential value of each data point is revised by the (3)


2*1*1

aiaePiPiP

−−−⇐

β (3)

where 2)/2( bc=β (4)

and cb is another positive constant. Now from each data point, a potential is subtracted as a function of its distance from the first cluster center. This effectively means that the data points which nearer to the first cluster center will have highly reduced potential and will not be selected as the next cluster center. With this it can be said that the next cluster center will be away from the first cluster center. The constant cb is effectively the radius defining the neighborhood that will have measurable reductions in potential. The value of cb is chosen more than the value of ca so that the cluster centers are not placed close to each other.

In the present work the value of ab cc 2= is

chosen so that good distance is maintained between two clusters. After revising the potential of all the data points using (3) the data point that has the highest remaining potential is chosen as the next cluster center. This process is continued till the potential of the data points fall below a fraction. In general the potential of each data point is revised by using (5)

2*

* kaiaekPiPiP

−−−⇐

β (5)

where *ka is the location of the kth cluster center

and *kP is its potential value.

3. IMPLEMENTATION OF FUZZY RULESDeduction of fuzzy rules using the proposed method is done by the procedure as explained in this section. At first the available data is divided into clusters with respect to their classes. The subtractive clustering then applied to the input space of each cluster of data independently for identification of each class of data [18]. The clusters created in the data of given group identify regions in the input space that map into the associated class. Hence, it can be stated that the each cluster center can be interpreted as a fuzzy rule for identifying the class. The fuzzy rule can be given as If A1 is Ai1 & A2 is Ai2 & A3 is Ai3 & ……. & Aj is Aij then class is c1 where Aj is the jth input feature and Aij is the membership function in the ith rule associated

with the jth input feature. The membership function Aij is given by (6)

−−=

2*

21exp)(

ij

ijajAjAijA

ρ (6)

where *

ija is the jth element of *ia and

αρ 21=ijWith the application of subtractive clustering to each class of data independently, a set of rules may be obtained for identifying each class. Hence, the set of rules obtained are combined to form the rule base of the classifier. At the time of clustering the output class of the cluster is established by the rule having the highest degree of fulfillment. The degree of fulfillment is given as

2*iaa

ei

−−=

αµ

(7)

The proposed approach of subtractive clustering is utilized in identifying the PI and PD type FLC. It is very much necessary and difficult to do the parameter settings for determining the universe ranges and required to perform many number of simulations until the satisfactory values are obtained.

PI type fuzzy controllers are physically related to classical PI controllers. A classical PI controller is define by (8)

∫+=t

dtteT

teKtu0

)(1

1)()( (8)

where K is the gain of PI controller, T1 is integral constant e(t)=r(t)-y(t) is error signal, r(t) is the desired output, y(t) is the obtained output of the process and u(t) is the output of the controller. Differentiating (8) we have

+= )(1)()(

1

teT

tedtdKtu

dtd

(9)

Equating (9) to 0 for getting local extreme location

we have

)(1)(1

teT

tedtd −= (10)

The discrete form of (9) is given as


+∆=∆ )(1)()(

1

keT

keKku (11)

where ( ) Tkukuku )1()()( −−=∆ , (12)

( ) Tkekeke )1()()( −−=∆ (13)

where T is the sampling period and k is the step. The

equation (11) can be re written as

( ))()(1)( 11

kekeTT

Kku +∆=∆ (14)

It is required to set the universe range for the error and change in error which is achieved by a scale factor M, for the universe ranges such that M > 0. By applying scale factor (14) can be written as

+∆=∆ )(

1)(1

1)( ke

Mke

M

T

T

MKku (15)

In the above equation )(ke and )(ke∆ are the two inputs to which fuzzification is to be applied and the output hence obtained is to be defuzzified. This fuzzification of input and defuzzification of output can be represented in the equation form as

+∆=∆ )(

1)(1

1)( ke

Mke

M

TFDF

T

MKku (16)

where F is for fuzzification and DF is for Defuzzification.Equating (12) and (16) and simplifying we have

)1()(1

)(1

1)( −++∆=

kuke

Mke

M

TFDF

T

TKMku (17)

The above equation in short can be written as

)1()()( −+∆= kukuku (18)

where Δu is the incremental change in controller output determined by the fuzzy rules. The rule base for computing Δu is given in Table 1.

Table 1: Rule base of FLCe

NB NM NS ZE PS PM PBΔeNB NB NB NB NB NM NS ZENM NB NB NB NM NS ZE PSNS NB NB NM NS ZE PS PMZE NB NM NS ZE PS PM PBPS NM NS ZE PS PM PB PBPM NS ZE PS PM PB PB PBPB ZE PS PM PB PB PB PB

The block diagram of the PI type fuzzy controller is shown in Fig. 1

Fig. 1 Block diagram of PI type fuzzy controller

In the Fig.1 the two inputs are e and its change Δe

are defined as

)1()()( −−=∆ kekeke (20)

where e(k) is the error at kth instant and e(k-1) is the error before the kth instant.In Fig. 1 G1, G2 are input gains and G3 is the output gain and are given as

13,1

2,1

1 T

KMTG

M

TG

MG === (21)

The membership functions of error, change in error and output are defined on the common domain [-1, 1] and is shown in Fig. 2. In Fig. 1 if the output of FLC is u rather than Δu and if there is no accumulation of controller output then the figure is a PD type FLC and the gains for this controller are given as

KMGMdT

GM

G === 3,2,1

1 (22)

where Td is derivative constant.The PI type fuzzy controller shown in Fig. 1 utilized 7 membership functions with 49 rules as can be seen from Table 1. The main aim of the proposed clustering method is to reduce the number of rules present in the rule base of the fuzzy controller without affecting the performance of the controller. The number of rules that are to be reduced utilizes the radius of influence or the constant ca given in (2)


which greatly influences the number of rules that will be generated.

Fig. 2. Membership functions of error, change in error and output

If the value of ca is large then the number of rules generated will be less and if the value of ca is lesser then the rules generated will be higher. Therefore, a choice is to be made between the value of ca and the number of rules to be generated. After applying the proposed subtractive clustering 12 rules are extracted and 8 membership functions are formed. Now by utilizing the similarity approach the number of membership functions is reduced to 3 for error and 5 for change in error.

1. APPLICATIONOFPROPOSEDMETHOD FOR PROCESS CONTROL

The proposed method of reducing the rules by subtractive clustering has been applied to two different control processes. The performance of the proposed FLC is compared with the conventional FLC. The main objective is to check whether the proposed FLC with reduced rule is able to provide the same level of performance as that of the conventional one. The two processes that have been considered for processes control are Process – 1

)2)(1(1)(

++=

sssG

Process – 2 6116

5.1)( 23 +++=

ssssG

2. SIMULATION RESULTS AND DISCUSSIONS

The simulations of the different controllers are performed on the two different processes and comparison is made between step responses of each controller. The input is taken as step input with 2.5 as the step value. The parameters considered for comparison of the different controllers are peak overshoot, rise time and settling time. The conventional PI type FLC is denoted by PIFLC, PD type FLC is denoted by PDFLC, reduced rule PI type FLC is denoted by RRPIFLC and the reduced rule PD type FLC is denoted by RRPDFLC. In Fig. 3 the response of the process -1 considered is shown for both PI type FLC and reduced ruled PI type FLC and as can be seen from the waveform the performance of the drive for RRPIFLC nearly remains same with that of conventional PIFLC with 49 rules and also the peak overshoot is reduced from 2.6 to 2.5.

0 2 4 6 8 1 0 1 2 1 4 1 6 1 8 2 00

0 . 5

1

1 . 5

2

2 . 5

3

T i m e

Out

put

P I F L C

R R P I F L C

Fig. 3. Response of process – 1 with PIFLC and RRPIFLC

In Fig. 4 the response of process -1 for conventional PD type FLC and proposed reduced rule FLC are shown. As can be seen from the waveform the response of the proposed RRPDFLC is same as that of the conventional PDFLC with 49 rules. In Fig. 5 and 6 the response of the process -2 considered is shown for PIFLC, RRPIFLC and PDFLC, RRPDFLC respectively. Fig. 5 shows the response of the process -2 with PIFLC and RRPIFLC and from this it is clearly seen


that the response of the drive with RRPIFLC is almost same as that of the conventional PIFLC. In Fig. 6 response of process -2 for PDFLC and RRPDFLC are shown and it can be inferred from the waveforms that the settling time, rise time remains same even though the rules are reduced.

0 2 4 6 8 1 0 1 2 1 4 1 6 1 8 2 00

0 . 5

1

1 . 5

2

2 . 5

3

T i m e

Outp

ut

F P D C

R R F P D C

Fig. 4. Response of process –1with PDFLC and RRPDFLC

In all the waveforms shown it can be stated that the response of the processes considered remains nearly the same even though the rules are reduced thereby taking less computational memory and less computational time.

0 2 0 4 0 6 0 8 0 1 0 0 1 2 00

0 . 5

1

1 . 5

2

2 . 5

3

T i m e

Outp

ut

P I F L C

R R P I F L C

Fig. 5. Response of process – 2 with PIFLC and RRPIFLC

0 2 0 4 0 6 0 8 0 1 0 0 1 2 00

0 . 5

1

1 . 5

2

2 . 5

3

T i m e

Outp

ut

P D F L C

R R P D F L C

Fig. 6. Response of process –2 with PDFLC and RRPDFLC

6.CONCLUSIONS

In this paper a novel method of reducing the rules has been proposed. In conventional PI or PD type fuzzy logic controllers to improve the performance of any process control the number of membership functions and number of rules have to be increased. With increase in number of rules the computational time taken will be increased considerably which is undesirable for any control system. Hence, in this paper a novel method utilizing the concept of clustering is proposed which reduces the 49 rules to 12 rules and the membership functions are reduced from 7 to 3 for error and 5 for change in error. At the same time even though the rules are reduced the performance of the process control nearly remains similar to that of the conventional FLC with 49 rules as can be observed from the waveforms which shows that the settling time, rise time and peak overshoot values nearly remains same for the conventional PI and PD type FLC and the proposed reduced rule PI and PD type FLC. As the number of rules and membership functions are reduced the computational effort taken will be reduced considerably thereby improving the performance of the process control. The above method of clustering can also be applied to any number of input and single output fuzzy logic controllers.


REFERENCES

[1] Wai, R. –J. and Lee, J. –D., “Intelligent Motion Control for Linear Piezoelectric Ceramic Motor Drive”, IEEE Transactions on Systems, Man & Cybernetics – Part B: Cybernetics, Vol. 34, No. 5, pp. 2100-2111, 2004.

[2] Llama, M. A., Kelly, R. and Santibañez, V., “Stable Computed-Torque Control of Robot Manipulators via Fuzzy Self-Tuning”, IEEE Transactions on Systems, Man & Cybernetics – Part B: Cybernetics, Vol. 30, No. 1, pp. 143-150, 2000.

[3] Ordóñez, R. and Passino, K. M., “Indirect Adaptive Control for A Class of Non-linear Systems with a Time-varying Structure”, International Journal of Control, Vol. 74, No. 7, pp. 701-717, 2001.

[4] Chang, Y.C., “Intelligent Robust Control for Uncertain Nonlinear Time-Varying Systems and Its Application to Robotic Systems”, IEEE Transactions on Systems, Man & Cybernetics Part B: Cybernetics, Vol. 35, No. 6, pp. 1108-1119, 2005.

[5] M. Sugeno, Industrial Applications of Fuzzy control, North-Holland, 1985.

[6] Kuldip S. Rattan, B.Chiu, V Feliu and H.B. Brown Jr., “Rule Based Fuzzy control of a single-Link Flexible Manipulator in the presence of Joint friction and Load changes,” American Control Conference, Pittsburgh, PA, 1989.

[7] Shapiro, A. F., Fuzzy Logic in Insurance, Insurance: Mathematics and Economics, Vol.35, No.2 , 2004, 399-424.

[8] Hayward, G. and Davidson, V., Fuzzy Logic Applications, Analyst, Vol.128, 2003, 1304-1306.

[9] E.H. Mamdani, “Application of Fuzzy Algorithms for control of simple dynamic plant, “ Proc. IEE 121 Vol. 12 pp 1585-1588, 1974

[10] Zadeh, L. A., “Fuzzy sets,” Information and Control, Vol. 8, pp. 338-353, 1965

[11] J. Bezdek, Pattern Recognition with Fuzzy Objective Function Algorithm, Plenum Press, New York, 1981.

[12] T. Kohonen, “The self-organizing map,” Proc. of IEEE, vol. 78, no. 9, pp. 1464-1480, September 1990.

[13] A. Kusiak and W. S. Chow, “An efficient cluster identification algorithm,” IEEE Trans. on Systems, Man, and Cybernetics-Part B: Cybernetics, vol. SMC-17, no. 4, pp. 696-699, Jul.-Aug. 1987.

[14] T. A. Runkler and R. H. Palm, “Identification of nonlinear systems using regular fuzzy celliptotype clustering,” Proc. of the Fifth IEEE International Conference on Fuzzy Systems, vol. 2, pp. 1026-1030, 1996.


[15] L. Tari, C.Baral, and S. Kim, “Fuzzy c-means clustering with prior biological knowledge”, J. Biomed. Inf., vol. 42, pp.74-81, 2009.

[16] A. Valls, M. Batet, and E. M. L´opez, “Using expert’s rules as background knowledge in the ClusDM methodology,” Eur. J. Oper. Res., vol. 195, no. 3, pp. 864–875, 2009.

[17] Witold Pedrycz, Vincenzo Loia, and Sabrina Senatore, “Fuzzy Clustering With Viewpoints” IEEE Trans. Fuzzy Syst., vol. 18, no. 2, pp. 274–284, April 2010.

[18] Stelios Krinidis and Vassilios Chatzis, “A Robust Fuzzy Local Information C-Means Clustering Algorithm” IEEE Trans. on Image Processing, vol. 19, No. 5, pp 1328-1337, May 2010.

[19] S. L. Chiu, “An efficient method for extracting fuzzy classification rules from high dimensional data,” J. Advanced Computational Intelligence, vol. 1, no. 1, pp. 31-36, 1997.

[20] S. Chopra, R. Mitra, and V. Kumar, “Identification of rules using subtractive clustering with application to fuzzy controllers,” Proc. of the Third International Conference on Machine Learning and Cybernetics, pp. 4125-4131, 2004.

[21] Zadeh, L. “Similarity relations and fuzzy ordering,” Inf. Sci., vol. 3, pp. 177-200, 1971.


FINITE LENGTH EFFECT ON CORDIC BASED OSCILLATORN.P.NARKHEDE1, S.S. LIMAYE2

1Sr. Lecturer, Department of Electronics & Communication Engineering, Ramdeobaba Engineering College, Gittikhadan, Katol Road, Nagpur-440013. M.S., INDIA.

[email protected], [email protected] 2Professor, Department of Electronics Engineering, Jhulelal Institute of Technology, Lonara,

Nagpur-441111 M.S., INDIA. [email protected], [email protected]

ABSTRACT: This article describes an effect of variation of register length on CORDIC based oscillator. We have implemented a complex oscillator based on the unfolded CORDIC algorithm which produces periodic sine and cosine samples for any specified angle increment. Where the frequency accuracy is achieved by residual angle correction unit and stability is achieved using AGC unit also phase jitter is avoided by rotating the X and Y registers directly. In this paper we are varying the register length and analyzing the effect on phase noise and amplitude stability by turning on and off the residual angle correction unit and gain control unit. The design is synthesized in SPARTAN 3E FPGA and the effect on hardware utilization is presented. Fourier analysis of the output for variation in length using MATLAB is also presented.

Keywords - DDS, CORDIC, Residual angle, gain correction, Finite length effect.

I. INTRODUCTIONMany hardware efficient algorithms exits, but these are not well known due to the dominance of software systems over the past quarter century. Among these algorithms is a set of shift – add algorithms collectively known as CORDIC for computing Sine and Cosine samples, trigonometric functions and other related mathematical functions. The trigonometric algorithm is called CORDIC, an acronym for Coordinate Rotation digital Computer. The trigonometric CORDIC algorithm was originally

developed as a digital solution for real time navigation problems. The original work is credited to Jack E. volder in 1959 [1], [2]. In this paper we have implemented a CORDIC algorithm in an efficient way and tend to maximize the speed of operation. The key aspect of the CORDIC algorithm is that the result is achieved using only shift, additions/subtraction and table look-ups which map well into hardware and are ideal for FPGA implementation. There are three methods to implement the CORDIC algorithm viz. Unrolled, Rolled and Pipelined architecture. Out


mailto:[email protected]




of these architecture Unrolled / unfolded architecture we have implemented to design a DDS (Direct Digital Synthesis) CORDIC oscillator [3]. This DDS CORDIC oscillator has applications in FFT and carrier recovery loops in software defined radios [4].

It is implemented using various register lengths from 20 to 8 bits, in case of 20 bits register to represent sine and cosine component and 14 numbers of CORDIC iterations and we are obtaining approximately 125 dB noise rejection. In describing the effect of variation of register length on our oscillator we first present the implementation of CORDIC based oscillator and generating the sine and cosine samples. [5] Then we deal with the tricks used to reduce the frequency error and obtain the

amplitude stability. Then we present the effect of length variation on spectrum and on sine and cosine samples. Finally Simulation and Synthesis result presents the finite length effect on CORDIC based oscillator.

II. CORDIC BASED OSCILLATOR

A) CORDIC AlgorithmCORDIC is a very interesting technique for calculating sine and cosine samples. The CORDIC method can be employed in two different modes: the “rotation” mode and the “vectoring” mode. In the rotation mode, the algorithm basic idea consists in decomposing rotation operation into successive basic rotations.

Each basic rotation can be realized by shifting and adding shift and add arithmetic operations. The rotation mode of the CORDIC algorithm could be used to compute sine and cosine of an angle θ. Outputs after “n” iterations are computed according to the following equations:

(1)

(2)

The variable d(i) is called the decision operator which is used to decide in which direction to rotate. It is nothing but the sign of the angle accumulator Z(i).The computation of sin (θ) and cos (θ) is based on the rotation of an initial vector of unit length, that is aligned with the abscissa (X0 =1 & Y0 = 0). Moreover, the accumulated angle is initialized with the desired rotation angle. For each iteration, a comparison is done between the initial angle and the resulting angle. Then, the comparison sign (represented by the variable d) is used to determine the sign of the next rotation. “Equation 1, 2 and 3” now represent the CORDIC algorithm for rotations in a Circular Coordinate System. Now these equations can be implemented by using iterative hardware. These N numbers of iterations in unrolled form can be described in VHDL. For details refer [6]The angle become progressively very small hence to represent them to integer the angle is scaled based on the number of input bits, e.g. scaling factor of 1000 is

used for 20 bit input and angle is represented in degrees as shown in table 1.

Table1.CORDIC rotation angle for different values of i

i 2-i θi = tan-1

(2-i)Scaling for 20 bit cordic

0 1 45.000000 450001 0.5 26.565051 265652 0.25 14.036243 140303 0.125 7.1250160 71254 0.0625 3.5763340 35765 0.03125 1.7899110 17896 0.015625 0.8951737 8957 0.0078125 0.7446142 7448 0.00390625 0.2238105 2239 0.001953125 0.1119057 11110 0.0009765625 0.0559529 5511 0.0004882812

50.0279764 27

12 0.000244140625

0.0139882 13

13 0.0001220703125

0.0069941 6

B) System Architecture

The architecture of system is shown in figure 1. X and Y are the sine and cosine registers that are initialized to 214 and 0 respectively. Zin is the angle increment and it decides the oscillation frequency.


The X, Y and Zin values are provided to the CORDIC processor which computes the next value of X and Y. It is stored back in the X and Y registers. Traditional DDS systems use an angle accumulator which is incremented by a fixed value with each clock.

Figure 1. Architecture foe CORDIC based oscillator

Our implementation is different in the sense that the sine and cosine registers keep track of the angle. The rotation sequence is identical at each step and the residual angle is also identical. As a result, there is no line structure in the spectrum. The internal architecture of CORDIC processor is shown in figure 2.

Figure 2. Internal architecture for CORDIC Processor

C) Residual Angle Correction

Note that after n numbers of iterations are over, a small residual angle may still remain in the angle accumulator. In our 14 stage implementation, it can be at the most .006 degrees (See table 1). If left unattended, it will give rise to frequency error. To take care of this residual angle, correction is applied to X and Y as suggested in [6] by using following equations.

(4)

(5)

At this stage α is very small therefore cos α = 1 and Sin α = α, where α is represented in radians. Therefore the above equation reduces to

(6)

(7)

These corrected values are given to the gain correction unit.

D) Gain Correction UnitWhen this DDS is used for carrier frequency generation in a software defined radio, the X and Y are subjected to indefinite rotations and there values are likely to expand or contract due to finite length effects. In ideal case, x and y should lie on the unit circle, i.e. 122 =+ yx . We ensure this in every step by multiplying x and y values with gain factor gn as suggested in [7]-[9]. This gain factor gn is

calculated as follows: 2

)]()([3 22 nynxgn +−=

Note that gn > 1 when the vector is inside the unit circle and gn < 1 when the vector is outside the unit circle. In case of 20 bits the x and y values are scaled by a factor of 214, therefore 3 in above equation is replaced with 3 * 228 and 2 is replaced with 2 * 228. But gn has to be scaled so that it will be an integer. Therefore, in above equation, the denominator is taken as 212 rather than 228. This has the effect of scaling the value of gn by 216.

After multiplying X and Y with gn, the result is right shifted by 16 places. This removes the above scaling effect.

III. FINITE LENGTH EFFECTThis CORDIC based oscillator using 20 bit register and angle and phase correction provides us approximately 125 dB noise rejections as shown in figure 3.


Figure 3. Gain plot for 20 bit CORDIC oscillatorTo analyze the effect on the performance of this CORDIC based oscillator we are varying the length of input bits and the number of iterations. The number of iterations in this oscillator is depends upon the register length used to represent angle. As shown in table 1, for 20 and 16 bits we are using a scaling of 1000 for representing angle in degrees and after 14 iterations angle becomes very small to represent. Similarly for 12 and 10 bit CORDIC we are using 10

numbers of iterations and the angle is scaled by 100 instead of 1000. And for 8 bit CORDIC we are using 4 numbers of iterations and here we are not scalingthe angle. The gain plot for these various register lengths is shown in figure 4.

(a)

(b)

(c)

(d)

(e)

Figure 4. Gain plot with gain and angle correction a) 20 bit 14 iteration oscillator b) 16 bit 14 iteration oscillator c) 12 bit 10 iterations oscillator d) 10 bit 10 iterations oscillator e) 8 bit 4 iterations oscillator


A) Effect of Angle CorrectionWe have implemented 15 degrees rotation in each step. As discussed in previous section if the correction is not applied to the residual angle correction unit it will give rise to frequency error so by applying correction we are reducing the frequency error. As a result of angle correction unit final angle after 1024 rotational steps is more accurate.The difference can be observed from the sine and cosine curves where the length of these curves are varied as shown in figure 5. The effect on the spectrum is slight shifting of the peak which is not visible. The results are shown for 16 bits CORDIC oscillator.

B) Effect of Gain CorrectionIn case of 20 bit CORDIC based oscillator effects of gain correction is negligible. In case of 10 bits the X and Y registers tends to expand and finally overflow as shown in figure 6 (a). In case of 8 bits the x and Y registers tends to decrease and finally becomes zero as shown in figure 6 (b). Gain correction unit stabilizes the amplitude as shown in figure 6.

(a)

(b)

Figure 5. Effect of angle correction on sine, cosine samples and gain plot for 16 bit CORDIC oscillator a) With gain and angle correction b) Without angle correction.

(a)


(b)

Figure 6. Effect of gain correction on sine, cosine samples and gain plot a) 10 bit CORDIC based oscillator without gain correction b) 8 bit CORDIC based oscillator without gain correction

IV. SIMULATION RESULTSThis design was coded in VHDL and simulated in Modelsim SE plus 6.2C. Main top level design was instantiated in a test bench which supplied clock, reset and Zin signals. Total 1024 Xout and Yout values were recorded in a text file in the test bench these values are read in MATLAB program. Each pair of X and Y is considered as a complex number vv. Blackman window (win) is applied to the vector vv and its FFT is calculated.

V. SYNTHESIS RESULATS

The VHDL code for the above design is synthesized using XILINX’s ISE 9.1i software. From Spartan 3E family XC3S1600E device is used for this purpose. We have synthesized the design for 20, 16, 12, 10 and 8 bits of input to the CORDIC oscillator. And for each bit length the numbers of iterations are also varied 14,10,8 and 4 numbers of iterations. The comparative study for finite length effect on a FPGA is presented in Table 2.

Table 2. Finite length effect on FPGA resources

Parameter Utilization

20 bit 14 iteration

16 bit 14 iteration

12 bit 10 iteration

10 bit 10 iteration

8 bit 4 iteration

Max. frequency 8.1 MHz 11.8 MHz 26.8 MHz 13.3 MHz 17.9 MHz

Min. period 122.8ns 84.26ns 37.3ns 75.02ns 55.83ns

No. of slices 5% 3% 1% 1% 1%

No. of LUT 5% 3% 1% 1% 1%

No. of IOs 64 52 40 34 28

Number of MULT18X18SIOs

83% 22% 13% 27% 27%

Number of GCLKs 4% 4% 4% 4% 4%

VI. CONCLUSION & FUTURE SCOPE


Finite length effect on FPGA resources requirement and spectral purity is studied successfully. It not only affects the spectrum but also the sine and cosine samples. This CORDIC based oscillator can be used in Software defined radio. The sine and cosine samples can be used for demodulation.

VII. ACKNOWLEDGEMENTThanks to the management of Ramdeobaba Kamla Nehru Engineering College and Jhulelal Institute of Technology, Nagpur, INDIA, for their constant inspiration and support.

VIII. REFERENCES[1] J.E. VOLDER, “The cordic trigonometric

computing technique”, IEEE Transactions on Electronic Computers, Sept. 1959.

[2] Walther, J.S., “A unified algorithm for elementary functions,” Spring Joint Computer Conf., 1971, proc., pp379-385

[3] Ultra Low Phase Noise DSP Oscillator [DSP Tips & Tricks] Harris, F.Signal Processing Magazine, IEEE Volume 24, Issue 4, July 2007 Page(s):121 – 124

[4] C. Dick, F. Harris, and M. Rice, “Synchronization in software defined radios—Carrier and timing recovery using FPGAs,” in Proc. IEEE Symp. Field-Programmable Custom

[5] Computing Machines, Napa Valley, CA, pp. 195–204, Apr. 2000.J. Valls, T. Sansaloni, A. Perez-Pascual, V. Torres, and V. Almenar, “The use of CORDIC in software defined radios: A tutorial,” IEEE Commun. Mag., vol. 44, no. 9, pp. 46–50, Sept. 2006.

[6] S S Limaye, VHDL A design oriented approach, Tata McGraw Hill Publishing company ltd., pp. 248-254, 2007.

[7] F. Harris, C. Dick, and R. Jekel, “An ultra low phase noise DDS,” presented at Software Defined Radio Forum Tech. Conf. (SDR-2006), Orlando FL, Nov. 2006.

[8] R. Lyons, Understanding Digital Signal Processing, 2nd ed. Upper Saddle River, NJ: Prentice Hall, pp. 576–578, 2004.

[9] C. Turner, “Recursive discrete-time sinusoidal oscillators,” IEEE Signal Processing Mag., vol. 20, no. 3, pp. 103–111, May 2003.

[10] Nitin P. Narkhede, S.S. Limaye, “FPGA Implementation of low phase noise oscillator” presented in International conference on Advances in computing, control and telecommunication technologies, ACT. pp 244-247,2009,http://doi.ieeecomputersociety.org/10.1109/ACT.2009/68,.

ADVANANCED SINGLE-PHASE BOOST CONVERTER SYSTEM FOR WIDE RANGE OF LOAD VARIATIONS

G.MURUGANAsst.prof/EEE [email protected]

J.MURUGANANTHAMAsst.prof/EEE [email protected]

ABSTRACT

Converters operated in discontinuous-Conduction- mode (DCM) and in continuous-conduction-mode (CCM) are

suitable for lighter and higher loads, respectively. A new, constant switching frequency based single-phase

rectifier system is proposed, which operates in DCM and in CCM for outputs less than and greater than 50%

rated load, respectively, covering a wide range of load variation. The power circuit and the control circuit of the

Proposed rectifier are easily configurable for DCM and CCM operations. The measured load current is used to

select the desired operating mode. The peak device current under DCM is limited to rated device current under

CCM without using a device of higher current rating. The input current shaping under CCM and DCM are based

on the comparison of measured input current with linear and nonlinear carriers, respectively. A load current


http://doi.ieeecomputersociety.org/10.1109/ACT.2009/68

http://doi.ieeecomputersociety.org/10.1109/ACT.2009/68

feedforward scheme is presented to improve the system dynamic performance and also to ensure a smooth

transition between the two operating modes. All the necessary control operations are performed without using

multiplication, division and square-root operation. The proposed rectifier shows improved input current

characteristics over the existing CCM converters for the above load range. This is validated on a 600-W rectifier

prototype. Simulation and experimental results are presented.

Key words—Continuous-conduction-mode (CCM), discontinuous- conduction-mode (DCM).

1. INTRODUCTION

INGLE-PHASE diode bridge rectifiers are gradually

being replaced by pulsewidth modulation (PWM)

rectifiers to maintain a sinusoidal input current at

near unity power factor and to satisfy the necessary

harmonic standards . A single-phase, single-switch

boost rectifier (Fig. 1) is a well-established topology

in the field of ac–dc power

conversion to comply with the above harmonic

standards. The rectifier system may be operated in

the continuous conduction mode (CCM), or in the

discontinuous conduction mode (DCM) . The CCM is

preferred over DCM because of continuous input

current and low conducted

electromagnetic interference (EMI) . However, it is

reported to have high input current distortion at

light load . For a particular switching frequency and

boost inductance, the amount of current distortion

increases as the load decreases . A high valued boost

inductor is necessary at light load to limit the input

current distortion . This increases the size, weight,

and cost of the converter and results in poor system

dynamic response. Hence, CCM is preferred at

higher loads. The above issues are not seen, when

the converter is operated

in DCM. However, DCM is always associated with

high device current stress and conducted EMI .

Therefore, a high current rated device and a costly

EMI filter are necessary at higher loads. Thus, DCM is

preferred for light loads The present work deals

with a constant output voltage application, where

the load current varies over a wide range (10% to

110% of rated load current) and the converter is

required to

comply with the necessary harmonic standards . It

can be seen from the above discussion that neither

of the operating modes (CCM and DCM) alone is

suitable and economical for the above application.

Therefore, the optimum choice is to operate the

converter in DCM during light loads and in CCM for

higher loads . The load boundary between DCM and

CCM operations can be set at a suitable level (say

50%) to limit the peak device current stress under

DCM up to the rated device


current under CCM without using a higher current

rated device. Similarly, the minimum load under

CCM, for which the converter is required to comply

is 50% rated load.

This permits us to use a low valued boost inductor

compared to the entire CCM case without any

degradation in the performance of the converter .

The main challenge associated with such a mixed-

mode operation is to realize the two distinct

operating modes (DCM and CCM) into a single

converter system without introducing any

appreciable dynamics during transition between the

two operating modes. There are two possible ways

to achieve this.

The first method suggests a single-valued boost

inductor with constant but two different switching

frequencies (a low switching frequency for DCM and

a high switching frequency for CCM) for the above

operation. The second method requires two

different boost inductors (a high valued inductor for

CCM and a low valued inductor for DCM) with a

constant switching frequency for above application.

The first method is simpler than the second method,

as it only requires the switching frequency of the

converter to be changed. However, the second

method requires the physical inductors to be

changed.

A converter system, using two different switching

frequencies(2.56 kHz for DCM and 25.6 kHz for CCM)

and a single valued boost inductor has been

reported in. The use of two different switching

frequencies introduces difficulties in designing the

EMI filter. The controller works in the principle of

voltage mode control without using any input

current sensor.

A current sensor is however required for over-

current protection of the converter. The input

current distortion under DCM is high as there is no

lowpass filter connected at the input to the

converter. The implementation of the above control

scheme involves complex mathematic operations,

such as multiplications, divisions and square root

operations. It also requires the peak value and the

zero crossing instants of the input voltage to

compute the unit vectors .These increase the

complexity and cost of the controller. Addressing the

above-mentioned issues and using two different

boost inductors (a high valued inductor for CCM and

a low valued

inductor for

DCM) a new,

constant-

switching-

frequency based

rectifier system

is proposed in

this paper. The


power circuit of the proposed converter system can

be configured either for CCM or for DCM by

performing a simple on-off control of an auxiliary

switch. A DCM power topology, with an input side

lowpass filter is obtained, when the auxiliary switch

is on. Again, a CCM power topology (without any

input filter) is realized, when the auxiliary switch is

off.

A simple, input voltage sensorless, current-mode

controller is proposed for the above rectifier

system. The controller works in the principle of one-

cycle control or the nonlinear carrier control

without using any of the above-mentioned

complex mathematical operations. The required

gating pulses for the converter switch are generated

by comparing the measured input current with one

of the two periodic carriers in a modulator. A linear

carrier is used under CCM, while a nonlinear carrier

is selected under DCM.

The measured load current is used to select the

desired operating mode (CCM or DCM). A simple

load current feedforward scheme is used to improve

the dynamic response of the converter system,

which also ensures a smooth transition from one

operating mode to the other. The proposed concept

has been simulated on MATLAB/SIMULINK platform

and experimentally validated on a 600-W prototype.

The simulation and experimental results are

presented.

2.PROPOSEDSINGLE-PHASE RECTIFIER SYSTEM

A single-phase, single-switch boost rectifier is shown

in Fig. 1, where Fig. 1(a) and (b) represent the DCM

and the CCM boost rectifier topologies, respectively.

A lowpass filter is used in the DCM topology for

filtering the switching current harmonics, which is

absent in the CCM topology. Further, it can be shown

that for the same switching frequency, the value of

boost inductor is much lower than . Thus, for the ame

switching frequency, the DCM topology is not

suitable for CCM operation and vice versa. Similarly,

it can be shown that the control scheme, suitable for

CCM application may not be useful in DCM

operation and vice versa. Therefore, a common

rectifier system (power circuit topology and control

scheme) is required to be developed, which is

suitable for both CCM and DCM. Such a rectifier

system is developed in this section.

A. Proposed Power Circuit Topology

The proposed power circuit topology is shown in Fig.

2,which is originally derived from the DCM topology

shown in fig. 1(a). There are two separate inductors

and and a filter capacitor used in the power circuit as

shown. The auxiliary switch may be turned on to

realize the DCM power topology same as Fig. 1(a).

The combination serves the purpose of the input

lowpass filter, while the inductor acts as the

effective low valued boost inductor. The switch can

be turned off, when the CCM topology is to be

realized. The filter capacitor remains ineffective in


the power circuit, while the series combination of

and acts as the effective high valued boost inductor .

A detailed design method for selecting ,

It is shown that by simple on-off control of the

switch , we can realize two different power

topologies [Fig. 1(a) and (b)] using a single converter

system, while maintaining a constant switching

frequency throughout. Now it is required to

understand the various issues associated with turn-

on and turn-off instants off of the thyristors are at

zero current (i.e. at ) ensuring a smooth transition

from DCM to CCM and vice versa.

It should be noted that it is also possible to use a

single thyristor and an anti-parallel diode to realize .

In that case, the thyristor (Fig. 2) may be replaced by

a diode of same polarity. The advantage of this

method is that no isolated gate drive is required to

drive the switch . Now, it is required to develop a

suitable controller for controlling the proposed

power topology, which is suitable for both CCM and

DCM operations. Such a controller is developed in

the following section.

B. Proposed Controller

The proposed controller works on the principle

of the switch and to propose a suitable semiconductor

switch for its realization. Let us consider a case,

when the converter system is required to be driven

into DCM from its original CCM operation. This

means that the switch , which was originally off is

now required to be closed. When the converter is

operated under CCM, the inductors and carry the

same instantaneous current , while the filter capacitor

does not carry any current. Under this condition can

be closed at any instant with zero current switching to

drive the converter system into

DCM. Now, let us consider the other possibility, i.e.,

the converter

system, which was initially operating under DCM and

is required to be driven into CCM. It can be seen

from Fig. 1(a) that, under DCM the instantaneous

currents carried by the inductors and are different.

Nevertheless, it can be shown that the current ,

through the switch passes through zero twice in a

switching cycle. Now, if we try to open at an instant


, the difference current will flow momentarily

through . This can cause over-voltage across

which can damage the switch. In order to avoid this

situation, we propose two anti-parallel thyristors,

each conducting during one half cycle of the

switching frequency current components,

for the realization of as in Fig. 2. Now both turn on

and turn

4.SIMULATIONAND EXPERIMENTAL RESULTS

The proposed control concept is verified through

simulation using MATLAB/SIMULINK and also

experimentally on a 600-W prototype shown in Fig.

10. Due to availability, inductors 6 mH and 0.4 mH

are used in the experimental setup. The control

circuit shown in Fig. 10 is implemented using

discrete integrated circuits (ICs). The op-amp full-

wave rectifier, amplifiers, voltage controller,

comparators,

summers and subtractors are implemented using op-

amp (TL084) based analog circuits . The input

lowpass filter, when connected in the ac side (Fig.

10) handles fundamental current along with

witching frequency current components. However,

when it is placed in the dc side (as in Fig. 6), it carries

dc, twice the line frequency and switching frequency

current

components. It is found that the placement of the

filter in the ac side gives better input current

waveform than its placement in the dc side.

however, no such significant change is noticed for

the inductor . In experimental setup, the inductors

and and the filter capacitor are connected in the ac

side as seen in Fig. 10. The auxiliary switch is realized

by two anti-parallel thyristors and as shown. The

input current is measured using an isolated current

sensor, while the output voltage and the load

current are measured using resistor based sensors as

shown.

State Input Voltage and Current Under DCM

The steady state input voltage and input current

waveforms, corresponding to 250Woutput power

are shown in Fig. 11. The converter is operated in

CM, where the auxiliary switch is turned on and the

nonlinear carrier is selected. The input voltage and

the current through the boost inductor are shown in

Fig. 11(a), while the input voltage and the input

current are shown in Fig. 11(b). The input current

THD and the input power factor are found to be

3.1% and0.988,respectively.


B. Steady State Input Voltage and Current Under

CCM

The steady state input voltage and input current

waveforms, corresponding to 600-W output power,

are shown in Fig. 12.

The converter is operated in CCM, where the

auxiliary switch is turned off and the linear carrier is

selected. The input current THD and the input power

factor are found to be 3.3% and0.98,respectively.

C. Variation of Input Power Factor and Input

Current THD With Load

In order to observe the above performances, the

output power of the converter is varied between 60

W and 600 W. Fig. 13 shows the variation of input

current total harmonic distortion the same figure. As

seen the DCM operation leads to less inputcurrent

distortion between 60-W and 300-W output power.

The maximum input current THD under DCM

operation is 5.5% (at 60 W). To obtain the same

input current THD at 60-W load under CCM, the

value of boost inductance required is 20mH(as seen

from simulation results). The experimental variations

of the measured input current THD and the input

power factor for different loads are shown inFig.

14(a) and (b), respectively. The THD increases as the

loaddecreases. It can be seen that the proposed

converter system is able to maintain a high power

factor and low input current THD over the entire

load range.

D. Dynamic Response of the Converter System

In order to test the dynamic response of the

converter system, the output load is suddenly varied

from 250 to 600 W and vice versa. Initially the

converter is tested without using the load current

feed forward scheme. The corresponding

experimental results

are shown in Fig. 15(a). The settling time of the

output voltage response is found to be around 160


ms. Now, the converter is tested with load current

feed forward scheme. The corresponding

experimental results are presented in Fig. 15(b). It

can be seen that, compared to the previous case

(i.e., without load current feed forward), there is no

overshoot or undershoot found in the output voltage

. The response of the input current as well as the

inductor current are also faster during the load

changes. It can be seen in Fig. 15(b) that the

transitions between CCM and DCM modes are

always smooth without any appreciable dynamics in

the output voltage as well as in the input current and

the boost inductor current .

The peak device current stress under DCM operation

is restricted to be close to the rated device current

under CCM operation.

E. Voltage and Current Stress in the Auxiliary

Switch

The voltage across the auxiliary switch and the

current through the switch are shown in Fig. 16. This

may be required to select the proper voltage and

current rating of the auxiliary switch . The maximum

possible voltage stress across is . This can appear

across for a short duration, if the filter capacitor is

charged at the peak input voltage, just before the

switch is turned off. The maximum factor (THD) with

load as obtained through simulation. The plot CCM

in Fig. 13 is obtained with the converter operating in

CCM at different loads over the entire load range

with a 6.4 mH boost inductor. This may be compared

against the plot DCM in current stress in the switch

is same as the rated device current.

5. CONCLUSIONA single-phase, constant-switching-frequency based rectifier system is presented for maintaining

sinusoidal input current at near unity power factor under wide range of load variations. The above rectifier system is operated under DCM for outputless than 330 W and in CCM for output greater than 330 W, exploiting the best features of both the operating modes. The power circuit of the proposed converter can be configured either for DCM or for CCM by simple on–off control of an auxiliary switch. Similarly, the proposed control circuit can also be configured either for CCM or for DCM simply by choosing the appropriate carriers (a linear carrier for CCM and a nonlinear carrier for DCM). The easured load current is used to select the desired operating mode. The required switching


instants are generated by comparing the measured input current with one ofthe above carriers in a modulator without using any multiplication, division, square root operation, and input voltage sensing. All the necessary design equations are provided to select the passive components. The averaged model of the proposed rectifier system is presented. Using such model, a design guideline for selecting the parameters of the voltage controller is presented. A simple load current feed forward scheme is presented to improve the dynamic response of the system against sudden change in loads. This also ensures a smooth transition from one operating mode to the other. The simulation and the experimental results show that the proposed rectifier system with mixed-mode operation gives better input current characteristics for a wide range of load variation compared to the case, when it is pirated completely in CCM for the given load range.

6.REFERENCES

[1] R. Ghosh and G. Narayanan, “Input voltage sensorless average current control technique for high-power-factor boost rectifiers operated in

discontinuous conduction mode,” in Proc. IEEE APEC’05, Mar. 2005,vol. 2, pp. 1145–1150.

[2] V. Grigore, J. Kyyra, and J. Rajamaki, “Input filter design for power factor correction converters operating in discontinuous onduction mode,” IEEE Trans. Electromag. Compat., vol. 1, no. 3, pp. 145–150, Aug. 1999.

[3] H. Paul and W. Hill, The Art of Electronics, 2nd ed. New York: Cambridge Univ. Press, 1989.

[4] IEEE Recommended Practices and Requirements for Harmonic Controlin Electrical Power Systems, IEEE std 519-1992, 1992.

[5] Electromagnetic Compatibility Part 3: Limits—Section 2: Limits forHarmonic Currents Emissions (Equipment Input Current _16 A perPhase), IEC 1000-3-2, 1992.

[6] J. Sebastian, M. Jaureguizar, and J. Uceda, “An overview of powerfactor correction in single-phase off-line power supply systems,”

CALCULATION OF STATIC TRANSMISSION ERROR FOR NYLON 6/6 SPURS HOUSER’S ANALYTICAL METHOD. - A CASE STUDY


PROF V.G. ARAJPUREdeptt of mechanical engg. bdcoe sewagram.

e-ma il- [email protected] . P. M. PADOLE

Deptt of Mechanical Engg. VNIT, Nagpur.e-mail- [email protected]

PROF.D.B.PAWAR

ABSTRACTThe performance of any machine tool is depending upon how accurately the motion istransmitted by gear drive. It is very important when the plastic gears are used in the drivearrangement. Plastic gearing is rapidly becoming a real contender in modular power gearingapplication and offering real breakthrough in performance also.The static transmission error is related in to the kinematic accuracy of gear train,dynamic load of gear tooth and noise, directly and strongly indicates the loss of powertransmitted by gear drive. The lower value of young modules of plastic responsible for largetooth deflection and so the static transmission error is more important and significant in geardesign.The paper presents the application of modified Houser’s analytical approach and appliesit to plastic (Nylon 6/6) gear material to calculate the static transmission error. The validity of Houser’s equation for Nylon spur gear is proved by comparing them with contact option ofFEM package MARC.

Keyword- plastic gear, static transmission error, compliance

1. INTRODUCTIONThe prediction of gear tooth dynamic load and gear noise has always been a major concern in gear designs. Greater emphasis has always been placed on the creature of analytical tools that may be employed to predict noise and to improve the dynamic performance of gears. High temperature resistant plastic helps to reduce maintenance cost and power consumption while ensuring longer wear. Plastic gears are corrosion resistant and can reduce noise significantly.2. TRANSMISSION ERROR:The study of error analysis will helpful for maximum power and torque transfer and reduce noise and vibr ation. Basically gear drive is for transmitting the power from one drive to other and having80% of efficiency compare to other power transmission device, like chain and belt.

The power transmission in gear is depending upon accurate profile of the gear tooth and actual gear tooth dimension, which are different. This difference is very small in amount called Gear transmission error which is further lead to noise and vibration in gearing system.

Plastic gears are used for low speed power transmission. To check the work performance of

plastic gear in light duty lathe machine and its durability and life testing error analysis is of


significant importance.The gear transmission error is because of mainly due to manufacturing limitation, material and accuracy and precision of machine tool used for gear manufacturing. The gear used in machine tool specifically lathe m/c. (feed gear box) is one of the source of noise and vibration in lathe m/c which further results in vibration of lead screw which is transmitting power from spindle shaft to tool post through service of gear and cause more friction at contact point of tool and work piece, which in result, more wear in tool nose radius and affect the quality of surface finish produce. This will also reflect in lower the tool life and accuracy of finish part in dimension. Finally, we can say it reduce the tool life time period and increase tool change timeperiod after a certain period of cycle of production and hence decrease the productivity of machine tool in terms of tool change time, tool wear, tool setup time, production accur acy, labour cost and lower the machinability index. The factors affecting the tool life are

1. Cutting speed

2. Depth of cut

3. Chip thickness

4. Tool geometry

5. Material of cutting fluid and

6. Rigidity of the machine

We consider a factor rigidity of machine tool and its effect on tool life. The rigidity of machine tool depends on the drive system in machine tool and our prime attention to use on the power transmission by gearing system. The vibration in cutting tool is because of due to powering system of that tool that means from spindle power to lead scr ew power and it’s given by feed box gear. More the gear transmission error in the feed box gear more the vibration in lead screw. This cause more wears in cutting tool nose radius.Measuring of tool life is based on the following factors:

1. Number of pieces machine between tool sharpening.2. Time of actual operation that means the time the tool is in contact With theaffect the quality of surface finish produce.job.3. Total time of operation.4. Equivalent cutting speed.5. Volume of material removed between tool sharpening.Hence by using steel gear lessvolume of material removed between toolsharpening in compare to plastic gearbecause less gear transmission error inplastic gear. This is because of due toaccurate profile of gear tooth of plasticmaterial i.e. machining capability, flowability in casting as well as easy plasticdeformation.On this basis, it is found thatmachinability which depends upon rate ofmaterial removal rate is good for plastic as compared to steel material.4. METHODOLOGY USED FOR MEASUREMENTOf Transmission Error:-Plastic gears are used for low speed powertransmission generally where the velocityof gear is less than 3.5 m/s. The Houser’sMethod is used for analysis of transmissionerror. The various dimensions required forcalculation of transmission error such as1) Base circle diameter,2) Pitch circle diameter,3) Addendum circle diameter and4) Deddundum circle diameter of gear and pinion in mesh respectively.


By using the computer progr am forcalculation of transmission error, the erroroccurred has been calculated. Initiallybending compliance, foundationcompliance and Hertzian compliance iscalculated and finally combine complianceat the point of contact at 0 and 1 has beenfound out. Along with this EQ separationhas been evaluated and finally transmissionerror (Etj) has beenfoundout. Thecomparativeanalysis of transmission error(Etj) of cast iron gear with Nylon66 gear ispresented as follows. Thermo plasticmaterial used isNylon66 for comparisonof transmission error with respect to cast iron gear.Gear Transmission Error for Plastic and Steel.

5.INPUT DATA FOR NYLON66 GEAR;1) Radius of Base Circle of Driving Gearin (cm) = 16.5152) Radius of Base Circle of Driven Gear(cm) = 34.5153) Radius of Pitch Circle of pitch Circle Driving Gear in (cm) = 17.554) Radius of Pitch Circle of pitch Circle of Driven Gear in (cm) = 35.325 5) Radius of Addendum Circle of DrivingGear in (cm) = 186) Radius of Deddundum Circle ofDriven Gear in (cm) = 367) Number of Teeth in Driving Gear = 50

8) Number of Teeth in Driven Gear = 100 9) Distance from pitch point in (cm) = 0.089

10) Center Distance in (cm) = 53.0111) Young’s modulus of Elasticity (E) in (MPa) 330012) Shear modulus of Elasticity (G) in (MPa) = 118713) Enter the value of passion ratio ( ) = 0.3914) Actual Element Width of teeth in (mm) = 0.36(mm) = 0.3615) Enter Different Values of Element Height and Arm Moment = 3.33, 0.943.16, 0.612.94, 0.222.72, 0.002.5, 0.52.0, 0.891.2, 1.2716) Distance of Contact Point Yj in (mm)= 1.28.RESULT :-1) Total Deflection (Dbj) = 5.5072) Effective Tooth length (Lf) = 1.9843) Effective Basic Thickness (Hf)= 4.004) Bending Compliance (Qbj) = 5.7805) Foundation Compliance at the Point ofContact (Qfj) = 0.0000696) Hertzian Compliance(Qh)= 0.0001097) Combined Compliance at the Pointof Contact (Qjo) = 5.7804498) Combined Compliance at the Point of Contact (Qj1) = 3.1214429) Calculated Value of (EQsep) = 0.2110)Transmission Error (Etj) = 0.0599. in mm.

6. INPUT DATA FOR STEEL GEAR:1) Radius of Base Circle of Driving Gear in (cm) =16.5152) Radius of Base Circle of Driven Gear (cm) = 34.5153) Radius of Pitch Circle of pitch Circle of Driving Gear in (cm) = 17.554) Radius of Pitch Circle of pitch Circle of Driven Gear in (cm) = 35.3255) Radius of Addendum Circle of Driving Gear in (cm) = 186) Radius of Deddundum Circle of Driven Gear in (cm) = 367) Number of Teeth in Driving Gear = 50 8) Number of Teeth in Driven Gear = 1009) Distance from pitch point in (cm)= 0.08910) Center Distance in (cm) = 53.0111) Young’s modulus of Elasticity (E) in(MPa) = 210 X10312) Shear modulus of Elasticity (G) in (MPa) = 25013) Enter the value of passion ratio ( ) = 1.49


14) Actual Element Width of teeth in (mm) =0.3615) Enter Different Values of ElementHeight And Arm Moment = 3.33, 0.94

3.16, 0.612.94, 0.22

2.72, 0.002.5, 0.52.0, 0.891.2, 1.2716) Distance of Contact Point Yj in (mm)= 1.28.Result :-1) Total Deflection (Dbj) = 12.3932) Effective Tooth length (Lf) = 1.9843) Effective Basic Thickness (Hf) = 4.004) Bending Compliance (Qbj) = 13.0085) Foundation Compliance at the Point ofContact (Qfj) = 0.0000016) Hertzian Compliance (Qh) = 0.0000037) Combined Compliance at the Point of Contact (Qjo) = 13.0083278) Combined Compliance at the Point of Contact (Qj1) = 7.0244979) Calculated Value of (EQsep) = 0.2110)Transmission Error (Etj) = 0.1349 in mm.7. RESULT AND DISCUSSION –Transmission error is one of the most important indications of gearing quality. The static transmission error describes the composite effect of any deviation of gear teeth from perfectly formed involutes surfaces. The static transmission error iswidely accepted as the principle source of vibration in gearing operations. Actualgears deviate slightly from perfect involutes surfaces and elastically deformunder load. These tiny surface deviationsresult in an unsteady force component ofthe torque. The unsteady forces aretransmitted as vibration from the gear,through the shaft and bearings to thegearing system. It is this unavoidablevibration that we exploit to noninvasivelyinvestigate the operating condition of gears while in operation.The transmission error cause due to manufacturing limitations causes due to material properties and characterizes in case of cast iron but such manufacturing limitations can be overcome by using plastic material due its better manufacturing capability than steel. The program involves the use of 2 D plain strain elements, coupled with contactelements at the points of contact between the meshing teeth. A simple strategy of how to determine an appropriate value of the penalty parameter of the contact elements (gap element) is also presented. It is found that with the increase in contact point value transmission

ratio decreases in case of plastic and slightly change in steel material. A comparative graph between steel and plastic material gear transmission error with respect to contact point is shown below.

8. CONCLUSION:The result obtained by using Houser’smethod for calculation of transmission error for different engineering thermoplastic materials like Nylon 66, Polycarbonate and Teflon are calculated. The results are compared with those obtained by existing method for cast iron gears. It is found that the tooth deflectionof thermoplastic spur gear is larger thanthat of cast iron gear. The influence ofpremature and delayed machine teethareoccurred is also consider for above threeengineering materials.9. REFERENCES-1.M.S. Tavakoli, D.R. Houser, optimumprofile modification for theminimization of static transmissionerror of spur gears. Journal ofmechanicaldesignandtransmission,A.S.M.E.,vol8,march-1986(86-95).2.Ming Huang Tsai, Ying-Chian Tsai, Amethod for calculating statictransmission errors of plastic spur gear,Finite element in analysis in design,vol-27,1997 (345-357).


3.Dr. Ir. H.G.H. van Melick, Toothbending effect in plastic spur gears,Gear technology, sep/oct-2007 (58-66)4.Clifford E. Adams , Plastic gearing –selection and application , Marcell-Dekker inc. Vol-49, 1986 (48-50, 99-106).

5.Earl Buckingham, Analyticalmechanics of gears , McGraw-Hill,March 1949 (474-479)6.S.S. Ratan , Theory Of Machines,McGraw-Hill,1997(350-353).

VIDEO INDEXING USING RADIAL BASIS FUNCTION NETWORK1S.PRASANNA, 2DR.A.JOTHI AND 3DR.S.PURUSHOTHAMAN

1S.Prasanna, Research Scholar, Dept. of MCA, VELS University, Chennai.2Dr. A. Jothi, Dean, VELS University, Pallavaram, Chennai, India

3Dr.S.Purushothaman, Principal, Sun College of Engineering & Technology, Kanyakumari – 629902,Tamil Nadu, INDIA, Email: [email protected]

ABSTRACT:

This paper indexes a video using the audio features and Radial Basis Function (RBF) neural network. The audio from the audio track corresponding to each shot in the video is extracted and compared with the audio query input by the users. Researchers have developed different methods to index a video. However, cepstrum analysis is used to extract features of the audio track and used as input to the RBF. The output of RBF will decide the presence of the audio query in the audio track of the video.

1. INTRODUCTION

Tools are required for representing, searching and retrieving content large digital video libraries One method is the query-by-example (QBE) approach, in which users provide (usually visual) examples of the content that they seek. However, such schemes have some obvious limitations and since most users wish to search in terms of semantic concepts rather than by visual content, John(1996), work in the video retrieval area has begun to shift from QBE to query-by-keyword (QBK) approaches which allow the user to search by specifying their query in terms of a limited vocabulary of semantic concepts. This paper focuses on QBK which presents the search engine, a keyword which is meaningful and understandable. This keyword is equivalently interpreted and searching takes place. Query using keywords representing semantic concepts has motivated recent research in semantic media indexing, Qian (1999), Korbus (2001), Michael(2001). Recent attempts to introduce semantics in the structuring and classi_cation of videos includes, Vasconcelos

(1998), Ferman (1999), Iyengar (1998), Wayne(1997), Michael (2001).

Probabilistic framework for semantic video indexing is presented by Naphade et al(1998) A library of examples approach, called semantic visual templates is given by Chang et al(1998). A rule-based system for indexing basketball videos is given by Zhang et al.(2000). Detecting sources of sounds in audio using cues are given by Ellies(1996). The semantic labeling is treated as a machine learning problem. We begin by assuming the a-priori definition of a set of atomic semantic concepts relevant to objects, scenes and events appearing in the video. The commentary in the video is closely related to various shots, scenes of the video. The set of video collection that meet the specification as mentioned will be suitable for analysis purposes. In case, where the video content and the audio content are different, then video indexing is not exact. A set of atomic semantic



concepts describing the audio event is established. (noise, water sound, lightening, roaring sound, whistling sound, animal sound, human voice, human speech( irrespective of words). These cannot be further decomposed. Concepts that can be described in terms of other concepts are then defined as high-level concepts. The definition of high-level concepts

will depend to some extent on the variety of atomic concepts defined.

2. PROPOSED METHODOLOGY

In this work different audio (sound) waves are considered. Features are extracted from each sound wave and stored as templates. In the search

engine, the name of the audio is given as semantic along with additional semantics that will express the real situation in which audio is present.

Fig.1 Schematic diagram

Explanation for the Figure 1 1. Inititally obtain features of different audio. Store the features and label the features corresponding to each audio.


QBK

Use local template details

Search the indexing details of a video (or) use audio processing to obtain features

Any matching for QBK

All video file searched?

Output the indexed video files

2. Finding the meaning of the semantics given as QBK and correspondingly access the features..

3. Access the already indexing details of the searching videos and locate the audio features (or) find the features for the audio.

4. Compare the details obtained in the step 2 with details of step 3 . If there is a close match then highlight the video as the indexed video.

5. verify if all the videos in the library have been searched. Else goto step 2.

3. EXTRACTION OF AUDIO FEATURES USING CEPSTRUM ANALYSISIn this work, the exactness with which features are extracted from continuous audio in successive frames depends on the type of signal processing methods used. Basically, cepstrum analysis has been used over a period of time by earlier researchers to extract low frequency information from high frequency noise. Incase of rocket launching, cracker blast, there cannot be a distinction among the low frequency information from high frequency noise. Both low frequency and high frequency will be the information representing the rocket launching or any other event. 3.1 Feature extraction using cepstral analysisThe result of signal analysis is a sequence of audio/speech frames, typically at 10 msec intervals, with about 16 coefficients per frame. These frames may be augmented by their own first and / or second derivatives, providing explicit information about speech dynamics; this typically leads to improved performance. The speech frames are used for acoustic analysis. It is convenient to assume that the signal consists of a discrete time sequence, so that the spectrum consists of a z transform evaluated on the unit circle. Let us consider a speech example, with X referring to the spectrum of the observed speech signal, E to the excitation component (for instance, the glottal pulse train), and V to the vocal tract shaping of the excitation spectrum. We begin with a multiplicative model of the two spectra (the excitation and the vocal tract). Thus, the spectral magnitude of the speech signal can be written as|X()| = |E()| | V()| (1)Taking the logarithm of above equation yieldslog|X()|= log |E()| + log |V()|. (2)Particularly for voiced sounds, it can be observed that the E term corresponds to an event that is relatively extended in time (e.g., a pulse train with pulses every 10 ms), and thus it yields a spectrum that should be characterized by a relatively rapidly varying function ; in comparison, because of the

relatively short impulse response of the vocal tract, the V term varies more slowly with function. With the use of this knowledge, the left-hand side of equation(2) can be separated into the two right-hand-side components by a kind of a filter that separates the log spectral components that very slowly (the so called high-time components) from those that very slowly (the low-time components). Such an operation would essentially be performing deconvolution.

Equation (2) has transformed the multiplicative formula (1) into a linear operation and thus can be subjected to linear operations such as filtering. Since the variable is frequency rather than time, notations must be changed. Thus, for example, rather than filtering (for time), we have filtering (for frequency); instead of a frequency response, we have a quefrency response; and the DFT (or z transform or Fourier transform) of the log |X()| is called the cepstrum. The cepstrum is computed by taking the inverse z transform of equation 2 on the unit circle, yieldingwhere c(n) is called the nth cepstral coefficient.Step 1: Input audio wave file from the videoA video is is demultiplexed to separate the embedded audio. The audio file is processed through the cepstrum analysis. The window frame considered is 10. Cepstrum values are extracted from the audio. The following steps are adopted to obtain cepstral values for words.

1. Input audio from the video track

2. Remove zeros and other noises which do not give any information.

3. Apply linear predictive analysis.


4. Apply fast fourier transform.

5. Apply log for the output in step 4

6. Apply inverse fast fourier transform.

7. Apply levinson Durban equation.

8. Repeat step 3 to step 7 for every samples of the data acquired from audio and average all values to finally get only 10 values.9. Repeat step 8 for the data collected from the other remaining audio. Hence after averaging there will be patterns with each pattern having 10 values. Label each pattern representing an audio.10. Repeat step 9 for other audio in remaining video file.

4. RADIAL BASIS FUNCTION (RBF)

RBF is a supervised artificial neural network (ANN) which works based on the distance concept. The distance is found between a pattern and each centre. The centre is also one of the patterns predefined. The square of the distance is a node in the hidden layer.

An exponential function is used as an activation function which will be the output of the particular node. The number of nodes in the hidden layer is

based on the number of centres decided in an implementation.

5. EXPERIMENTAL SETUP Sixteen audio semantics were considered in this work. They are given in Table 1. The corresponding frames are shown in Table 1.

Semantic Number of video files in which the events are present

Average number of occurrences in the video files

Number of video retrieved

A single frame

Rocket

launch

5 2.5 4

Airplane

landing

10 3.5 8

Airplane

takeoff

10 3.5 8


Racing car 20 5 17

Racing two

wheeler

20 6 19

Ship horn 3 1.5 3

Avalanche 4 2.5 4

Nuclear

explosion

1 5 1

Vulcano 3 4.5 3

Train horn 50 75 48


Bird sound 100 250 100

Waterfalls

sound

75 25 74

Telephone

ringing

30 100 28

Police whistling 30 300 25

Rain Sound 100 500 89

Building

collapse

15 100 14


Temple Bells 45 200 43

Cepstrum values have been generated for the audio of the above semantics. These cepstrum values are used as inputs for the RBF. Ten centers are used in hidden layer. One hundred video files have been used for experimentation. Each video file is of 50min at an average length with 25 frames/sec. The average percentage of video files retrieved 84.0%

CONCLUSIONS

This work presents the advantages of RBF in retrieving the video documents using 16 semantic audio words. The audio words that are in the form of wave file are converted into cepstrum features which are further trained by RBF. In the video retrieval, a combination of few or more out of 16 semantic words are input to the system. Subsequently, with equivalent translation, the frames of each video is analysed and the correct video that match the semantic words are retrived.

REFERENCE:

Chang S. F., Chen W., and Sundaram H.,Semantic visual templates -linking features to semantics, in Intl. Conf. Image Processing, Chicago, IL, October 1998, vol. 3, pp. 531-535, IEEE.

Ellis D., Prediction-driven computational auditory scene analysis, Ph.D. thesis, MIT, Cambridge, MA, 1996.

Ferman A. M. and Tekalp A. M., Probabilistic analysis and extraction of video content, in Intl. Conf. Image Processing, Kobe Japan, October 1999, IEEE.

Iyengar G. and Lippman A. B., Models for automatic classification of video sequences, in Storage and Retrieval from Image and Video Databases. Jan 1998, vol. VI, SPIE.

John R. Smith and S-F Chang, Visualseek: a fully automated content-based query system, In Proc. fourth intl. conf. multimedia. May 1996, pp.87-92, ACM.

Korbus Barnard and David Forsyth, Learning the Semantics of Words and Pictures," in Intl. Conf. on Computer Vision. 2001, IEEE.

Michael A. Casey, \Reduced-rank spectra and minimum-entropy priors as consistent and reliable cues for generalized sound recognition," in Proceedings of Eurospeech, 2001.

Naphade M., Kristjansson T., Frey B., and Huang T. S., Probabilistic multimedia objects (multijects): A novel approach to indexing and retrieval in multimedia systems," in Intl. Conf. Image Processing, Chicago, IL, October 1998, vol. 3, pp. 536-540, IEEE.


Qian R., Hearing N., and Sezan I., A computational approach to semantic event detection," in Intl. Conf.

on Computer Vision and Pattern Recognition, Fort Collins, CO, June 1999, vol. 1, pp. 200-206, IEEE.

Vasconcelos N. and Lippman A., Bayesian modeling of video editing and structure: Semantic features for video summarization and browsing, in Intl. Conf.

Image Processing, Chicago IL, October 1998, vol. 2, pp. 550-555, IEEE.

Wayne Wolf, \Hidden Markov model parsing of video programs," in Intl. Conf. On Acoust., Sp., and Sig. Proc. 1997, IEEE.

Zhang T. and Kuo C., An integrated approach to multimodal media content analysis, in Storage and Retrieval from Image and Video Databases, San Jose, CA, January 2000, vol. 3972, pp. 506-517, SPIE.


REVIEW PAPER ON ANALYSIS AND SYNTHESIS OF ANTENNA ARRAY PATTERNS USING CONVEX OPTIMIZATION

AMANPREET KAURThapar University, Patiala M.Tech, PhD PursuingProfessional Membership if anyIETE

E Mail [email protected] No 09815601313 Work experience: 4YearsM.Tech thesis guided:7 Area of interest; Wireless & Antennas

Publications: National: 6JASPREET KAUR

Baba Banda Singh Bahadur Engg. College, Fatehgarh Sahib M.Tech,Professional Membership if anyNA E Mail [email protected] No 09463058507

Work experience: 4YearsM.Tech thesis guided: 6 Area of interest; WirelessPublications: International Journals: 4International Papers: 2 National Papers: 1

GURDEEP MOHALRIMT-MAEC Mandi Gobindgarh M Tech PursuingProfessional Membership if any NA E Mail

[email protected] No09501021605 Work experience: 4Years 8 monthsArea of interest; Wireless & AntennasPublications: International Papers: 1 National Papers: 5

ABSTRACT:

Antenna Array patterns can be analyzed and synthesized in a variety of ways and are very important in almost every field of Antenna applications. Antenna Array patterns can be expressed as Convex Optimization problem and can be optimized by a number of methods available nowadays e.q. Interior Point Methods, Genetic Algorithm, SCHELKUNOFF'S METHOD, These methods usually analyzed the Antenna Patterns in terms of Side Lobe Level, Main Lobe, Noise Power, Signal to Noise ratio.In this paper, we consider a uniform linear Antenna array with 40 elements with two equality constraints. It has been verified through the simulation results, using Convex Optimization that the side lobes and the major lobes and beam width of side lobes can now be optimized with precision as well as visible on the plots

Keywords: convex, uniform antenna, optimization, side lobe level, Beam width

I INTRODUCTION



A single element Antenna is unable to meet the required gain and radiation pattern. So the solution is, to combine several single element antennas to form an Antenna Array. Antenna arrays are used to direct radiated power in a particular direction. There are many types of Antenna Arrays. In this paper, we are considering only Linear and Broadside Antenna Arrays. After finding the array response, signals are weighted and summed up to obtain the beam pattern.

Antenna Array Pattern synthesis problem consists of finding the optimum weights to satisfy the given requirements. The 1st synthesis problem has been studied by Schelkunoff [1] or Dolph [2]. Literature is available on antenna array optimization by direct methods Such as genetic algorithms [11] and particle Swarm Optimization [ 12]These methods are feasible because of increased computing power. An important comment in [3] is that there is no guarantee to reach at an optimum solution unless the problem is convexBut as stated in [4] not all antenna array problems are convex.

A Linear Convex Optimization problem can be solved by many available optimization algorithms. Even Nonlinear Convex Optimization problems can be solved by using Karmarkars Interior Point method [5]

In this paper, we formulate the importance of Convex Optimization in synthesis and analysis of Linear and Broadside Antenna arrays and show that the side lobe level can be minimized by Convex Optimization.

In section II, we formulate the array factor of two basic Antenna Arrays. Section III deals with the Convex Optimization and briefs about the methods given above. The section IV is equipped with some examples of Antenna Array Pattern optimization And, in the section V we describes the analysis of results.

II. MATHEMATICAL FORMULATION OF ANTENNA ARRAY PATTERN

Linear Array

Consider a linear array of n isotropic elements of equal amplitude and spacing d as shown in fig. The total electric field E in direction ø is given by

ψψψψ )1(321 ............1 −+++++= njjjj eeeeE

Where ψ is the total phase difference of fields from adjacent sources and is given by:

αφλψ +Π= cos*)/(2 dwhere φ is the phase difference between the feed

currents of adjacent sources. φ will be same for all the sourcesIn case of linear arrays , elements are equally spaced at λ/2 distance.

The array factor of this linear antenna depends upon the number of elements, the element spacing, amplitude, and phase applied to each element Array factor is given by


Y

ThetaZ

BROADSIDE

XENDFIRE

∑=

+∏=N

iii yxjA

1

))sincos(/2exp( φφλ

The N signal outputs are converted to complex numbers weighted by weights Wi and summed up to give the linear array beam pattern

)1......())sincos(/2exp()(1

∑=

+∏=N

iiii yxjwH φφλφ

where TNi wwwww ]........,,[ 321=

is the complex weight vector to be designed

Beam width or HPBW of Antenna:

Beam width is a measure of directivity. It is angular width in degrees, measured on the radiation pattern between points where the radiated power has fallen to half its maximum value called Half Power Beam Width or the Gain is one half the maximum value. Beam width is the product of individual antenna pattern and array factor shape.

For reflector antenna Beam width is given byDkHPBW /λα ==

K is the factor that depends on the shape of the reflector and the method of illumination. For a typical antenna k=70o

Angular width between the first nulls or first side lobes is called Beam Width between the First Nulls (BWFN).The half power levels occurs at those angles for which

707.02

1),( ==φθE

Gain may be expressed in terms of Beam width

2)/( αη Π Κ=G

III. CONVEX OPTIMIZATION

In Mathematics, the term Optimization refers to the study of problems in which we minimizes or maximizes a real function by choosing the real values from allowed set. A mathematical optimization problem has the form:

minimize fo(x)subject to fi(x) <bi i=1,........m

Here vector x is the optimization variable of the problem, the function fo : Rn R is the objective function, the function fi: Rn R i=1....m are constrained functions

A Convex Optimization problem is a problem where all of the constraints are convex functions and the objective is a convex function if minimizing or concave if maximizing. Linear functions are convex that’s why we are using convex optimization.

The following are the analysis about the Convex Optimization problem

• if a local minima exists, then is a global minima.

• the set of all global minima is a convex.• if the function is strictly convex, then there

exists at most one minimum.The various Convex Optimization problems are: Least square, Linear programming, Conic optimization, Geometric programming, Semi definite programming etc.Convex optimization problems can be solved with a variety of methods like Ellipsoid method, Sub gradient method, cutting plane method, Interior-point method,


Side lobe pattern

Schelkunoffs Method:

In this method making the substitutions ψjew = in the array factor given by (1) we obtain

∏=

−=N

inN wwIH

1

)2....(..........).........()(φ

If IN=1 ,the magnitude of (2) may be expressed as

)3........(.........)( 21 NwwwwwwH −−−=φ

Null steering can be achieved by varying the positions of the roots

Interior Point Method: In very large-scale optimization problems, it is not possible to form Jacobian and Hessian matrices. In that case, interior point method is used. Actually Interior Point method is very much efficient in obtaining the optimal solution of non linear equations. There are a number of interior point methods like: Barrier method, Inequality constrained minimization, Logarithmic barrier method, Primal interior point method These methods gained great popularity after Karmarkar [6] given these methods for polynomial. A large amount of work has been done on this, given by Gonzage[7] and Nesterov [8] who developed a very general framework for solving nonlinear convex optimization problems using IPM.

Interior point method moves in the interior of the feasible region hoping to bypass many corner points on the boundary of the region.

Bisection Method

Bisection method is a practical method to find the roots of an equation. It repeatedly bisects an interval then selects a subinterval in which a root must lie for further processing.The method is applicable when we wish to solve the equation f(x)=0 for the scalar variable x, where f is a continuous function.Suppose that 0)( afc = and 0)( ℏbfd = . If f is continuous then it must be zero at some x between a and b. The bisection method then

consists of looking half way between a and b for zero of f

i.e. let 2

bax += and evaluate )(xfy =

Unless this is zero, then from signs of c, d, y we can decide which new interval to subdivide.

This is done by flow control in most common way of if……..else……end statement.

IV. PROBLEM STATEMENT

We have to minimize the beam pattern level given by (1) with the possible constraints

min max H(ø)Wi i=1,2,....N

subject to H( øo)=1

The formulation of Convex Optimization problem is given by Lasdon [9]. Now to apply the Interior point method, we have to eliminate the equality constraint first by expressing the last weight with the first one

)1(1

1

)cos()/2()cos()/2( ∑−

=

Π−Π −=N

i

xji

xjN

NN ewew οο φλφλ

Now replace the objective function with a new variable p by adding new constraints Therefore it is easy to express the original problem in the form of convex problem which can be easily solved with interior point method.

Further we have to find out the feasible solutions for beam width of array. This can be done by Bisection method and then using optimization we have to find the most feasible solution for the beam width.

V. SYNTHESIS OF ANTENNA ARRAY BEAM PATTERN

Antenna synthesis problem consist of determining the structure of antenna for given set of specifications. Antenna array synthesis consists of finding the currents of array elements and some parameters of desired radiation pattern .The synthesis of antenna is same as to finding the


impulse response of FIR filter for desired frequency response. In this paper, we tried to minimize the side lobe levels and beam width of a linear and broadband antenna with nulls constraint and finding the weight vector which satisfies the desired specifications

VI. SYNTHESIS RESULTS

Linear array analysis

As given in the problem statement, we worked to optimize a linear antenna array. The array has 40 elements with inter element spacing of 0.45 λ . The main lobe direction is 60o with unit sensitivityThe original diagram without optimization is given in Figure 1. In which side lobe level comes out to be -8.91dB i.e. the minimum value of side lobe magnitude in dB is -8.91dB. This pattern is not suitable for many applications that’s why we optimize this pattern and all the side lobes of all the elements we have taken out are positioned at one feasible value.

Figure 2 gives the optimal pattern which minimizes the side lobe level between 0 and 180 degrees. The main lobe shape and position is exact as without optimization. As shown in the figure that all the side lobes have one magnitude i.e. -25.3dB. This level of -25.3dB is taken out by satisfying the two constraints one is that the target direction is 1 and the nulls we were considering are zero.

We have taken only two constraints so the result is not satisfied. With some new constraints we can find the more optimal solution.

0 2 0 4 0 6 0 8 0 1 0 0 1 2 0 1 4 0 1 6 0 1 8 0- 2 0

- 1 8

- 1 6

- 1 4

- 1 2

- 1 0

- 8

- 6

- 4

- 2

0

l o o k a n g l e

mag

y(t

heta

) in

dB

Fig 1Sidelobe pattern of linear array with nooptimization and with SLL of -8.91dB

In Fig. (1), the main beam direction is 60 degree and the width of the main beam is 50 to 70 degrees. We have taken the nulls at five different points shown by pink lines The plot shows the various side lobes with unequal amplitude at various positionsThe main beam is sharpened as the no of elements increases.Total no of lobes =n Main lobe width=4л/nMinor lobe width=2л/n

0 2 0 4 0 6 0 8 0 1 0 0 1 2 0 1 4 0 1 6 0 1 8 0- 5 0

- 4 5

- 4 0

- 3 5

- 3 0

- 2 5

- 2 0

- 1 5

- 1 0

- 5

0

l o o k a n g l e

mag

y(t

heta

) in

dB

Fig 2 Optimized array pattern with SLL of -25.3dB

0 5 0 1 0 0 1 5 0 2 0 0 2 5 0 3 0 0 3 5 0-4 0

-3 5

-3 0

-2 5

-2 0

-1 5

-1 0

-5

0

lo o k a n g le

ma

g y

(th

eta

) in

dB

Fig 3 Optimized beam width array pattern

Beam Width Analysis: In this paper we have shown that, antenna gain is degraded if a low side lobe level is sought, so that expansion of antenna beam width is evitable. In this paper we have used weighting with window functions to achieve maximum gain for desired beam width . Fig 3 shows the minimized beam width pattern of an antenna array with 36 elements and the target direction for the main lobe is 60 degrees. In this example we have taken the minimum side lobe level


of -20 dB and maximum half beam width is 60.We use the bisection algorithm to find out the feasible value for beam width which comes out to be 12 degrees.In this we are only considering the integer values of half beam width and sampling the angle with 1 degree resolution.The reference side lobe level we have taken is -20dB.By adjusting the phase of central element we can increase the directivity and decrease the beam width as well as minimize the side lobe level.

Parameter

Without Optimization

With Optimization

Method used for Optimization

Side lobe Level

-8.91dB -25.3 dB Optimization is done using Matlab

Beam width

2 degrees 12 degrees Optimization is done using bisection method and reference SLL taken is -20 dB

VII CONCLUSION

We have shown that how Convex Optimization can be used to design Optimal Antenna Arrays. We have used interior point method which finds the global optimum values with precision. So, Convex

Optimization is an excellent tool for pattern synthesis. We have a number of convex optimization algorithms but still, a lot of work can be done in this field. We have shown in this paper that Convex Optimization can solve linear as well as non-linear problems and we have chosen simple numerical examples. This can be used in almost every kind of application of optimizationWe have taken the SLL and antenna beam width ,one can optimize the pattern in terms of other antenna parameters.

VIII. ACKNOWLEDGMENT

We would like to thank, Herve Lebret and Stephen Boyd, without whom we would have never been able to understood Convex Optimization. We would also like to thank, L. Vandenberghe, for the review of Interior Point method. We would finally thank the reviewers, whose able suggestions, helped us in writing this paper.

REFERENCES:

[1] S. A. Schelkunoff, “A mathematical theory of linear arrays,” Bell Syst.Tech. J., vol. 22, no. 1, pp. 80–107, Jan. 1943.

[2] C. L. Dolph, “A current distribution for broadside arrays which optimizes the relationship between beam width and side-lobe level”, Proc. IRE, vol. 34, pp. 335–348, June 1946.

[3] O. M. Bucci, D. D’Elia, G. Mazzarella, and G. Panatiello, “Antennapattern synthesis: A new general approach,” Proc. IEEE, vol. 82, pp.358–371, Mar. 1994.

[4] Herv´e Lebret and Stephen Boyd“Antenna Array pattern synthesis via convex optimization” IEEE transactions on signal processing, vol. 45, no. 3, march 1997


[5] wright, stephen (1997). primal-dual interior-point methods. philadelphia, pa: siam. isbn 0-89871-382-x.[6] N. Karmarkar, “A new polynomial-time algorithm for linear programming”, Combinatorica, vol. 4, no. 4, pp. 373–395, 1984.

[7] C. C. Gonzaga, “Path-following methods for linear programming,” SIAM Rev., vol. 34, no. 2, pp. 167–224, June 1992.

[8] Yu. Nesterov and A. Nemirovsky, Interior-Point Polynomial Methodsin Convex Programming, vol. 13 of Studies in Applied Mathematics.Philadelphia, PA: SIAM, 1994.

[9] L. S. Lasdon, J. Plummer, B. Buehler, and A. D. Warren, “Optimaldesign of efficient acoustic antenna arrays,” Math. Programming, vol.

39, no. 2, pp. 131–155, 1987

[10] S. Boyd, L. Vandenberghe, and M. Grant, “Efficient convex optimization for engineering design,” in Proc. IFAC Symp. Robust Contr. Des., Sept. 1994, pp. 14–23.

[11] Mahanti, G.K. et al., “Discrete Phase-Only Synthesis of A Dual Beam Collinear Dipole Antenna Array Using Genetic Algorithms”, International Journal of Theoretical and Applied Computer Science, Vol. 1, No. 1, pp. 63-70, 2006.

[12] Gies, D. and Rahmat-Samii, Y., “Particle Swarm Optimization For Reconfigurable Phase-Differentiated Array Design,” Microwave And Optical Technology Letters, Vol. 38, No. 3, 168-175,August 5, 2003.

TIME REDUCTION FOR WIRELESS COMMUNICATIONUSING OFDM

1BALRAJ B, 2SIVAKUMAR D, 3THAMARAI SELVI D

1Departement of Electronics & Instrumentation Engineering,Annamalai University,Annamalai Nagar – 608002, Tamilnadu, India. Email : [email protected]

2Professor / Electronics & Instrumentation Engineering,Annamalai University,Annamalai Nagar – 608002, Tamilnadu, India. Email : [email protected]

3Assistant Professor / Electronics & Communication Engineering,Krishnasamy College ofEngineering and Technology, Cuddalore – 607109, Tamilnadu, India.

Email : [email protected]

ABSTRACT Orthogonal Frequency Division Multiplexing (OFDM) is a multi–carrier modulation system employing Frequency Division multiplexing of orthogonal sub–carrier, each modulating a low bit – rate digital stream. OFDM symbol is generated by computing IFFT. The design of IFFT / FFT is the main consideration in OFDM transceiver design. The parallel – pipelined FFT architecture based on multiplier less implementation targeting in High – Speed Wireless Communication application such as IEEE 802.11 wireless base band chip and CDMA. These have advantages of high throughput and high power efficiency. The multiplier – less architecture uses shift and addition operations for complex multiplication. By using low power butterfly the resulting power and area savings are increased upto 20%. 64 point FFT is used compared to Wallace tree multiplier.

Key-Word : FFT, IFFT, OFDM, CDMA

INTRODUCTION


http://en.wikipedia.org/wiki/Special:BookSources/0-89871-382-X

http://en.wikipedia.org/wiki/Special:BookSources/0-89871-382-X

http://en.wikipedia.org/wiki/International_Standard_Book_Number

OFDM is a multichannel modulation system employing FDM of orthogonal sub-carriers each modulating a low bit rate digital stream. In OFDM, to overcome the problem of bandwidth wastage, N overlapping but orthogonal sub-carriers each carrying a baud rate of 1/T and spaced 1/T apart are used. For recent wireless systems, such as IEEE 802.11a providing 54Mbps data rate, increased throughput requires further parallelization. It means more than one Processor Element needs to be assigned per column to the FFT. Parallel - pipelined FFTs are suitable for both high throughput and high power efficiency. Parallel pipelined FFTs have not significant area and it can operate at lower frequency and low power consumption. This paper discus the application of common sub expression sharing across coefficients to the second stage of 64-point or the first stage of 16-point FFTs. In the past complex multiplication shifters and adders were used and some special constant coefficients. In this FFT architecture, the Complex multiplications are replaced by minimum number of shift and addition operations.

Hence both area and power consumptions for the multiplier unit are reduced.

ALGORITHMSA. FFT Algorithm

The Discrete Fourier Transform of N complex data point x(n) is defined by

∑−

==

1

0)()(

N

n

nkNWnxkX

k = 0, 1,……N-1;

Where )/2( Nj

N eW π−= , WN is twiddle factor or

coefficient. Previously [7] R4SDC pipelined FFT algorithm for word sequential data radix r, the equation (1) can be written as

∑ ∑−

=

−

=+= 11

01

1

0 111

1)1()( N

q

r

p

pk

rN wqpNxkqWkX

This N point can be decomposed into V stages where N=r1, r2,………..rv. The final stage is X(r1r2….rv-1mv+r1r2….rv-2mv-1+r….r1m1+m1)

= ∑ −

−

−−−−1

1

1)1,1(1v

v

vr

qmvq

rvvvv WmqX .

The intermediate stages are given by recursive equationXt(qt-mt)

∑ −

= −− +=−

1

0 11 ),(1

tt

t

r

ppmt

rtttttmtq

N WmqPNxW0≤qi≤ni-1, 2≤i≤v and Nt=N1(r1r2…..rv), 2≤t≤v-1,0≤mi≤ri-1. For r1=4, the 16 point FFT is shown in figure. Dots define stage borders. Open cycle denotes summations. Number outside the open circle is twiddle factor. Architecture in figure, It has 75% utilization of complex multiplier and 100% Butterfly.

2.2 Common Subexpression SharingCommon subexpression sharing shares the subexpression among several multiplication-accumulation operations in order to reduce the total number of operations. This approach is very effective for reducing the hardware cost of multiple constant multiplications, especially for the filter-like operation. For example, for a 3-tap FIR filter, the output Y(2) is given as follow.

∑ = −×= 2

0)2(

i ini XAYThe weights A1 are the filter coefficients. Suppose the coefficients are given as A0 = 00111011, A1= 00101011, and A2 = 10110011. The coefficients are represented in two’s complement format. According to the equation Y(2) = A2 ×X2+A1×X1+A2×X0.Using shifts and additions to replace the multiplications, gives: Y(2) = X2+X2 << 1+X2 << 3+X2 << 4+X2 << 5+X1+X1 << 1+X1 << 3+X3 << 5+X0+X0<<1+X0<< 4+X0<<5-X0<< 7.

(Signal Flow Graph of a radix – 4, 16 point FFT)


(N-point radix-4 pipelined FFT processor architecture)

The computation required twelve additions, one subtraction and eleven shifts. However, if pre-computing X02 = X0+X2; X12 = X1+X2; X012 = X12+X0, the output can be shown as : Y(2) = X012+X02 << 4+X012 << 1+X012 << 5+X12 << 3- X0 << 7.

(16-point 2-parallel pipelined FFT architecture)This computation only needs seven additions, one subtraction and five shifts. X02, D12, X012 are the common subexpression for this case.. From the above examples, it can be shown that common subexpression sharing can reduce the number of additions and subtraction from 13 to 8 (i.e. 38% reduction).

2.3 Canonic Signed Digit (CSD)It is common to use the redundancy of signed digit code to replace the conventional multiplier digits, such that addition operations in a multiplication can be reduced with the increase of average shift length across the zeros in the multiplier. Canonical Signed-Digit (CSD) is a widely used signed digit approach. In CSD code of a number, each bit is set to 0, 1 or -1 and no two consecutive bits are nonzero. The advantages of CSD form is that no value has more than (N+1)/2 nonzero bits, often fewer, and so the multiplication by a constant requires no more than that number of additions for its implementation.

IMPLEMENTATION3.1 Reduction Multiplier R4SDC FFT

In this algorithm it consists of four real multipliers, one adder and one subtractor. The complex co-efficient for all stages are precomputed. The calculated co-efficients are shown in table. For the trivial co-efficient (7FFF, 0000) and (0000, 8000), the complex multiplication is not necessary. An additional unit, which swaps the real and imaginary parts of input data and inverts the imaginary part for (0000, 8000). The rest of the co-efficients are composed of only 6 constants (7641, 5A82, 30FB, A57D, 89BE, CF04). For example, a multiplication with the constant A57d could be realized by first multiplying the data with 5A83, and then two’s complementing the result.

The coefficients for 16 point R4SDC FFT Table

Coefficient sequence m1=0, 1

Original quantized coefficient

Coefficient sequence m1=2, 3

Original quantized coefficient

W0 7FFF, 0000 W0 7FFF, 0000W0 7FFF, 0000 W2 5A82, A57DW0 7FFF, 0000 W4 0000, 8000

W0 7FFF, 0000 W6A57D, A57D

W0 7FFF, 0000 W0 7FFF, 0000W1 7641, CF04 W3 30FB, 89BE

W25A82, A57D

W6A57D, A57D

W3 30FB, 89BE W9 89BE, 30FB

Note that a multiplication by the constant 5A82 already existents. Therefore, the multiplication with the constant 5A83 can simply be obtained by adding the data to the already existing multiplication with 5A82. The other two constants (89BE and CF04) can be realized in a similar manner, using constants 7641 and 30FB respectively. 5A82 is represented by two’s complement format, 7641 and 30fb are represented by CSD format as follows,


5A82010110101000001076411000-10-100100000130FB010-1000100000-10-1The mixed use of CSD and two’s complement is for minimizing the number of addition/shift operations. We can use shifters and adders based on the three constants to carry out those nontrivial complex multiplications as shown below:

5A82X =5X<<12+5X<<9+65X<<1

7641X = X<<15+65X-5X<<930FBX = 65X<<8-X<<12-5X

Where X means input data. The common subexpressions for the three constants are 101(5) and 1000001(65).

3.2 Conventional ButterflyThe conventional butterfly architecture consists of 6 adder / subtracters. In this paper, we proposed a low power butterfly architecture which employs two 5-input summation blocks to replace six adder / subtracters.

(Block Diagram for Conventional Butterfly)Figure shows the conventional butterfly architecture. Inverters (CI1 to CI6) are used to generate the normal or the one’s complement form under the control of c5, c6 and c7. The signal C4 controls the four multiplexers (M1 to M4) for directing appropriate data to the inputs of the summation blocks. Two 5-input summation blocks (SUM0 to SUM5) are employed to generate the real and imaginary parts of

the output respectively.

An additional decoder unit is used to generate compensation for eliminating the error which results from the one’s complement inversion controllable inverters.

SIMULATION RESULTS

(Simulation Result for FFT)

(Simulation Result for (IFFT)

CONCLUSION


The Paper presents parallel pipelined architecture for 64 point FFTs. This has multiplier-less and Low power butterfly. By synthesizing the power consumption can be compared with the normal FFTs.

REFERENCES

[1] S. He and M. Torkelson, “Design and implementation of 1024-point pipeline FFT processor” Custom Integrated Circuits Conference, 1998. Processing of the IEEE 1998, 11-14 May 1998. pp. 131-134.

[2] K. Maharatna, E. Grass and U. Jagdhold “ A 64-point Fourier Transform Chip for High speed Wireless LAN Application using OFDM” IEEE Journal of Solid-State Circuit. Vol. 39, No.3, March 2004.

[3] Wei Han, T. Arslan, A.T. Erdogan and M. Hasan “ Multiplier Less based Parallel Pipelined FFT architecture for Wireless Communication Applications” in IEEE 2005, CASSP 2005, Pp V-45 – V-48.

[4] G. Bi and E.V. Jones, “A pipelined FFT processors for word-sequential data”, IEEE Transactions on acoustics, speech and signal processing, Vol.37, no:12, Dec.1989, pp.1982-1985.

[5] Wei Han, T. Arslan, A.T. Erdogan and M. Hasan “ A novel low power pipelined FFT based on subexpression sharing for wireless LAN applications”, in IEEE signal processing systems workshop, 2004. (SIPS 2004), Oct 2004,pp. 83-88,

EFFECT OF SIC CONTENT ON MECHANICAL AND TRIBOLOGICAL PROPERTIES OF AL ALLOY SIC COMPOSITESSANJAY SONI, S. DAS, G. DIXIT Sanjay Soni is with the Maulana Azad National Institute of Technology, Bhopal, India phone: 0755-2670416; fax: 0755-2670562; e-mail: sansoni_in@ yahoo.com. S. Das is with the Advanced Materials and Processes Research Institute, Bhopal , India, e-mail: [email protected]. Dixit is with the Maulana Azad National Institute of Technology, Bhopal, India, e-mail: [email protected]

ABSTRACT—The present work deals with Al-Si alloy (ADC-12) and its composite developed by reinforcing 10 wt%, and 15 wt% SiC particles in the Al matrix. The effect of change in volume fraction of SiCp has been studied on the mechanical and tribological properties of alloy and composites. The mechanical properties such as tensile strength, 0.2% proof stress, hardness, impact strength of ADC 12 aluminium alloy and composites were studied in order to achieve the maximum properties. In tribological study, the detailed examination of worn surface is done to obtain information on the mechanism of material removal during sliding wear behaviour. The subsurface deformation behavior during sliding wear of the Al alloy & composites is also studied through SEM analysis, which provides information on the extent of subsurface deformation and formation of mechanically mixed layer (MML). Formation of MML is also related to material transfer mutually between pin surface and counter disk. It is observed that the mechanical properties such as hardness and yield stress are improved due to SiC addition. Impact strength of the composites remains constant on increase in SiC content. There is a decrease in tensile strength of composite as compared to ADC12 alloy. The wear rate of the composites has improved considerably and the seizure pressure and seizure temperature has also increased. The wear mechanism and sub surface deformation studies were evaluated. Scanning electron micrograph of subsurface shows MML, a thin plastically deformed and bulk undeformed zone; the extent of subsurface deformation is less. With increased applied pressure, the


extent of subsurface deformation increases. A highly deformed region of thickness 40 µm, just below the MML, can be seen at higher applied pressure.

Key words — Al-Si alloy; SiC particles; tensile strength; 0.2% proof stress; hardness; impact strength; subsurface;Mechanicallymixedlayer.

1. INTRODUCTIONcomposites are becoming important as high performance materials combining properties not obtained in individual material. In the recent technological innovation, high strength, light weight, wear resistant and cost competitive material aluminium alloy reinforced with hard ceramic particles (Al composites) are being synthesized. These are having combination of properties which are achieved by tailoring the matrix & reinforcement (dispersoid) [1-4].The present work deals with the aluminium – silicon in view of the fact that hard silicon phase(s) impart(s) low thermal expansion coefficient, excellent wear resistance, hardness, strength etc. Further, the alloy possesses good fluidity in molten state enabling to make intricate shaped castings. Aluminium – Silicon alloys are used for making automobile components subjected to friction and wear like pistons, cylinder liners, bearings etc, however, binary Al-Si alloys could not offer the level of strength properties required by the components [5-9]. As a result, they are alloyed with elementslikecopper, nickel, magnesium etc. leading to theattainment of desired combinations . of physical,mechanical and tribological properties. Improvement in the properties of the alloys through this technique could be attained to the required level and thus the alloys became more popular and acceptable. [10].Addition of second phase particles to aluminium based alloys (A1-Si, A1-Cu, A1-Si-Mg, A1-Zn alloys) exhibited tremendous improvement in the mechanical, tribological and elevated temperature properties [11]. Alloying is limited to phase diagram, and hence the gaining of properties is also restricted. In an another process, where hard ceramic particles are added externally to Al-alloy to improve the mechanical, tribological and high temperature properties. In Al- alloy composites one can achieve the advantage of both light weight and high ductility of Al alloy and high hardness and stiffness of ceramic particles. Thus, in Al-alloy composites, one can achieve a combination of properties [12-16]. In the present investigation AMCs have been synthesized through Stir Casting technique using Al-Si as matrix alloys and SiC particle as reinforcement The size range of SiCp lies between

40 - 80 microns and the amount of SiC used is 10 wt% and 15 wt%.2. EXPERIMENTAL PROCEDURE2.1 MaterialAluminium-Silicon alloy (ADC12) is selected as matrix alloys for synthesis of AMCs. The chemical compositions of the aluminium alloy was analyzed using glow discharge spectrometer (model: GDS 500A, Leco) which is shown in the Table 1.

Table 1Chemical composition of ADC 12 alloy

Element

Si Mn

Mg

Cu Fe Ni Al

ADC-12 10.3

0.12

0.47

1.98

0.75

0.80

Rest

2.2 Reinforcement The SiC particles, used as reinforcement in the aluminium matrix for synthesizing the composites, were obtained from M/s. Grindwell Norton ltd., Bangalore, India. The particles are sieved using standard sieving practice through different grades of sieves in a vibratory sieving machine, with an aim to get particles within the size range of 40-80 µm. The various properties of SiC particles are shown in table2.

Table 2Properties of SiC reinforcement

Particle

Elastic Modulu

s GPa

Densit

y gm/cc

Coefficient

of Thermal expansion

K-1

Specific Hea

t Kg-1

K-1

Thermal

Conductivity

Wm-1 K-

1

Poisons Ratio

SiC 420-450

3.2 4.3x10-6

840 10-40 at

1100/o

C

0.17

2.3 Composite PreparationMelting of the ADC-12 alloy was carried out in graphite crucibles using a oil-fired furnace. When the


alloy reaches semi pasty stage the surface is covered with a fluxing agent (coverall 11). About 70 g of flux is added to 10 kg of the molten alloy. The dross is removed from the surface of the melt using a refractory skimmer. The dissolved gases were removed by purging dry nitrogen gas into the melt for 5 minutes. After degassing the surface is again cleaned and the temperature of the melt is increased to 660oC to take care of the minimum gas absorption during bubbling.

The composite was prepared by the stir casting method developed and patented by AMPRI (formerly RRL), Bhopal. The method comprises of dispersing the second phase particles (α-SiC) on to the vortex of the molten alloy. The vortex was created with the help of a rotating mechanical stirrer. The speed of rotation was maintained around 500 to 600 rpm. The SiC particles 10 wt% and 15 wt% of the desired size range (40 to 80 µm) were heated to 1000oC in a graphite crucible in a muffle furnace. The SiC particles were introduced into the melt by churning action of the stirrer. It produces great amount of turbulence in the flow pattern and induces better mixing of the SiC particles in the aluminium alloy melt. After the complete addition of SiC particles, the speed of the stirrer is brought down to 200-400 rpm and the stirring continued for 3-5 minutes. The composite melt was then solidified into a cast iron permanent die in the form of fingers (10 mm/20mm diameter and 200 mm long) for preparation of specimens for evaluation of different mechanical properties. The alloy was also cast in the same die.

2.4 Specimen PreparationThe specimens for various mechanical properties evaluation were prepared in accordance with the BIS code mentioned in table 3.

Table 3Test Code Size No of

specimensTensile 1608 10 mm Ø 3 x 3Charpy Impact

IS 1757 55x10x10 V notch

2mm

3 x 3

Vickers Hardness

IS 1501 20x20 mm 5 x 3

Wear (Pin on Disc)

--- 10 mm Pin 3 x 10

2.5 Mechanical Testing: The specimens in as cast condition for alloy and the composites of varying volume fraction were then tested for various tests. Mechanical properties such as UTS, and 0.2% proof stress, were evaluated on computerized Universal testing machine (INSTRON M/c, Model 5586) 300KN Capacity at a strain rate of 0.5 mm/min at Metallurgical services Lab, Mumbai, hardness and impact were evaluated on Optical Vickers hardness tester (FIE make, model VM 50) and Charpy Impact Tester (FIE make, model FIT 30) at MANIT, Bhopal. Wear property of the alloy and the composite was determined using Pin on Disc tester (DUCOM make, model TR-20LE) at different load conditions for 5000 m run.

The test was conducted till the Seizure of the specimen was observed. The wear rates (m3/m) of the samples were calculated on the basis of Volume loss. Variation of other factors like temperature rise, change in friction force and coefficient of friction was also measured. Finally the wear surface was examined using SEM for wear surface study.

2.6 Microstructural Examination:Samples for Microstructural characterization were polished according to standard metallographic procedures. The samples were etched with Keller’s reagent and observed in SEM (JEOL make, JSM-6390)

3. RESULTS AND DISCUSSION3.1 Mechanical Properties

Figure (3.1) shows the variation of tensile strength of ADC12 alloy and Al -composites in as cast conditions. It is noted from the figure that in case of ADC12 alloy the tensile strength is around 151 MPa which decreases to 126 MPa in case of ADC12 -10%SiC and further to a value of 118 MPa for ADC12 -15%SiC composites. The reason for decrease in tensile strength of composites as compared to ADC 12 alloy is that during the tensile deformation of the composite the soft aluminium matrix will elongate and the SiC particles which are very hard in nature are not deformed at the same rate as the matrix does. This non-uniform deformation of matrix and hard particles results into development of cracks at the interface. Hence, decohesion and void formations contribute to lowering the tensile strength of aluminium alloy particle composites then the aluminium alloy.


Tensile Strength v/s Alloy & Composite

80

90

100

110

120

130

140

150

160

ADC 12 Alloy ADC 12 - 10%SiC ADC 12 - 15%SiC

Tens

ile S

treng

th, M

Pa

Fig 3.1Fig (3.2) shows the variation of 0.2 % proof stress of ADC12 alloy & composites. The 0.2% proof stress of the ADC12 alloy-10%SiC composite and ADC 12-15% SiC composite are found around 160 MPa. This clearly indicates that alloy and composites yield at more or less same stress level. Which indicates that the elastic to plastic transformation of the material is solely depends on matrix materials and it has got not much significance in effecting the yield point of the materials.

Fig 3.2Fig (3.3) shows the variation of hardness of ADC12 alloy and composites in as cast conditions. It is noted from the figure that in case of composites the hardness increases with increase in SiC content. The value of hardness being 92 HV in as cast condition which is increased to 101 HV & 108 HV in the case of 10%SiC and 15%SiC composites respectively. The reason for increase in hardness value of composite is due to the fact that dispersion of SiC particles in aluminium alloy offer resistance to deformation of the ductile Al matrix. A good interface between the SiC and aluminium matrix always give rise to higher hardness value. Incorporation of 15% SiC particles in aluminium matrix shows highest hardness values amongst all.

Hardness v/s Alloy & Composite

80

85

90

95

100

105

110


Har

dnes

s, H

v

Fig 3.3Fig (3.4) shows the variation of impact strength in composites as compared to ADC 12 alloy. It is observed from the figure that the impact strength is almost constant and does not decrease with addition of SiC particles. The present results clearly indicate that impact strength of the aluminium alloy and composite remain almost constant by addition of SiC particles.

Impact Strength v/s Alloy & Composite

0

0.05

0.1

0.15

0.2

0.25


Impa

ct s

tren

gth,

Kgm

Fig 3.4

3.2 Wear PropertiesSliding wear behavior of Al-alloy & composites was studied as a function of applied pressure & volume fraction in as cast conditions. The response parameters such as wear rate, temperature rise, seizure pressure & coefficient of friction were measured during the test.

3.2.1. Effect of Applied Pressure on Wear RateFig 3.5 shows the wear rate of ADC12 alloy and its composite as a function of applied pressure at a sliding velocity of 4.4 m/sec. It is observed from the figure that the wear rate initially increases slowly with increase in applied pressure irrespective of the material. Further increasing beyond a certain applied pressure, the wear rate shoots up rapidly. It is further noted that in case of composites, the wear rate remains constant up to a certain applied pressure before seizure of the specimens. But in case


0.2 % Proof Stress v/s Alloy & Composite

60

70

80

90

100

110

120


0.2%

Pro

of S

tress

, MPa

of alloy the wear rate shoots up rapidly with applied pressure. The transition pressure from mild wear to severe wear and seizure pressure, in case of composites, are considerably higher than that of the alloy.

0

2

4

6

8

10

12

0 0.5 1 1.5 2

Applied Pressure, MPa

Wea

r Rat

e x

10-1

2 m

3 /m

ADC12 Alloy

ADC12 Alloy+10%SiC

ADC12 Alloy+15%SiC

Fig 3.5

3.2.2 Effect of Applied Pressure on Temperature Rise

Fig 3.6 shows the variation of maximum temperature for alloy and composites as a function of applied pressure for alloy and composites. It is evident from the figure that the temperature increases with increase in applied pressure At a certain applied pressure, temperature increases suddenly to a considerably higher value. This pressure corresponds to the seizure pressure and temperature corresponds to the seizure temperature. The temperature rise can be divided into three regimes for example at a pressure of 0.25 MPa and 0.5 MPa the temperature were in the rang of 50 C to 85 C (regime –I), at applied pressure of 0.75 MPa to 1.0 MPa the temperature were in the range of 100oC to 150oC (regime – II), and at an applied pressure of 1.25 MPa to 1.75 MPa the temperature were above 150 oC (regime –III).

0

50

100

150

200

250

0 0.25

0.5 0.75

1 1.25

1.5 1.75

2 2.25Applied Pressure , M Pa

Tem

per

atu

re, o C

ADC 1 2 AlloyADC 1 2 Alloy+10%SiCADC 1 2 Alloy+1 5%S iC

Fig 3.6

3.2.3 Effect of Applied Pressure on Coefficient ofFriction

The variation of coefficient of friction of ACD 12 alloy and its composites are shown in Fig (3.7) It is noted that the coefficient of friction of the alloy and the composites varies up & down with in a narrow band with increase in applied pressure. The range of coefficient of friction for ADC 12 alloy is primarily between 0.3 to 0.4. For 10%SiC composite its value lies between 0.3 to 0.4 but at the seizure pressure it goes below 0.3. For 15%SiC composite there is a sharp increase in COF and its value lies between 0.5 to 0.6. Again at the seizure pressure there is a sharp decrease in the value which recorded below 0.4.

Fig 3.7

3.3 Wear Surface StudyThe wear surface of aluminum alloy (ADC 12 Alloy) and ADC 12 composites is studied using Scanning Electron Microscope (SEM). In order to understand the mechanism of material removal during sliding wear, the worn surfaces of Al alloy & composites were cut, cleaned with acetone and fixed on a copper holder using double – sided copper tape and coated with Platinum, and observed in a Scanning Electron Microscope. 3.3.1 Wear Surface Study of Cast ADC 12 AlloyThe wear surface of ADC 12 alloy studied at a sliding speed of 4.4 m/sec, was observed under SEM. It shows that during the sliding wear, there is asperity to asperity contact between the sliding surfaces. The harder asperity, the steel surface in this case (62 HRc), slides over the softer aluminum surface and results into formation of grooves. A typical wear surface of (as Cast) ADC 12 Alloy at a pressure of 0.25 MPa is shown in (Fig 1). It shows continuous grooves and some damaged regions. It is interesting to note that during sliding action, cracks are formed on the surface which propagates along the longitudinal as well as transverse directions (arrow marked, Fig 2). It may be noted that transverse


0.1

0.2

0.3

0.4

0.5

0.6

0.7

0 0.25 0.5 0.75 1 1 .25 1 .5 1 .75 2

Applied Pressure, MPaC

oef

fici

ent

of

Fri

ctio

nADC 12 Alloy ADC 12+10% SiC ADC 12+15% SiC

cracks join the grooves together and form debris particles. At lower pressure, the wear surface is damaged in such a manner that it forms finer sized debris particles. On increasing the pressure to 0.75 MPa, the depth of grooves is increased and more damaged regions are observed (Fig3). Further increasing the pressure to 1.0 MPa, the sample tends to seize. At higher pressure, the material softens because of temperature, and tends to weld on a micro-scale. The as cast ADC 12 alloy seized at an applied pressure of 1.25 MPa and the wear surface is characterized by the formation of parallel lips (Fig 4).

Fig 1 SEM micrograph showing a typical wear surface of as cast ADC 12 alloy at a pressure of 0.25 MPa

Fig 2 SEM micrograph showing damaged regions of as cast alloy

Fig 3 SEM micrograph showing increased depth of grooves and damaged regions

Fig 4 SEM micrograph showing formation of parallel lips on seizure

3.3.2 Wear Surface Study of ADC 12 CompositeThe wear surface of 10% SiC and 15% SiC composite shows similar features as those of base alloy. It is observed that presence of SiC particles in aluminum matrix enhances the seizer pressure. This clearly indicates that by dispersing SiC particles in aluminum alloy matrix, the thermal stability of the alloy is increased. In case of composites, at lower pressure regimes, the groove formation is not very prominent and the wear surface is dictated by a smooth surface with some scratches only. When we increase the pressure to 0.75 MPa, some shallow grooves are seen. It is also seen that the rate of wear is comparatively less in the case of 15% SiC composites. On increasing the pressure to 1.5 MPa, the wear surface is characterized by shallow wear grooves, pittings and some damaged regions (Fig 5).


The sample tends to seize at a pressure of 1.625 MPa, as seen by the formation of parallel lips (Fig 6).

Fig 5 SEM micrograph showing shallow wear grooves, pittings and some damaged regions

Fig 6 SEM micrograph showing formation of parallel lips on seizure

3.3.3 Mechanism of Material Failure.

3.3.3.1 Wear surface of Al alloyThe SEM photograph (Fig 1) of a wear surface shows the continuous grooves running from one end of the sample to the other end. This happens when the hard asperities of the counter surface (in the present case it is steel surface) scratches the softer surface (in this case it is aluminium) and produces continuous grooves. The depth of grooves depends upon the applied pressure used during the wear test. The width of the grooves is dictated by the width of the asperities of harder surface. In the initial stage there is asperity to asperity contact and during the sliding action the softer asperity is broken down by the interaction with the harder asperity and later on there will be a direct contact between the harder

asperities and softer surface. During the sliding action there will be a groove formation on softer surface which results into work hardening of the Al surface.In the initial stage the wear of material is dictated by the scratching action. On prolonged sliding action the asperities of the hard surface will also be broken down and a situation comes when there will be a direct surface to surface contact. This leads to formation of oxide layer more on the softer Aluminium surface and the black colour oxide surface is discernable. This happens in the mild wear regime. Further the oxide layer will be removed by continuous sliding action. Further enhancing the applied pressure the stress on the softer surface will be more which leads to brake down of oxide layer and fresh surface will be evolved. At high load regime there is a every possibility of formation of sub surface cracks which join together and lead to higher wear by formation of plate shaped debris. The wear surface at high load regime is more of damaged regions. This is represented by (Fig 7). Further enhancing the load the friction become higher and the temperature rise also increases to a higher level which leads to incipient melting at the contact junctions. This melting results into sticking behavior of the materials and not allowed for free sliding action. This phenomenon is called the inception of seizure. The wear surface at the time of seizure is represented by parallel lips formation (Fig 4). As compared to other Aluminium alloy Al-Si alloy (ADC12 alloy) always show better wear resistance properties. It is known that Si which is a harder material then Al and because of its independent nucleation it is distributed uniformly in the matrix as discreet particle with physically identified feature. Because of this nature of Si it improves the bearing properties of Al when it is added to Al.

The interphase between Al and Si is quite good and during wear process there is no debonding at the interface and during wear process Si goes along with the Al.


Fig 7 SEM micrograph showing wear surface with more damaged regions at high load regime

3.3.3.2 Wear Surface of CompositesIn case of Al-Si-SiC composite, the SiC particles which are harder particles contributing to enhancing the wear resistance of the material. During the sliding operation initially the softer Al phase wear out and SiC particles protrudes in due course of time (Fig8). These protruding SiC particles comes in contact with harder counter surface and there will be a direct contact between harder steel surface and SiC particles. This results into lowering the wear of Al Si composites as compared to Al alloys. It is also interesting to note that the SiC particles are intact on the surface and there is no removal of particle from the surface. This clearly suggests that Al and SiC bonding is quite good. While increasing the concentration of SiC to 15 wt % the wear of material is reduced to an appreciable amount which is mainly due to the presence of hard SiC particle on the surface. The wear surface examination reveals neither fragmentation of SiC particles nor interface decohesion. In the case of sliding wear the shear stress is developed on the surface and there is a chance of stress concentration at the Al/Si particle interface. If the interfacial bonding is not strong enough then there will be a chance of interface decohesion and the particle will be removed from the surface due to shear stress. On the other hand if the interface bonding is good enough then the chance of load transfer from the particle to the metallic phase through the interface will be higher without damaging the interface. The later phenomenon is prevailing in the present situation. There is always a question of enhancing the concentration of SiC particles in Al matrix for the improvement of seizure resistance. The optimization of SiC loading in Al matrix depends upon processing

techniques and also to retain the ductility of the material. The presence of SiC particles no doubt enhances the brittleness behaviour of material which increases the probability of crack nucleation in the matrix due to generation of residual stress.

Fig 8 SEM micrograph showing protruded SiC particles

4.CONCLUSIONFrom the investigation of mechanical properties and sliding wear of the alloy and the composite the following conclusion can be made. 1. The hardness of the composite increases by 9% and 15% in case of 10%SiC and 15%SiC composite respectively as compared to as cast ADC 12 alloy.2. The impact strength of alloy and composites remains same.3. The tensile strength of 10%SiC and 15%SiC composite decreases to around 16% and 22% respectively as compared to as cast ADC 12 alloy.4. The yield stress (0.2% proof stress) remains the same in both the cases.5. The wear resistance of the alloy is improved significantly due to particle addition. The wear rate decreases with increasing SiC content and increases with increasing sliding distance and applied pressure. The coefficient of friction increases with increase in SiC content. 6. The seizure pressure of 10%SiC and 15%SiC composite increases by 20% and 30% respectively.7. The seizure temperature of 10%SiC and 15%SiC composite increases by 15% and 12% respectively.

8. Scanning electron micrograph of subsurface of an alloy shows MML (Mechanically Mixed Layer) which is extended upto 40µm below the surface. In case of


composite the SiC particles are found intact in the subsurface region but fragmentation of particle was observed in the form of trapped particle in MML at higher applied pressure.9. Sliding wear mechanism in case of alloy and composite are almost same. The stages associated with sliding wear are (i) contact of the counter surfaces primarily at the asperities (ii) asperity-asperity interaction, deformation and fracture of asperities, micro grooving by higher and stronger asperities leading to formation of fewer wear debris which get entrapped in the wear track, and surface of the pin and counter surface (iii) oxidation of wear debris and surfaces of pin and counter surface, (iv) compaction and mixing of wear debrises with the surface material of pin and counter surface and thus formed MML of MMOL (v) thermal and plastic incompatibility in addition to frictional heating and frictional force(shear force) leads to longitudinal and transverse cracking of MML, (vi) relatively more softer and partially melted subsurface material (form high deformed region) came out to the pin surface and create more pressure on MML, (vii) MML gets cracked and removed and consequently softer and partially melted material flows in the form of waves along the sliding direction and finally adhered with the counter surface. 10. I 10. In case of composite, the stages of wear mechanism are same but protruded SiC particles cause less contact between the counter surfaces. Additionally, SiC particles cause more abrasive action over counter surface and thus more amount of counter surface material transferred to the pin surface leading to thicker MML, which protect the composite surface more effectively. The SiC particles resist the subsurface material flow and thus turbulence like plastic flow of matrix material around these particles is noted.ACKNOWLEDGEMENTSThe authors would like to thanks Director, MANIT, Bhopal and Director, AMPRI (RRL), Bhopal for the facilities extended to carry out the present research work.

REFERENCES

1. I. J. Polmear, Light Alloys: Metallurgy of the Metals, Edward Arnold (publishers) Ltd., London (1981) 1. 2. Carl Zweben, Composite Materials and Mechanical Design, Mechanical Engineer’s Handbook, 2nd Ed., Myer Kuts, Ed., John Wiley & Sons, Inc., New York, (1998).3. T. Suganuma, A. Tanaka, Jr. of Iron and Steel Inst. of Japan, 75 (1989) 376.4. P. K. Rohatgi, “Cast Aluminium Matrix Composites for Automotive Applications” J.Metals, (1991) 35. 5. K. K. Chawla, Composite materials, 2nd edition, 204, Springer, (1988).6. B. Decelis, Wear, 116 (1987) 287.7. B.K. Prasad, O.P. Modi and A.K. Jha, Tribology Interenational, 27 (1994) 153.8. V.S. Aigbodion, S.B. Hassan, Mater. Sci. & Engg., A 447 (2007) 355-360.9. A.R. Bunsell, Fibre reinforcement for composite materials, vol.2, composite material series, 1988.10. A.R.Rehman, Ph.D thesis on Aluminium alloy composite for automobile brake application, (2004), MANIT, Bhopal.11. R. Asthna, Solidification Processing of Reinforced Metals, Trans. Tech. Publications Ltd. Cleveland, Key Engineering Materials, 151-152 (1997).12. O.M. Schuster, M.D. Skibo, W.R. Hoover., Light Metal Age., (1989) 15.13. V.G. Gorbunov, V.D. Parshin, V.V. Panin, Russian Cost Production, 8 (1971) 348.14. P. K. Rohatgi, “Cast Aluminium Matrix Composites for Automotive Applications” J.Metals, (1991) 35. 15. Alcan Aluminium Corporation (Duralcan), US Patent 4,786 (1988) 467.16. T. Sridharan, Extended Abstract of Conference on Metal Matrix Composites: Property Optimization and Application UK, IOM. (1989).


“F-TEST APPROACH FOR EVALUATION OF PROCESS PARAMETER EFFECT IN MANUFACTURING FOR L25 ORTHOGONAL

ARRAY BASED EXPERIMENTATION”PROF.V.R. NAIK,

Head, Mech.Engg.Dept., Textile & Engineering Institute, ‘Rajwada’, Ichalkaranji. Dist.Kolhapur (M.S.) Mobile : 9323268945,

I.S.T.E. Life Member LM 22832. email ID : [email protected]

PROF.DR.S.D. MADNAIK,Ex.Principal, Walchand College of Engineering, Sangli. (M.S.)

email ID : [email protected]

ABSTRACT :

It is important to uncover how the various process parameters and environmental factors affect the ultimate performance of the product. Performance may be affected by design parameters, some factors in the environment, and also possible interaction among design and environmental factors. Various methods have been devised to study this interaction like – Chi – Square Test, T – Test & F – Test. The present work deals with experimentation carried out on sample component on lathe machine. Process parameters are defined and L25 orthogonal array method is used to set up design of experiments. F Test based approach is used and discussed in this paper which enable understanding of interaction effect of process parameters. Also actual validation of the experimentation output is carried out by using VGA camera and Image processing technique.

Keywords:Orthogonal Array, Design of Experiments, F-Test, Chi-Square Test, T-Test.

INTRODUCTION :In any manufacturing process there are many input parameters, which contribute to output of the process, in form of variability in the product. To carry and experimentation problem formation is based in the output, produced on all geared lathe for turning operations with following specification –

Make – Panther, Spindle HP – 3Model No. – 1350/2,16 speed all geared headstockRange – 13 to 1250 rpm

In conventional manufacturing process current trend is to use lathe or similar operations or special purpose machines as a part of rough or semifinish operations and carry out finishing operations on very sophisticated machines like CNC or VMC.Since later group of machines have high accuracy and repeatability level, exhibited by very good process capability indices. (Cp) and (CpK), the root cause of introduction of variability is preprocess like rough/semifinish operations. Such operations were taken as problem definition providersand accordingly

manufacturing operations were carried out for sample component under consideration.

The 3D Cad model of drawing is shown in Fig. No.1.Fig. 1 : Sample component for experimentation.Basic input parameters identified were speed, feed, depth of cut and cutting tool material, which vary as a part of experiment. Fixed parameter was Material of Construction (MOC) of component which is M.S.




IS:1570 Fe C40. The attribute under consideration was max. taper diameter having dimension 10 + 0.55.

Approach :Process parameters identified were speed, feed, depth of cut and tool material. These factors were varied in five levels. L25 orthogonal array based design of experimentation[1] was devised.Accordingly sample component was manufactured on the lathe machine of said specifications. The output was tabulated for pin diameter under consideration as shown in Table No.2.F-Test based[1] approach was used to study effect of parameters and within the parameters, as elaborated subsequently.L25 Orthogonal Array :

L25 Orthogonal array was constructed for various parameters under consideration as specified belowSpeed (rpm) 0 192, 1 384, 2 572, 3

796, 4 1250Feed (mm/rev) 0 0.468, 1 0.5, 2

0.562, 3 0.525, 4 0.70

Depth of cut (mm) 0 0.1, 1 0.125, 2 0.150, 3 0.160, 4 0.17

Tool material 0 HSS, 1 Brazed, 2 TC, 3 WC, 4 CC

Tool 2: Insert – CPMT 060204R - PF4015,

Holder – A 12 M SCLPR 06Tool 3: Insert – N 123 g20 300 0003 –TF 4125,Holder – RF 123 G 10 2020 BTool 4: Insert – CNMG 12008 WM 4015,Holder – DC LNR 2020 K 12

Table 1 : Design of Experiment

Exper-iment

No.

Speed (rpm)

Feed (mm/rev)

Depth of Cut (mm)

Tool

(A) (B) (C) (D)

1 0 0 0 0

2 0 1 1 2

3 0 2 2 3

4 0 3 3 4

5 0 4 4 1

6 1 0 1 1

7 1 1 2 4

8 1 2 4 0

9 1 3 0 3

10 1 4 3 2

11 2 0 2 2

12 2 1 4 3

13 2 2 3 1

14 2 3 1 0

15 2 4 0 4

16 3 0 3 3

17 3 1 0 1

18 3 2 1 4

19 3 3 4 2

20 3 4 2 0

21 4 0 4 4

22 4 1 3 0

23 4 2 0 2

24 4 3 2 1

25 4 4 1 3

The outcome of ANN for L25 orthogonal array is shown in Table below

Table 2 : ANN Output Table

Exper

-iment No.

Dia measured actually using digital vernier (mm)

(Y)

1 10.130

2 10.030

3 10.060

4 10.010

5 9.760


6 9.950

7 10.050

8 9.940

9 9.920

10 9.970

11 10.040

12 9.980

13 9.980

14 10.080

15 10.080

16 9.930

17 10.050

18 10.130

19 9.980

20 10.020

21 10.110

22 10.030

23 9.970

24 9.970

25 9.920

For L25 Orthogonal Array –For Factor A :A0 = Σ observations (where one sets factor A at 0)

= (Y1 + Y2 + Y3 + Y4 + Y5 )NA0 = No. of observations with factor A set at 0.

= 5Abar0 = Average of observation with factor A set at 0.

= AO/NA0

= ( 10.130 + 10.030 + 10.060 + 10.010 + 9.760 ) / 5

= 9.998 mm

A1 = Σ observations (where one sets factor A at 1)



= A1/NA1

= ( 9.950 + 10.050 + 9.940 + 9.920 + 9.970 ) / 5

= 9.966 mm

A2 = Σ observations (where one sets factor A at 2)= (Y11 + Y12 + Y13 + Y14 + Y15 )NA2 = No. of observations with factor A setat2.= 5Abar2 = Average of observation with factor A set at 2.

= A2/NA2

= ( 10.040 + 9.980 + 9.980 + 10.080 + 10.080 ) / 5

= 10.032 mm




= A3/NA3

= ( 9.930 + 10.050 + 10.130 + 9.980 + 10.020 ) / 5

= 10.022 mm




= A4/NA4

= ( 10.110 + 10.030 + 9.970 + 9.970 + 9.920 ) / 5

= 10.00 mm

For Factor B :

B0 = Σ observations (where one sets factor B at 0)

= (Y1 + Y6 + Y11 + Y16 + Y21 )NB0 = No. of observations with factor B set at 0.

= 5Bbar0 = Average of observation with factor B set at 0.

= B0/NB0


= ( 10.130 + 9.950 + 10.040 + 9.930 + 10.110 ) / 5

= 10.032 mm



= 5Bbar1 = Average of observation with factor B set at 1. = B1/NB1

= ( 10.030 + 10.050 + 9.980 + 10.050 +10.030 ) / 5= 10.028 mm




= B2/NB2

= ( 10.060 + 9.940 + 9.980 + 10.130 + 10.030 ) / 5

= 10.028 mm




= B3/NB3

= ( 10.010 + 9.920 + 10.180 + 9.980 + 9.970 ) / 5

= 9.992 mm




= B4/NB4

= ( 9.760 + 9.970 + 10.080 + 10.020 + 9.920 ) / 5

= 9.95 mm

For Factor C :

C0 = Σ observations (where one sets factor C at 0)

= (Y1 + Y9 + Y15 + Y17 + Y23 )NC0 = No. of observations with factor C set at 0.

= 5Cbar0 = Average of observation with factor C set at 0.

= CO/NC0

= ( 10.130 + 9.920 + 10.080 + 10.050 + 9.970 ) / 5

= 10.03 mm



= 5Cbar1 = Average of observation with factor C set at 1. = C1/NC1

= ( 10.030 + 9.950 + 10.080 + 10.130 + 9.920 ) / 5

= 10.022 mm




= C2/NC2

= ( 10.060 + 10.050 + 10.040 + 10.020 + 9.970 ) / 5

= 10.028 mm




= C3/NC3

= ( 10.010 + 9.970 + 9.980 + 9.930 + 10.030 ) / 5

= 9.984 mm




= C4/NC4


= ( 9.760 + 9.940 + 9.980 + 9.980 + 10.110 ) / 5

= 9.954 mm

For Factor D :

D0 = Σ observations (where one sets factor D at 0)

= (Y1 + Y8 + Y14 + Y20 + Y22 )ND0 = No. of observations with factor D set at 0.

= 5Dbar0 = Average of observation with factor D set at 0.

= DO/ND0

= ( 10.130 + 9.940 + 10.080 + 10.020 + 10.030 ) / 5= 10.04 mm

D1 = Σ observations (where one sets factor D at 1) = (Y5 + Y6 + Y13 + Y17 + Y24 )

ND1 = No. of observations with factor D set at 1.= 5

Dbar1 = Average of observation with factor D set at 1. = D1/ND1

= ( 9.760 + 9.950 + 9.980 + 10.050 + 9.970 ) / 5

= 9.942 mm




= D2/ND2

= ( 10.030 + 9.970 + 10.040 + 9.980 + 9.970 ) / 5

= 9.998 mm




= D3/ND3

= ( 10.060 + 9.920 + 9.980 + 9.930 + 9.920 ) / 5

= 9.962 mm




= D4/ND4

= ( 10.010 + 10.050 + 10.080 + 10.130 + 10.110 ) / 5

= 10.076 mmAverage Effect A =

Abar2 – Abar1 = 10.032 – 9.966 = 0.066

Average Effect B =

Bbar0 – Bbar4 = 10.032 – 9.950 = 0.082

Average Effect C =

Cbar0 – Cbar4 = 10.030 – 9.954 = 0.076

Average Effect D = Dbar4 – Dbar1 = 10.076 – 9.942 = 0.134

The “two factor interactions” are calculated as follows –Let,A1C1 = Sum of observations with factor A set at 1 and factor C set at 1

= Y6 = 9.950

A1C2 = Sum of observations with factor A set at 1 and factor C set at 2= Y7 = 10.050



Interaction AC = [ (A1C1 + A2C2) – (A1C2 + A2C1) ] / 4 = [ (9.950 + 10.040) – (10.050 + 10.080) ] / 4= ( 19.99 – 20.13 ) / 4= ( – 0.14 ) / 4= – 0.035

To carry out ANOVA[3] of observations the sum of squares of certain deviations are required. This can be determined as follows–T = Σ of all observations


= ( Y1 + Y2 + …… + Y25 )= 250.09

The correlation factor (CF)[4] is defined as –

CF = T2 / N (N–Total no. of observations)= (250.09) / 25

= 10.0036

N

Total sum of squares (ST) = Σ[Yi]2–10.0036 i=1

= (250.09)2–10.0036 = 62535.00

The factor sum of squares –

SA = [A1]2 / NA1 + [A2]2 / NA2 – CF = (49.83)2 / 5 + (50.16)2 / 5 – 10.0036 = (2483.02/5) + (2516.02/5) – 10.0036 = 496.60 + 503.20 – 10.0036 = 989.79

SB = [B1]2 / NB1 + [B2]2 / NB2 – CF = (50.14)2 / 5 + (50.14)2 / 5 – 10.0036 = (2514 / 5) + (2514 / 5) – 10.0036

= 502.80 + 502.80 – 10.0036 = 1005.60 – 10.0036 = 995.59

SC = [C1]2 / NC1 + [C2]2 / NC2 – CF = (50.11)2 / 5 + (50.14)2 / 5 – 10.0036 = (2511.01 / 5) + (2514.01 / 5) – 10.0036 = 502.20 + 502.80 – 10.0036 = 994.99

SD = [D1]2 / ND1 + [D2]2 / NDC2 – CF = (49.71)2 / 5 + (49.99)2 / 5 – 10.0036 = (2471.08 / 5) + (2499.00 / 5) – 10.0036 = 494.21 + 499.80 – 10.0036 = 984.00

And, the interaction sum of squares are –

SAxC = [A1C1+A2C2]2 / (NA1C1+NA2C2) + [A1C1+A2C2]2 / (NA2C1+NA1C2) – CF= (9.950+10.040)2 / (1+1) + (9.950 +10.040)2 / (1+1) – 10.0036= (19.99)2/2 + (19.99)2/2 – 10.0036= 199.80 + 199.80 – 10.0036

SAxC = 389.59

Similarly,

B1C1 – Y2 = 10.030

B1C2 – Y7 = 10.050

B2C1 – Y18 = 10.130

B2C2 – Y3 = 10.060

SBxC = [B1C1+B2C2]2 / (NB1C1+NB2C2) + [B1C1+B2C2]2 / (NB2C1+NB1C2) – CF= (10.030+10.060)2 / (1+1) + (10.030+10.060)2/ (1+1) – 10.0036= (20.09)2/2 + (20.09)2/2 – 10.0036= 201.80 + 201.80 – 10.0036= 403.60 – 10.0036

SBxC = 393.60

We can find the sum of Squares for error (Se) as follows –Se = ST – SA – SB – SC – SD – SAxC – SBxC

= 62535 – 999.80 – 995.59 – 994.99 – 984 – 389.59 – 393.60

= 57,777.43

We then determine the respective dof. The total dof

it is given by –

ƒT = Total no. of observation – 1= 25 – 1 = 24

The other dof are as follows –

ƒA = No. of distinct levels of A – 1= 5 – 1 = 4

ƒB = No. of distinct levels of B – 1= 5 – 1 = 4

ƒC = No. of distinct levels of C – 1= 5 – 1 = 4

ƒD = No. of distinct levels of D – 1= 5 – 1 = 4

ƒAxC = ƒA x ƒC = 4 x 4 = 16

ƒBxC = ƒB x ƒC = 4 x 4 = 16

The dof for error (the influence of uncontrolled factors influencing the response) may be found as –Ferror = ƒT – (ƒA+ƒB+ƒC+ƒD+ƒAxC+ƒBxC)

= 24 – ( 4 + 4 + 4 + 4 + 16 + 16 )= – 24

The Mean Sums of Squares[5] are given by formula –


Mean SSA = SA / ƒA

= {[A1]2/NA1 +[A2]2/NA1–CF}/ƒA

= (989.79) / 4= 247.45

Mean SSAxC = SAxC / dofAxC

= (389.59) / 16= 24.34

Mean SSA 249.95F-Statistic = --------------- = ------- = - 0.10

Mean SSerror –24

F-SSerror = ---------- = ---------- = -2,407

ƒerror – 24

Since F, F calculated = 0.10 which is less than F test. Critical value for given dof, it concludes that the effect of one treatment is not equal to other treatment and the interaction effect is significant.

Cross Validation using Image Processing Technique[2]:As an alternative approach to the earlier methodology cross validation of dimensional measurement is carried out by using image processing technique.

For this, image is grabbed by using VGA camera for the mounting of the same alongwith image grabbing scheme and job fixture. Refer Fig. Nos. 2, 3 & 4.Edge detection is carried out by using appropriate algorithm and writing down code in

Matlab 7.3. The outcome is shown in Fig. Nos. 5 & 6.Another programme is written in Matlab 7.3 which is a m file that gives diameter of this image using two point diameter form equation of the circle. The results of the above are tabulated in Table 3.Thus cross validation of the experimentation is carried out using image processing technique also.

IMAGING SCHEME :Good quality image is the basic input for further image manipulation. Hence a set up is devised for grabbing image. It consists of a fixture to firmly locate the component specimen. This fixture in addition to locating the component also provides fine movement of the component such that the plane of location can be ensured exactly horizontal

by a spirit level. This fine adjustment is provided by 4 adjustment screws situated on the bottom side of the fixture.

CAMERA MOUNTING FIXTURE –

For high quality imaging, it is also necessary that the camera should be mounted or clamped in a stand. This ensures movement of camera from object specimen as per requirement and facilitating effective focusing with the said objective.

Camera mounting stand consists of parts as : a) C clamp, b) Vertical guide bar, c) Horizontal bar, d) Bar holding hinge, e) Camera Mounting fixture, etc.

C clamp is used to firmly grip the entire assembly rigidly to only platform, providing necessary support to this structure. It can be clamped and unclamped using the clamping spindle provided on the bottom side.

Fig. No. 2 : Camera Mounting Fixture

Fig.No.3: Illumination Adjustment Stand

A guide bar is a vertical cylindrical high finish bar mounted on top of the said C clamp to facilitate horizontal movement of camera along X plane (horizontal plate) to centre the component under consideration.Bar holding hinge, this component facilitates clamping of camera mounting fixture to horizontal


bar. This hinge also provides rolling and pitching movement of camera to fine-tune it for effective imaging. All bar sections are chrome plated and have good finish for effective movement.

Camera mounting fixture, as shown in Fig. No. 2, holds camera firmly providing it required stability for imaging purpose.

This entire assembly consisting of above components provides desired camera orientation in any direction for better imaging purpose.

BACKLIGHT ILLUMINATION –

After a lot of experimentation a backlight scheme of illumination is proposed. This arrangement ensures adequate light on the surface of component specification under consideration. It has been

observed that, diffused lighting scheme with fluorescent tube provide desired illumination for

To suit above requirement, Illumination Adjustment Stand is devised as shown in Fig No. 3

FRAME GRABBER HARDWARE –

Supporting frame grabber hardware is used to support image display on PC with following specification.

Odyssey PCL TV tuner card supporting still images in bmp, jpeg, tiff formats. effective image grabbing, with required contrast level.

Table 3 : L25 Experimentation

No. Image No. Extremity Readings

EstimatedDia (mm)

% error = (estimated diameter – actual diameter)

÷ actual diameter

1 125_16_10130 X378Y325

X858Y325

1236.00 18.04

2 125_17_10030 X340Y348

X823Y348

1163.00 13.75

3 125_18_10060 X304Y340

X781Y340

1085.00 7.28

4 125_19_10010 X327Y408

X816Y408

1143.00 12.42

5 125_20_9760 X318Y301

X783Y301

1101.00 11.36

6 125_21_9950 X289Y318

X757Y318

1046.00 4.87

7 125_22_10050 X289Y307

X757Y307

1046.00 3.91

8 125_23_9940 X292Y306

X751Y306

1043.00 4.69

9 125_24_9920 X333Y369

X786Y369

1119.00 11.34

10 125_25_9970 X312Y310

X795Y310

1107.00 10.38

11 125_26_10040 X306Y432

X789Y432

1092.00 8.05

12 125_27_9980 X339Y465

X814Y465

1158.00 13.81

13 125_28_9980 X357 X831 1188.00 15.99


Y430 Y430

14 125_29_10080 X307Y472

X783Y472

1090.00 7.52

15 125_30_10080 X391Y516

X864Y516

1255.00 19.68

16 125_31_9930 X342Y508

X810Y508

1152.00 13.80

17 125_32_10050 X342Y475

X816Y475

1158.00 13.21

18 125_33_10130 X370Y474

X852Y474

1222.00 17.10

19 125_34_9980 X390Y304

X750Y304

1040.00 4.03

20 125_35_10020 X343Y364

X828Y364

1171.00 14.43

21 125_36_10110 X378Y484

X857Y484

1235.00 18.13

22 125_37_10030 X315Y514

X792Y514

1107.00 9.39

23 125_38_9970 X312Y421

X783Y421

1095.00 8.94

24 125_39_9970 X390Y478

X868Y478

1258.00 1178

25 125_40_9920 X366Y502

X831Y502

1197.00 17.12

% Avg. error = 5.28 %

Average of experimentation method = 5.28 %

From above calculation it can be concluded that the % error between actual diameter of pin component using OA technique implemented through L16 and L25 orthogonal array experimentation and estimated diameter derived by grabbing the image,

carrying out edge detection algorithm very with 6.62%.Thus the experimentation methodology is validated using image processing technique.


Fig. No. 4 : Original Grabbed Image

Fig. No. 5 : Original Image after Edge Detection


Fig. No. 6 : Grey Scale Image after Edge Detection

CONCLUSION :

I. Traditional ‘Vary one factor at a time’ based experimentation is incapable of uncovering interaction among factors.

II. Statistically designed experiments constitute sound and scientific approach to study such phenomena and carry out experimentation.

III. Since the calculated value of F is less than the critical value for the degree of freedom under consideration. It can be concluded that significant interaction is existing between the parameters, speed, depth of cut, feed and tool material.

IV. Such evaluation enables understanding of variation within the parameters and between the parameters under consideration.

V. From above experimentation, it can be observed that for a component under consideration, various alternatives can be proposed for finding out some key geometric parameter like diameter. Effect and significance of each parameter that contribute to the manufacturing process, like tunning can be studied by using F-Test Technique and also cross validation can be

carried out using image processing technique as described in the present work.

REFERENCES :

1. Tapan B. Bagchi, “Taguchi Techniques”, (ISBN No. 0-87692-808-4), Prentice Hall Publication, page 79-90,.

2. Digital Image Processing using MATLAB by Rafeel C. Gonzalez, Richard E.Woods and Steven L.Eddins, Peerson Education, 2nd Edition 2005, ISBN 81-297-0515-X, pp. 379-400.

3. Juran J.M. and Gryna F.M. (1988), Juran’s Quality Control Handbook, 4th Ed., McGraw Hill Pub., pp 100-150.

4. Wadsworth H.M. (1989), Handbook of statistical Methods for Engieners & Scientists, McGraw Hill Publications, pp 150-170.

5. “Applied Statistics & Probability for Engineers”, Douglas C. Montagomery, George C. Renger, John Wiley & Sons Pub. ISBN 9812-13-058-4, pp.600-625.


RECENT ADVANCES IN KNOWLEDGE DISCOVERY DATABASE AND FUZZY LOGIC.

M.SURESH BABU,MCA, M.Phil, PGDBA, MISTE, Principal, INTEL Institute of Science, Anantapur – 515004. Email :

[email protected] S.SURESH BABU,

M.Tech,MISTE, Assistant Professor, NBKR Institute of Technology, Vidyanagar, Nellore Dist, Andhra Pradesh. Email: [email protected].

K.N.SUDEEPTHI, MCA, MISTE,Assistant Professor, INTEL Institute of Science, Anantapur – 515004.

Email : [email protected] are many activities of both technology and its application in data mining and knowledge discovery in database area. In this paper, actual applications of data mining will be presented from its development point of view, and a generalized conceptual model of data mining will be proposed. Additionally, to make the data mining process more applicable, important notes at the performing process against real world problems will be discussed.

Keywords: fuzzy logic, fuzzy database, data mining, knowledge discovery in database

1 INTRODUCTION

The more database systems are developed, the more strategic use of them is necessary. The more needs to use database efficiently are increased, the more applicable technologies are developed This positive and evolutionary cycle is now occurring in the area named data mining or knowledge discovery in database. Informative knowledge discovering and new valuable data finding in database are very attractive in various business scene. In this paper, as these examples, real world applications of data mining will be presented from its development point of view, and a generalized conceptual model of data mining and its process will be proposed. Additional notes on the process will be discussed and some useful solutions using fuzzy logic will also be proposed.2 Data Mining ApplicationsVarious sorts of knowledge-based systems have been developed as real world applications. In their development stage, some kinds of data mining and knowledge discovery technologies, not explicitly recognized yet, have been used. In this chapter, we give attention to the development process of applications and pick up two representative actual applications. The knowledge used in them are represented by

Case 1: A direct mailing system

This system is for obtaining data set of high potential customers who will buy more products at a certain sale such as a winter sale.

Their motivation to buy is expected to be triggeredwith advertisement mails directly sent to them. The purpose of using this system is to increase sales or profit. In general, there are few cases of aiming just sales because a mailing budget is previously fixed and the number of mails is controlled to increase profit. By using fuzzy retrieval technology, a retrieval result from customer database is ranked and sorted by the matching degree values for a given fuzzy query condition. The list of retrieved results makes it easy to select high potential customers data set because the decision making of data selection can be done by both quantity and quality aspects according to the matching &gm value. However, the problem here is the method to get an appropriate fuzzy query condition for obtaining the data of good customers. We can get the condition through the following data mining process:Clarifying the purpose of advertisement mailing for a certain sale; that decides the balance between sales and profit. Collecting the customers data who bought products from the shop by using the customer database and the yearly sales database of two years ago and then plotting the data onto a visualized graph screen. The graphs are two dimensions.The examples of the attributes are:





“frequency of visiting the shop”, “sum of shopping”, “date of the latest visit”, and “term of being memberships”. Analyzing the dat

the graph to find the clusters which includes good customers’ data.Discovering the attributes relationships to decide the good clusters to define a query condition. This query condition is represented by the pairs of attributes and fuzzy linguistic values, and the Connectives for the pairs with weighting values for each pair. Verifying the query conditions by using the customer database and the yearly sales database of one year ago to refine the conditions not to omit the good customers.Case 2: A vibration sensorThis product is for recognizing what causes specific vibrations; e.g., by an earthquake or others. The problem here is to improve the accuracy of sense result. A conventional sensor was designed using very simple rule such as threshold level setting on one or two dimensions. Because of the simple sensing mechanism, the conventional version often gives incorrect output; e.g., when vibrations are caused by automobile or human, it gives output as a earthquake vibration. The error rate by the conventional one is 12%. However, it is improved to 1% by using fuzzy rules which is founded by the data mining process as follows: Clarifying the performance goal of the

sensor, e.g., focusing on recognizing difference of earthquake vibration from others.

Collecting actual data in the specified cases of earthquake, automobile, human, etc., then visualizing them by a time series graph.

Analyzing them to find the feature parameters that can classify each of the data.

Finding the relationship between the feature parameters and the vibration sorts to build up fuzzy rules. Testing, evaluating and refining the fuzzy rules by using the actual data.

3 A Conceptual Model of Data Mining

Many useful studies have been one in data mining and knowledge discovery in database. A technical overview and a process description of knowledge discovery in database are useful to understand this area comprehensively.

By basing on the concept that features the process aspects of data mining, we gives attention to the interaction between a human and a machine and the purpose clarification.

In this chapter, a conceptual model of data mining is proposed by generalizing the actual application development process.

Data mining is the process which extracts knowledge from real world environment according to a certain purpose. In this process, both top-down and bottom-up approaches are performed as problem solving methods. The top-down approach clarifies purpose, defines problems to be solved, then breaks down the problems into elements until solvable level. On the other hand, the bottom-up approach collects data from the real world, analyzes them, then integrates the findings. Both approaches are combined into data mining to find solvable goal, to select a suitable method for the goal and then to develop knowledge based on the method.


The data mining process is shown in Figure 2. The steps below are the generalized data mining process. Before applying the process, we should define the benefits of developing target applications clearly to give the purpose.Purpose Clarification: Clarifying the purpose, the problems to be solved, and the hypothetical goal of solution through the top-down approach.

Data Collection: Collecting data from the real world and visualizing them through the bottom-up approach. Data Analysis: Analyzing the data collection to verify the hypothetical goal of solution through the combination of the top-down and the bottom-up approach.Knowledge Development: Selecting a suitable method for the goal and developing knowledge based on the method.Knowledge Refinement: Testing and refining the knowledge. If necessary, back to the previous steps.4 For More Applicable Data MiningIn order to make the data mining process to be more applicable for a real world problem, two important viewpoints are described in this chapter. When we see the problem solving technique such as data mining, many of problems are originally complex and not easily solved. Therefore, the previously explained data mining process which describes a useful general procedure is not always completely effective by itself. Thus, we should add following two kinds of points which are useful to make efficient use of the data mining process in the real world.4.1 Importance of Data CollectionReliability of data representing the real world is important to get inductive knowledge. However, using data includes problems, practically; e.g., 1) no data, 2) few data, 3)redundant data, 4) incorrect data, 5) uneven distributed &a, etc. To solve these problems, considering the purpose and the hypothetical goal, it is necessary to collect data of both quality and quantity carefully. In the case of supervised machine learning, the class distribution of data for

learning should be considered Dividing original data into for learning and testing is also important. Besides on these notes, some possible solutions are thinkable by using fuzzy logic; i.e., data cleaning based on similarity retrieval. The data cleaning is a solution for the problems regarding redundant data An example for data redundant is that a customer database system accumulates plural data of the same customers caused by miss-typing, including middle name or not, etc. The first step of data cleaning procedure is to retrieve data which has similar attribute values, specifying a candidate of redundant data for similarity retrieval. Similarity degree values are calculated by using fuzzified data of the candidate. The next is to eliminate the redundant data from the retrieved data set.

Because each retrieved data is ranked by similarity, a human can easily select the redundant data which should be eliminated.

4.2 Human-Machine Skill IntegrationBecause a human and a machine have their own strong skill for each, integrating their skills effectively will exploit the efficiency of the data mining process. The examples of the strong skills of a human are: recognizing two or three dimensions of information, processing deep knowledge, etc. It is easy for a human to recognize objects which are existing in a photograph or to understand meanings from just short conversation with using their background knowledge. On the other hand, machine ability has excellent to handle large quantity, many interaction, and high performance information processing. Regarding information representation, linguistic information for a human and numerical for a machine will the appropriate grouping. The skill integration is put into practice by the interaction between a human and a machine in the data mining process. The importance here is to put the right skill in the right post. Fuzzy technology can support the cooperative interaction because it has capability of transferring linguistic and numerical data into each other. For instance,


although human can extract knowledge from complex data set, it is hard to write down his imprecise, incomplete, and ambiguous knowledge with numerical representative information. In the case of explanation of knowledge discovered by a machine, it is hard for a human to understand the knowledge which is extracted by a neural network although the neural network has strong learning ability to build up the model from sample data. By using fuzzy logic, its linguistic representative power can easily explain the knowledge to a human. It is useful to extract knowledge via data from the real world and to represent it in practical use form.5 Summary and ConclusionData mining process is discussed and a conceptual model of data mining is proposed. This model features that the purpose clarification and the human-machine interaction

are taken on importance. Machine learning algorithms are important and really useful to help discovering knowledge. However, the keys to perform data

mining in the real world successfully are the purpose and the interaction. For integration of the top-down and bottom-up approach, the feature of the fuzzy logic that can process imprecise information is really efficient. In the interaction between human and machine skills in the data mining process, linguistic representation power of fuzzy logic makes it easy to write down the knowledge into efficient fuzzy rules.


REFERENCES1.H. Nakajima, T. Sogoh, and M. Arao, “Fuzzy Database Language and Library : Fuzzy Extension to SQL”, Proceedings of 2nd IEEE International Conference on Fuzzy Systems, (1993), pp477-482.2. U.M.Fayyad, G.Piatetsky-Shapim, and P. Smyth, “From Data Mining to Knowledge Discovery:An Overview”, Advances in Knowledge Discovery and Data Mining, AAA1 Press/ The MIT Press, (1996), ppl-34.3. R.J.Brachman andT.Anand, ”The Process of KnowledgeDiscovery in Database”, Advances in KnowledgeDiscovery andDataMining, AAAIPress/ TheMIT Press, H. Nakajima and Y. Senoh, ‘Development of Fuzzy Add-in Macros with Spread Sheet”, Proceedings of the FUZZ-IEEEAFES’95 Workshop on Fuzzy Database Systems, (1995), pp55-60.4..Krantz, D.H., Luce, R.D., Suppes, P., Tversky, A. (1971, 1989, 1990), Foundations of measurement, Vol. 1,2,3 - NY, London: Acad. press, (1971) 577 p., (1989) 493 p., (1990) 356 p.5. Kovalerchuk, B., Vityaev, E. (2000), Data Mining in finance: Advances in Relational and Hybrid Methods, Kluwer Academic Publishers, 308 p.6. Kovalerchuk, B., Vityaev, E., Ruiz, J.F. (2001), ‘Consistent and Complete Data and "Expert" Mining in Medicine’. In: Medical Data Mining and Knowledge Discovery, Springer, pp.238-280.7. Scientific Discovery website http://www.math.nsc.ru/AP/ScientificDiscovery8. Zighed, DA (1996): SIPINA-W, http://eric.univ-lyon2.fr/~ricco/sipina.html, Université Lumière,Lyon.9. Michie, D., Bain, M., Hayes-Michie, J. Cognitive models from subcognitive skills. In: Grimble, M., McGhee, J., Mowforth, P. (eds .) Knowledge- Based System in Industrial Control, Stevenage: Peter Peregrinus, 1990.10. Quinlan, J. R. Combining instance-based and model-based learning. Proc. Tenth Itit. Cot$ Machine Learning, pp. 236-243, Aniherst, Massachusetts. Morgan Kaufniann 1993.About Authors

About Authors:

1. M.Suresh Babu, born on 27th December 1970 obtained his MCA from Osmania University and M.Phil Degree in Computer Science from Bharathiar Univeristy. At Present he is doing research in Computer Science on Inductive Data Mining.

He served Intel Institute of Science, Anantapur in various capacities like Lecturer, Head of the Department and at present is working as Principal. He has organized several workshops and training Programmes in the field of Computer Science and attended a number of Workshops and seminars. He is a life member of ISTE and Science & Society & ISCA. He is the State Convenor for State Education Sub Committee, Jana Vignana Vedika.

2. K.N. Sudeepthi, obtained her MCA from Sri Krishnadevaraya University, Anantapur. She is working as Assistant Professor in Department of Computer Applications, INTEL Institute of Science. She is a life member of Indian society for technical education and Indian society for computer applications. Her area of interest are Datamining,Artificial Intelligence, Cryptography and network security.

3. S.Suresh Babu, has obtained B.Tech (ECE) from JNTU, Kakinada and Master of Technology in Computer Science Engineering from Bharath University, Chennai. He has registered for PhD in 2009. He is working as Asst Professor

ENGINEERING TODAY QUARTERLY JOURNAL 69 DECEMBER 2010 VOLUME II ISSUE 4 ISSN :2180-0995

in Department of Computer Science & ngineering at NBKR Institute of Science Technology, Vidyanagar, Nellore Dist . He has rich and varied experience in teaching since 1999. His area of interest are Data mining, Data Communications, Artificial Intelligence, Cryptography, Network Security, Programming Languages.

EFFECT OF LPG CONTENT ON THE PERFORMANCE AND EMISSION

CHARACTERISTICS OF A DIESEL-LPG DUAL-FUEL ENGINE

G.A.RAO, A.V.S.RAJU, C.V.MOHAN RAO, K.GOVINDA RAJULU

1 Dr.Paul Raj Engineering College, 2 Professor of Mechanical Engineering, BHADRACHALAM, Dist.Khammam JNTUniversity,Kukatpally,Pin507111,IndiaMobileNumber09849524535Hyderabad500085,IndiaEmailid:[email protected] University College of Engineering,4 Professor of Mechanical EngineeringKOTHAGUDEM,

Dist.Khammam,JNT University, Anantapur. India.

ABSTRACT -:

In the present work, LPG, a by-product of petroleum refining process is used to replace conventional diesel fuel, partially, for improved combustion efficiency and clean burning. A conventional diesel engine was operated on the dual-fuel mode, using LPG as the primary fuel and diesel as the pilot fuel. A four-stroke, single-cylinder diesel engine, most widely used in agricultural sector, has been considered for the purpose of experimentation. The engine was operated at a constant speed of 1500 rpm at a low engine load of 20% and a high engine load of 80%.

Under both these operating conditions, combustion, performance and emission characteristics of the engine have been evaluated and compared with that of baseline diesel fuel operation. It is found that at 20% engine load, diesel fuel operation was superior to the dual-fuel operation, whereas at 80% engine load, the reverse was true.

Key Words:

Dual-Fuel, LPG, Diesel, Combustion, Performance, Emissions and Load.

1. INTRODUCTIONThe purpose of internal combustion engines is the production of mechanical power from the chemical energy contained in the fuel. In internal combustion engines, as distinct from external combustion engines, this energy is released by burning the fuel inside the engine cylinder. The fuel-air mixture before combustion and the burned products after combustion are the actual working fluids. The work transfers which provide the desired power output

occur directly between these working fluids and the mechanical components of the engine. Because of their simplicity, ruggedness and high power-to-weight ratio, the internal combustion engines have found wide applications in transportation and power generation. In spite of many unconventional sources of energy developed, petroleum based fuels remain the primary source of energy in the field of power and propulsion all over the world. The demand for


the petroleum products in India is increasing at an alarming rate.For instance the demand for the petroleum products was about 3.5 Million Metric Tons (MMT) in 1950-51, and in 1997-98 it has risen to about 84.3 MMT, in 2004-05 it has reached an alarming value of 111 MMT(1). At this rate it is believed that the demand may reach a staggering 234 MMT by 2019-20. The approximate petroleum reservoirs in the world are 700 MMT. The domestic production of crude oil is about 33 MMT and the

diesel consumption is 40 MMT, the petroleum oils worth Rs.102500 crores were imported (1).Conventional fuels for internal combustion engines are getting dwindled at an alarming rate, primarily due to exponential rise in the population of automobile users world-wide. These fuels are likely to prevail for another 50-75 years unless newer reservoirs are explored. Further, these fuels, upon their combustion release toxic gases into the atmosphere, causing severe environmental

pollution and degradation of the quality of air of the atmosphere.Dual-fuel operation is found to be one of the attractive ways of conserving precious conventional fuels, diesel and petrol. In this mode of operation, two fuels would be used, normally a gaseous fuel and a liquid fuel.

Ghazi A.Karim (2), has underlined the importance of understanding the combustion processes in dual-fuel engines with regard to enhanced engine performance and reduced air pollution. In this article the importance of various gaseous fuels with regard to their availability, performance, cost aspects and pollution aspects is discussed.

Dong Jian (3) and et al, have developed have developed a new type of duel supply system, which could able to economically convert conventional diesel engines into dual-fuel engines like LPG/Diesel engines and CNG/Diesel engines. These are capable of using either single diesel fuel or dual-fuel including both diesel and LPG and diesel and CNG. These diesel-LPG engines have been applied to the diesel buses in the public transportation of Guangzhou city, one of the biggest cities in China. Compared with the diesel baseline engine, it was found that there were significant reduction in soot emission and an improvement in the fuel consumption with the diesel-LPG engine. Also the strategy on LPG content is discussed in order to meet the demands for soot emission, fuel economy, transient performance and the output power at the same time.

Poonia (4) and et al, have investigated experimentally the factors that affect the performance of a LPG-Diesel dual-fuel engine. In this work, the effects of intake charge temperature, pilot fuel quality, exhaust gas recirculation and throttling of the intake on the performance of a LPG-diesel dual fuel engine have been studied. It is found that at

low outputs an increase in intake temperature and pilot quality is advantageous as HC levels are reduced.

Sahir Salman (5) and et al, have investigated the reduction in exhaust gas emissions from a diesel engine under dual-fuel operation. For this purpose, a single-cylinder, direct injection diesel engine was modified to operate with dual-fuel operation (30% LPG and 70% diesel fuel by weight). During the experiments, the engine speed was kept constant (1650 rpm) and the load was changed.

It was observed that the NOx and smoke emissions were reduced with the dual fuel operation. Further, it is found that the fuel property is one of the most important parameters, which effects the exhaust emissions.

Srinivasa Rao (6) and et al, have made experimental investigations on single-cylinder vertical water cooled compression ignition engine run on the dual-fuel mode with diesel as the pilot fuel and LPG as the main fuel. The engine is run under different

operating conditions and in each case the optimum combination of the inducted to inject fuel energy proportions are determined for the best efficiency.

Amarendar Rao (7) and et al, have performed experimental investigation of a single-cylinder, 4-stroke diesel engine operating on the dual-fuel mode using LPG as the main fuel and diesel as the pilot fuel. The experiments have been carried out at a constant speed of 1500 rpm under varying load conditions. The results are compared with those of pure diesel operation. They have indicated that the by dual-fuel operation precious diesel could be conserved even up to 80%, however in their work it could be done only up to 45% due to severe engine vibrations.


EXPERIMENTAL SET-UP:

The experimental setup consists of a single-cylinder, four-stroke diesel engine connected to an eddy-current dynamometer for loading of the engine. It was provided with necessary instruments for combustion pressure and crank angle measurements. The signals were interfaced to a computer through an engine indicator to obtain pressure-crank angle and pressure-volume diagrams. Provision was also made for interfacing air flow, fuel flow, temperatures and load measurement. The brake power of the engine was

measured by coupling the engine to an eddy-current dynamometer. It consists of a stator, on which number of electromagnets was fitted, and a rotor disc, and coupled to the output shaft of the engine. When the rotor rotated eddy-currents are produced in the stator due to magnetic flux set up the passage of the field current in the electromagnets. These eddy-currents oppose the rotor motion, thus loading the engine. These eddy-currents dissipate lot of heat, thus proper cooling is required for the dynamometer. A moment arm measures the torque; the load on the engine was controlled by regulating the current to the electromagnets.

The engine set up has stand-alone panel box, consisting of an air box, a fuel tank, a manometer, a fuel measuring unit, transmitters for air and fuel flow measurements, a process indicator and an engine indicator. Rotameters are provided for cooling water and calorimeter water flow measurement. An NDIR AVL exhaust gas analyzer was used to measure the emissions of CO,

HC and NOx. A smoke meter was

employed to record the smoke intensity of the exhaust gases. The schematic layout of the experimental set up is shown in figure1. The specifications of the engine are depicted in table 1. In this work, LPG has been used as the primary fuel and diesel as the pilot fuel. The role of the pilot fuel is to initiate the combustion, while that of the primary fuel is to generate mechanical power.

RESULTS AND DISCUSSION

A conventional diesel engine could be successfully converted to run on LPG-diesel dual-fuel operation. The influence of addition of LPG on the engine performance and emissions has been evaluated. Initially, baseline tests have been performed using diesel as the fuel. The tests have been conducted at a constant speed of 1500 rpm at two distinct loads of 20% and 80% of full load. Then, the tests have been performed on the engine at the same speed under the same operating conditions on dual-fuel mode. The content of LPG was gradually increased and diesel supply was reduced without affecting the engine power. At 20% engine load, 50% of diesel consumption could be replaced by LPG, but at 80% load, it could be done only up to 20%. This was due to engine operating difficulties, like undue vibrations of the engine parts, and excessive heating.The effect of LPG content by energy on brake thermal efficiency of the engine at low (20%) and high engine loads is depicted in figure 2. From this figure it is

evident that at low engine load, brake thermal efficiency has decreased with an increase in the LPG content. With zero LPG content (diesel fuel operation) it was found to be 14.4% and with 50% LPG (dual-fuel operation), it was reduced to 9.7%, a decrease of 32.6%. At 80% engine load, the brake thermal efficiency has increased with an increase in the LPG content.

Figure 3 represents the effect of LPG content on the mechanical efficiency of the engine at low and high loads of the engine. At 20% engine load, mechanical efficiency of the engine has reduced with an increase in LPG content. It has reduced from 33.3% with zero LPG to 28.3% with 50% LPG content, a drop of 15.01%. At 80% engine load it has increased with an increase in LPG content. On diesel fuel operation it was 69.2% and on dual-fuel operation it has increased to 70.3%.The volumetric efficiency of the engine has reduced with an increase in LPG content at both the loads, as shown in figure 4. At 20% engine load, it was 77.8% on diesel fuel operation and reduced to 73.8% with dual-fuel operation. This is depicted in figure 4. Figure 5 represents the effect of LPG content on the brake specific fuel consumption of the engine. It is evident that, as the LPG content is increased, it has increased at 20% load, making the engine operation more costly. However, at 80% load, bsfc has decreased with an increase in the LPG content.With regards to emissions, smoke density has remained almost constant at 20% engine load, but reduced significantly at 80% engine load. This could be due to better combustion of air-LPG mixture under high load conditions. This trend is represented in figure 6. On the other hand both HC and CO emissions have increased with


an addition of LPG content under all operating conditions, as shown in figures 7 and 8.

CONCLUSIONS

A conventional diesel engine could be successfully converted to LPG-diesel dual-fuel operation with minimum alterations. At 20% engine load, diesel could be replaced by LPG up to 50% without any operational difficulties; however, at 80% engine load it could be done only up to 20%. Further, at 20% engine load, diesel fuel operation is

found to be superior to dual-fuel operation, with regards to combustion, performance and emissions. On the other hand at 80% engine load, the reverse has been found to be true. However, emissions of HC and CO have increased on dual-fuel operation under both the loading conditions. At higher engine loads, conventional diesel engines can be switched over to dual-fuel operation, in order to have an overall improvement in the performance of the engine. The scheme developed can be extended for CNG-diesel dual-fuel operation.

Figure 1 Layout of the Experimental Set-Up

Figure2 Variation of Brake Thermal Efficiency with LGP Energy at 20% and 80% Loads

Figure 3 Variation of Mechanical Efficiency with LGP Energy at 20% and 80% Loads


Figure 4 Variation of Volumetric Efficiency with LPG Energy at 20% and 80% Loads

Figure 5 Variation of Brake Specific Fuel Consumption with LPG Energy at 20% and 80% Loads

Figure 6 Variation of Smoke Density with LPG Energy at 20% and 80% Loads

Figure 7 Variation of Emissions of Hydrocarbons with LPG Energy at 20% and 80% Loads


Figure 8 Variation of Emissions of CO with LPG Energy at 20% and 80% Loads

Table 1 Specification of the engine set up

REFERENCES1. M.K.Chaudhari and N.Karupppaiah, “Fuel

Efficiency Policies for HDVs-Indian Perspective”, International Energy Agency, Paris and other Policy Instruments, 21-22 June, 2006, Paris.

2. Ghazi A.Karim, “A Review of Combustion Processes in the Dual-Fuel engine-The Gas Diesel Engine”, Progressive Energy Combustion Science, Volume 6, pp.277-285, 1980.

3. Dong Jian, Gao Xiaohong, Li Gesheng and Zhang Xintang, “Study on Diesel-LPG Dual-Fuel Engines”, SAE International, 2001-01-3679.

4. Poonia, M.P., Ramesh, A., and Gaur, R.R., “Experimental Investigation of the Factors Affecting the Performance of a LPG-Diesel Dual-

Fuel Engine”, SAE International, 1999-01-1123, 1999.

5. Sahir Salman, Can Cinar, Can Hasimoglu, Tolga Topgul and Murat Ciniviz, “ The Effects of Dual-Fuel Operation on Exhaust Emissions in Diesel Engines”, TEKNOLOJI, Volume 7, 2004, Issue 3, pp.455-460.

6. Srinivasa Rao B.R.,Samaga B.S., and Mohanan R.,” Combustion Studies on LPG-Diesel Dual-Fuel Engine”, Proceedings of the 19th National Conference

on Internal Combustion Engines and Combustion, Annamalai University, Chidambaram, December 21-23,2005,pp.125-130.

7. Amarendar Rao, G., Sita Rama Raju, A.V., Mohan Rao, C.V., and Govinda Rajulu, K.,


“Experimental Investigation of a Single-Cylinder, 4-Stroke Diesel Engine Operating on Dual-Fuel Mode (Diesel + LPG)”, International Journal of Scientific Computing, 2(2) July-December 2008, pp.145-152.

8. Ganesan V., “Internal Combustion Engines “, Third Edition, Tata McGraw-Hill Publishing Company, New Delhi, 2007

9. Heywood, John B.,” Internal Combustion Engine Fundamentals”, McGraw-Hill Book Company, 2007.

ACKNOWLEDGEMENTS

The authors wish to express their sincere gratitude to the managements of Kakatiya Institute of Technology and Science, Warangal and Dr.Paul Raj Engineering College, Bhadrachalam for their support and encouragement. The facilities provided by them during the above work is highly solicited.

UNSTEADY MHD PULSATILE GENERALIZED COUETTE FLOW OF A COUPLE STRESS FLUID THROUGH A POROUS MEDIUM UNDER THE

INFLUENCE OF PERIODIC BODY ACCELERATION

S.V. SUNEETHA, Dr. M. VEERA KRISHNA, Dr. S. VENKATESWARLU and

Prof. R. SIVA PRASAD

*Department of Mathematics, Rayalaseema University, KURNOOL (A.P) - 518002 (INDIA).

** Department of Mathematics, Ragiv Gandhi Memorial College of Engineering and Technology, Nandyal, Kurnool (A.P) - 518501 (INDIA)

***Department of Mathematics, Sri Krishnadevaraya University, ANANTAPUR (A.P)-515002 (INDIA)

ABSTRACT:

In this paper, we discuss an analytical study of unsteady magneto hydro dynamic generalized couette flow of an incompressible electrically conducting couple stress fluid through a porous medium between parallel plates, taking into account pulsation of the pressure gradient effect and under the influence of transverse magnetic field and periodic body acceleration with phase difference. The solution of the problem is obtained with the help of perturbation technique. Analytical expression is given for the velocity field, and the effects of the various governing parameters entering into the problem are discussed with the help of graphs. The shear stresses on the boundaries are also obtained analytically and their behaviour computationally discussed with different variations in the governing parameters in detail.

Keywords:

Couette flow, couple stress fluid, periodic body acceleration and porous medium.


1. INTRODUCTION:

Many attempts have been made by several authors to describe blood as a simple model but failed to reach their attempts. In further investigation many authors have used one of the simplification is that they have assumed blood to be a suspension of spherical rigid particles (red cells), this suspension of spherical rigid particles will give rise to couple stresses in a fluid. Stokes [16] introduced the theory of couple stress fluid, which is a special case of micro-polar fluid. Valanis and Sun [17] have proposed a mathematical model for blood flow by assuming blood as a couple stress fluid. It seems that their work contained some serious errors that have been corrected by Chaturani [2]. Further, Chaturani [3] has proposed a method to determine couple- stresses parameters with the help of relative viscosity and velocity profiles. Chaturani

and Upadhya [4] Investigated the pulsatile flow of couple stress fluid by using perturbation method. They have suggested two methods for the determination of the value of puslatile Reynolds’s number. The important conclusion of their analysis is a method (geometrical) that has been developed for studied a theoretical model for pulsatile flow of blood with varying cross sectional tube and its applications to cardiovascular diseases. It is observed that an increase in finding the precise value of non dimensional couple stress parameter. A simple mathematical model depicting blood flow through permeable tube by assuming blood as couple stress fluid has been studied by Pal et al [8].

Sagayamary and Devanathan [11] have studied two dimensional flow of couple stress fluid through a rigid tube of varying cross section for low Reynolds numbers. Padmanabha [7] analyzed pulsatile flow of viscous fluid through a curved elastic tube. Batra and Jena [1] have studied the steady, laminar flow of a Casson fluid in a curved tube of circular cross section. Smith [15] has studied on flow through bends and branching. Schneck [12] has obtained an approximate analytical solution for a pulsatile flow through a diverging channel. Using perturbation method, Rao and Devanathan [9] have analyzed pulsatile flow of blood through varying cross sectional tube. Schneck and Ostrich [13] studied the pulsatile flow of blood in a channel of small expontical divergence. Schneck and Walburn [14]

have investigated the pulsatile flow of blood with low Reynolds number assuming blood as a Newtonian fluid, through a channel of diverging cross section. They have observed a phase-lag between flow rate and pressure gradient. The steady flow of an incompressible micro polar fluid in a diverging channel has been studied by Kamel [5]. Misra and Ghosh [6] used a micro continuum approach to determine the velocity and pressure distributions in an exponentially diverging channel. Rathod [10] studied the pulsatile flow of couple stress fluid through slowly diverging tubes and its applications to cardiovascular diseases. In this paper, we discuss an analytical study of unsteady magneto hydro dynamic generalized couette flow of an incompressible electrically conducting

Where u is the axial velocity, is the density of the fluid, p is the pressure, is the coefficient of viscosity and is the coefficient of couple stress. The boundary conditions are


couple stress fluid through a porous medium between parallel plates, taking into account pulsation of the pressure gradient effect and under the influence of transverse magnetic field and the periodic body acceleration with phase difference.

2. FORMULATION AND SOLUTION OF THE PROBLEM:

We consider the unsteady flow of a couple stress fluid through a porous medium in a parallel plate channel. The flow of a couple stress fluid in a parallel plate channel of width 2h bounded by a porous medium under the influence

of transverse magnetic field. The flow takes place with uniform axial pressure gradient and under the influence of periodic body acceleration with phase difference. The upper plate moves with a constant velocity U in its own plane and bottom plate is at rest. We choose a Cartesian frame of reference o(x, y) with. The flow in the porous region is assumed to be fully developed. The periodic body acceleration is assumed to be where, is the amplitude of the body acceleration and is its phase difference. The flow through a porous medium governed by Brinkman equations. The unsteady hydro magnetic equations governing the couple stress fluid under the influence of transverse magnetic field, uniform pressure gradient and periodic body accelerations with reference to a frame

Conditions (2.3) and (2.4) specify the non-slip conditions at the bounding walls. However condition (2.3) specifies that the fluid adjacent to the non accelerating upper boundarywith velocity U

where as the lower boundary is fixed. Condition (2.5) specifies the vanishing couple stress conditions.Introducing non-dimensional variables are

Using the non-dimensional variables (dropping asterisks), we obtain

Where is the couple stress parameter

is the Hartman numberis the inverse Darcy parameter

is the body acceleration parameterCorresponding the non-dimensional boundary conditions are given by


For the pulsation pressure gradient

Equation (2.6) reduces to the form

The equation (2.11) can be solved by using the following perturbation technique

Substituting the equation (2.12) in (2.11) and equating like terms on both sides

)13.2(cos)( 222122 φ+

∂∂−=++− − Ga

xpa uaDM

yduda

dyud

ss2

s2

4s

4

And

Subjected to the boundary conditions

)17.2(10

)16.2(10)15.2(1

±==

−====

yatydud

yatuyatUu

2s

2s

s

Let


The solutions of the equations (2.13) and (2.14) subjected to the boundary conditions (2.15) to (2.20) give the velocity distribution of the fluid under consideration.

Where, the constants are given in appendix.

The shear stresses on the lower and upper plates are given in dimension less form as

3. RESULTS AND DISCUSSION:

In general the magnitude of velocity u increases from zero

the state of rest on the lower boundary (y=0) to a maximum

in the upper half region and later gradually reduces to rest

on the upper boundary (y=1). The flow governing the non-

dimensional parameters namely viz. a couple stress

parameter, D-1 the inverse Darcy parameter, M the Hartmann number, G body acceleration parameter, P

o the

amplitude of pulsation pressure gradient. Fig 1-3 represent

the velocity profiles for the pulsation pressure gradient

dominates the body acceleration parameter and which

corresponds to with variations in the governing parameters

while fixing the other parameters and the figures (4-6)

represents the reverse case with flow taking place from

right to left. Fig (1 and 4) illustrates the magnitude of

the velocity u enhances with increasing the couple stress

parameter “a” and fixing the other parameters. From

figures (2 and 5), it is evident that the magnitude of the

velocity u decreases with increasing the inverse Darcy

parameter D-1. Hence lesser the permeability of the porous

medium greater the retardation experienced by the flow in

the entire flow field. The magnitude of the velocity u

reduces with increasing the intensity of the magnetic field

for the irrespective of the flow takes place from left to right

and vice-versa. This has been displayed in the figures 3

and 6. The velocity profile (7) exhibits how the velocity u

influenced with the body acceleration parameter G. We

may observe that the negative pressure gradient in the

momentum equation balances the body acceleration term

and hence in the absence of any other extraneous forces the

fluid is at rest, since the channel walls are at rest. However,

when the body acceleration dominates the pulsation

pressure gradient, the magnitude of the velocity component

u enhances with increase in G in the entire flow field.

Likewise it is interesting to note that when the pulsation

pressure gradient dominates the body acceleration, an

increase in G the magnitude of the velocity u reduces in the

entire flow field. The Fig (8) illustrates the magnitude of

the velocity u enhances with increase in the amplitude of

pulsation of pressure gradient. The shear stresses have been

evaluated on the boundaries and tabulated in the tables I

and II. The magnitude of the stresses on either plate

enhances with increase in body acceleration parameter G

and the Hartmann number M, and it reduces with increase

in the amplitude of pulsation pressure gradient and the


inverse Darcy parameter D-1 fixing the other parameters.

Thus lesser the permeability lower the stresses on the

boundaries, also the magnitude of the stresses on the lower

boundary is lesser than the corresponding magnitudes on

the upper boundary. We observe that the stresses reduces

on the upper boundary while enhances on the lower

boundary with increase in the couple stress parameter ‘a’.

-0.8-0.6

-0.4-0.2

00.2

0.40.6

0.81

-1 -0.5 0 0.5 1

y

u

D‾¹=1000D‾¹=2000D‾¹=3000D‾¹=4000

Fig. 1: The velocity profile for u with a

M=2, G=1, D-1 =1000, Po=Ps =10

Fig. 2: The velocity profile for u with D-1

M=2, a=0.5, G=1, Po=Ps =10.


Fig. 3: The velocity profile for u with M

a=0.5, G=1, D-1 =1000,Po=Ps =10.

Fig. 4: The velocity profile for u with a

M=2,G=1, D-1 =1000, Po=Ps =10

Fig. 5: The velocity profile for u with D-1

M=2, a=0.5, G=1, Po=Ps =10,


Fig. 6: The velocity profile for u with M

a=0.5, G=1, D-1 =1000, Po=Ps =10

Fig. 7: The velocity profile for u with G.

M=2,a=0.5, D-1 =1000, Po=Ps =10,

Fig. 8: The velocity profile for u with Po.


M=2, a=0.5, G=1, D-1 =1000, Ps =10,

Table I: The shear stresses on the upper plate.

Table II: The shear stresses on the lower plate.

4. CONCLUSIONS:

1. The magnitude of the velocity enhances with increase in the couple stress parameter ‘a’ and the amplitude of pulsation pressure gradient. 2. The magnitude of the velocity reduces with increase in the Hartmann number M and the inverse Darcy parameter D-1. 3.

When the body acceleration dominates the pulsation pressure gradient,the magnitude of the velocity u enhances with increase in the body acceleration parameter G, while pulsation pressure gradient dominates body acceleration the magnitude of the velocity reduces with increase in G. 4.The magnitude of the stresses on either plate enhances with increase in body acceleration parameter G and it reduces with increase in the amplitude pulsation

pressure gradient and the inverse Darcy parameter.

The stress reduces on the upper boundary and

enhances on the lower boundary with increase in the

couple stress parameter ‘a’. 5. The magnitude of the

lower boundary lesser than the corresponding values

of the upper boundary.

REFERENCES:

1. Batra, R. L. and Bigyani Jena. “Flow of a Casson

fluid in a slightly curved tube”, Int. J. Engg. Sci., 29,

10, 1245-1258 (1991).

2. Chaturani, P. “Some comments on Poiseuille flow

of a fluid with couple-stress, applications to blood

flow”, Biorheology, 6, 85-97 (1979).

3. Chaturani, P. “Viscosity of Poiseuille flow of a

couple stress fluid with application to blood flow”,

Biorheology, 15, 119-128 (1978).

4. Chaturani, P. and Upadhya, V.S. “Pulsatile flow of

a couple stress fluid though circular tuves with


applications to blood flow”, Biorheology , 18, 235-

244 (1981).

5. Kamel, M.T. “Flow of micro polar fluid in a

diverging channel”, Int. J. Engg. Sci. 25, 6, 759-768

(1987).

6. Misra, J. C. and Ghosh, S.K. “A mathematical

model for the study of blood flow through a channel

with permeable walls”. Acta Mechanica, 122, 137-

153 (1997).

7. Padmanabhan, N. “Pulsatile flow of a viscous fluid

of a viscous fluid through curved elastic pipes” Proc.

Ind. Nat. Sci. Acad A. Part A. 53, 208 (1987).

8. Pal L, Rudraiah, N. and Devanathan R. “A couple

stress model of blood flow in the Microcirculation”,

Bull. Math.Biol. 50, 329 (1988).

9. Ram Chandra Rao and Devanathan, R. “Pulsatile

flow in tubes of varying cross sections”. ZAMP, 24,

203-213 (1973).

10. Rathod, V.P. “Mathematical modeling of blood

flow through tubes with various cross sections and

its applications, Ph.D. Thesis, IIT, Bombay (1984).

11. Sagayamary, R.V. and Devanathan, R. “Steady

flow of couple stress fluid through tubes of slowly

varying cross sections, Applications to blood flow”,

Biorheology, 26, 753-769 (1956).

12. Schneck, D.J. “Pulsatile blood flow in a diverging

circular channel”, Ph.D. Thesis, Case Wstern Reserve

University, Cleveland, Ohio, USA (1973).

13. Schneck, D.J. and Walburn, F.J. “Pulsatile blood

flow in a channel of small exponential divergence-I.

The linear approximations for low mean Reynolds

number”. J. Fluids. Engg. 97, 353-360 (1975).

14. Schneck, D.J. and Walburn, F.J. “Pulsatile blood flow in a channel of small exponential divergence Part-II: Steady streaming due to the interaction of viscous effects with convected inertia”. J. Fluids Engg., 98, 707-714 (1976).

15. Smith, F. T. “On flow trough bends and branchings”. Biorheology, 39, 373-378 (2002).

16. Stokes, V.K. “Couple stresses in fluid”, The Physics of Fluids, 9, 1709-1715 (1966).

17. Valanis, K.C. and Sun, C.T., “Poiseulle flow of fluid with couple stress with applications to blood flow”, Biorheology, 6, 85-97 (1979).

APPENDIX:


CONTROL OF THREE PHASE CASCADED MULTILEVEL INVERTER USING VARIOUS NOVAL PULSE WIDTH MODULATION

TECHNIQUES

P.PALANIVEL, SUBHRANSU SEKHAR DASH

Department of Electrical and Electronics Engineering

SRM University Chennai, India


[email protected]:

In this paper, various novel pulse width modulation techniques are proposed, which can minimize the total harmonic distortion and enhances the output voltages from five level inverter to multilevel topologies. Multilevel inverters are important for power electronics applications such as flexible ac transmission systems, renewable energy sources, uninterruptible power supplies and active power filters. Three methodologies adopting the constant switching frequency, variable switching frequency multicarrier, phase shifted carrier pulse width modulation concepts are proposed in this paper. The above methodologies divided in to two techniques. The subharmonic pulse width modulation cascaded multilevel inverter strategy, minimized total harmonic distortion and switching frequency optimal pulse width modulation cascaded multilevel inverters strategy, enhances the output voltages. Field programmable gate array has been chosen to implement the pulse width modulation due its fast proto typing, simple hardware and software design. Simulation and Experimental results are provided.

Keywords:

Constant switching frequency, variable switching frequency, multicarrier pulse width modulation, phase shifted carrier pulse width modulation, subharmonic, switching frequency optimal, cascaded multilevel inverter

INTRODUCTION:

Recently, for increasing use in practice and fast developing of high power devices and related control techniques, multilevel inverters have become more attractive to researches and industrial companies. Multilevel inverters have achieved an increasing contribution in high performance applications. The different multilevel inverter structures are cascaded H-bridge, diode clamped and flying capacitors multilevel inverters [1]-[4]. Increasing the number of levels in the inverter without requiring high ratings on individual devices can increase the power rating [5]. In this paper, constant switching frequency, variable switching frequency multicarrier and phase shifted carrier pulse width modulation methods are used for proposed inverter control methods, since, they are all based on the carrier concept. The control objective is to compare reference with multicarrier

and phase shifted carrier wave using three phase five level cascaded inverter. The multilevel inverter advantages are improved output voltage, reduced output total harmonic distortion, reduced voltage stress on semiconductors switches and decrease of EMI problems [6]-[10].In this paper, three novel carrier pulse width modulation schemes are presented which take advantage of special properties available in multilevel inverter to minimize total harmonic distortion and increases output voltage [11]-[15]. The total harmonic distortion value is high for multicarrier subharmonic pulse width modulation and multicarrier switched frequency optimal pulse width modulation and output voltage level is below actual value [16]. Illustrative examples are given to demonstrate the feasibility of the proposed methods.

II.THREE PHASE CASCADED MULTILEVEL INVERTER


Fig. 1: FPGA based three phase cascaded five level inverter

A Field programmable gate array based three phase cascaded five level inverter is illustrated in fig.1. Each dc source is connected to an inverter. Each inverter level can generate three different voltage outputs, +Vdc, 0, and –Vdc using various combinations of the four switches. The ac outputs of the different full bridge inverter levels are connected in series such that the synthesized voltage waveform is the sum of the inverter outputs. The number of output phase voltage levels m in a cascaded inverter is defined by m=2s+1, where s is the number of separate dc sources.

III. CONSTANT SWITCHING FREQUENCY MULTICARRIER PULSE WIDTH MODULATION

A. Constant Switching frequency Multicarrier Subharmonic Pulse width modulation (CSFMC-SH PWM)

Fig 2: CSFMC-SH PWM modulating signal generation

Fig 3: CSFMC-SH PWM signal generation

Fig.2 shows the Constant switching frequency multicarrier subharmonic pulse width modulation (CSFMC-SH PWM) modulating signal generation. Fig.3 shows the CSFMC-SH pulse width modulation signal generation.


Fig.4 shows an m-level inverter, m-1 carriers with the same frequency fc and the same amplitude Ac are disposed such that the bands they occupy are contiguous. The reference waveform has peak to peak

amplitude Am, the frequency fm, and its zero centered in the middle of the carrier set. The reference is continuously compared with each of the carrier signals. If the reference is greater than s carrier signal, then they active device corresponding to that carrier is switched off.

Fig 4: Constant switching frequency multicarrier subharmonic Pulse width modulation.

In multilevel inverters, the amplitude modulation index Ma and the frequency ratio Mf are defined as

Ma = Am / (m-1)Ac (1)

Mf=fcfm (2)

B. Constant Switching frequency Multicarrier Switching frequency optimal Pulse width modulation(CSMC-SFO PWM)

Fig 5: Multicarrier switching frequency optimal PWM modulating signal generation.


Fig.5 shows the Constant switching frequency multicarrier Switching Frequency Optimal PWM modulating signal generation. Fig.6 shows the CSFMC-SFO PWM in which triplen harmonic voltage is added to each of the carrier

waveforms. The method takes the instantaneous average of the maximum and minimum of the three reference voltages (Va, Vb, Vc) and subtracts the value from each of the individual reference voltages to obtain the modulation waveforms.

Fig 6: Constant switching frequency multicarrier switching frequency optimal pulse width modulation

Voffset = {max (Va,Vb,Vc ) + min (Va,Vb,Vc )} / 2 (3)

VaSFO = Va – Voffset (4)

VbSFO = Vb – Voffset (5)

VcSFO = Vc – Voffset (6)

The zero sequence modification made by the SFO PWM technique restricts its use to three phase three wire system, however it enables the modulation index to be increased by 15% before over modulation or pulse dropping occurs.

In this Paper to increase output voltage, MC-SFO PWM technique is used and by Third harmonic injection, the output voltage Vac can be achieved to 10V with THD value 21.40%.

IV. VARIABLE SWITCHING FREQUENCY MULTICARRIER SUBHARMONIC PULSE WIDTH MODULATIONA. Variable Switching frequency Multicarrier Subharmonic Pulse width Modulation (VSFMC-SH PWM)


Fig 7: VSFMC-SH PWM signal generation

For a multi level inverter, if the level are m there will be m-1 carrier set with variable switching frequency multi carrier Pulse width modulation when compared with sinusoidal reference. The carriers are in phase across for all the bands. In this technique, significant harmonic energy is concentrated at the carrier frequency.

But since it is a co-phasal component, it doesn’t appear line to line voltage. In this paper, we proposed a five level inverter whose levels are 0, ± V/2 and ± V, its carrier set are assigned to have variable switching frequency of 2000 Hz and 4000Hz as shown in the fig.8 and PWM generation as shown in fig.7.

Fig 8: Variable Switching Frequency Multicarrier subharmonic Pulse width modulation.

B. Variable Switching frequency Multicarrier Switching frequency Optimal Pulse width Modulation (VSFMC-SFO PWM)For a multilevel inverter, if the level is m there will be m-1 carrier set with variable switching frequency multi carrier Pulse width modulation when compared with third harmonic injection reference. For third harmonic injection given as

Y = 1.15 sinθ + 1.15 / 6sin3θ. (7)

The resulting flat topped waveform allows over modulation while maintaining excellent AC term and DC term spectra. This is an alternative to improve the output voltage without entering the over modulation range. So any carriers employed for this reference will enhance the output voltage by 15% without increasing the harmonics.

Fig 9: Variable Switching Frequency Multicarrier switching frequency optimal pulse width modulation.

In this paper, there are five level inverter is proposed whose levels are 0, ± V/2 and ± V, its carrier set are assigned to have variable switching frequency of 2000 Hz and 4000Hz as shown in the fig. 9.

V.PHASE SHIFTED CARRIER PULSE WIDTH MODULATION

A. phase shifted carrier subharmonic pulse width modulation (PSC-SH PWM)

Fig.10: Phase shifted carrier subharmonic pulse width modulation

Fig.10 shows the Phase shifted carrier subharmonic pulse width modulation. Each cell is modulated independently using sinusoidal unipolar pulse width modulation and bipolar pulse width modulation


respectively, which provides an even power distribution among the cells. A carrier phase shift of 180°/m for cascaded inverter is introduced across the cells to generate the stepped multilevel output waveform with lower distortion.

Fig. 11: PSC-SH PWM Modulating signal generation

Fig.11 shows the modulating signal generator for the PSC PWM.Optimum harmonic cancellation is achieved in PSC PWM. Phase shifting for carrier is given by,

(K-1)Π/n (8)

Where, k is the kth inverter. n is the number of series

connected single phase inverter.N = (L-1)/2 (9)

Where, L is the number of switched DC levels that can be achieved in each phase Leg.

Fig. 12: PSC-SH PWM signal generation

Fig.12 shows the phase shifted carrier pulse width modulation. The average output voltage for a phase shifted carrier pulse width modulation to a particular power cell i is given by,

Voi = 1/ Tcr . +”Voi(t)dt (10)Voi = Ton/Tcr . Vdc

(11)Voi = V (12)

Where, Voi is the output voltage of cell i, and Ton is the time interval, determined by the comparison between the reference and the carrier signals.

B. Phase Shifted Carrier Switching Frequency Optimal Pulse Width Modulation (PSC-SFO PWM)

Fig.13 shows the phase shifted carrier SFO PWM modulating signal generation. The method takes the instantaneous average of the maximum and minimum of the three reference voltages (Va, Vb, Vc) and subtracts the value from each of the individual reference voltages to obtain the modulation waveforms, which is shown in fig.14.

Fig. 13: PSC-SFO PWM modulating signal generation

From the above criteria we obtain the following equationVcarrier = {max (Va,Vb,Vc ) + min (Va,Vb,Vc )} / 2 (13)

VaSFO = Va – Vcarrier (14)

VbSFO = Vb – Vcarrier (15)

VcSFO = Vc – Vcarrier

(16)


Fig.14: Phase shifted carrier switching frequency optimal pulse width modulation

The carrier voltage is the average of maximum and minimum value of Va,Vb,Vc. The phase voltage using SFO is the difference between reference voltages to carrier voltage. The zero sequence modification made by the SFO PWM technique restricts its use to three phase three wire system, however it enables the modulation index to be increased by 15% before over modulation or pulse dropping occurs.

VI. RESULTS

The total harmonic distortion (THD), which is a measure of closeness shape between a waveform and its fundamental component, is defined as “

THD = 1/V01 (ª Von2 ) (17)

n = 2,3

Where,

V01 is the fundamental rms output voltage.

Von is the rms value of nth harmonic component.

The table.1 shows the THD value and Vac value using CSMC-SH PWM, CSMC-SFO PWM, VSMC-SH PWM and VSMC-SFO PWM. Using PSC-SH PWM and PSC-SFO PWM the THD and Vac values are reduced respectively.

The table.2 shows the THD and output voltage value for PSC-SH PWM and PSC-SFO PWM. The THD value for PSC-SFO PWM was seen to be high. Inspite of this high value the output voltage was improved. So, PSC-SFO PWM technique can be used where high output voltage is needed.

TABLE 1

VARIOUS MODULATION INDEX OUTPUT VOLTAGE AND THD FOR MC-

SH PWM AND MC-SFO PWM


TABLE 2

VARIOUS MODULATION INDEX OUTPUT VOLTAGE AND THD

FOR PSC-SH PWM AND PSC-SFO PWM

A. Simulation ResultsTo verify the proposed schemes, a simulation model for a three phase five level cascaded H-Bridge inverter is implemented. The simulation parameters for constant switching frequency multicarrier pulse width modulation are as following, 5KW rating, three phase load R = 100 ohms, L = 20mH, each source Vdc = 5V, switching frequency 2KHz. Phase leg voltages have been calculated and drawn for CSFMC-SH PWM Method in Fig.15, 16, 17. Phase leg voltages have been calculated and drawn for CSFMC-SFO PWM Method in Fig.18, 19, 20.


Fig.15: CSFMC-SH PWM signal generation

Fig.16. CSFMC-SH PWM output voltage

Fig 17.CSFMC-SH PWM harmonic spectrum


Fig 18: CSFMC-SFO PWM signal generation

Fig 19: CSFMC-SFO PWM output voltage

Fig 20: CSFMC-SFO PWM harmonic spectrum

The simulation parameters for variable switching frequency multicarrier pulse width modulation are as following, 5KW rating, three phase load R = 100 ohms, L= 20mH, each source Vdc = 5V, switching frequency 2KHz and 4KHz. Phase leg voltages have been calculated and drawn for VSFMC-SH PWM Method in Fig.21 ,22, 23.Phase leg voltages have been calculated and drawn for VSFMC-SFO PWM Method in Fig.24, 25, 26.

Fig 21: VSFMC-SH PWM signal generation


Fig 22: VSFMC-SH PWM output voltage

Fig 23: VSFMC-SH PWM harmonic spectrum

Fig 24: VSFMC-SFO PWM signal generation

Fig 25: VSFMC-SFO PWM signal generation

Fig 26: VSFMC-SFO PWM harmonic spectrumThe simulation parameters for phase shifted carrier pulse width modulation are as following, 5KW rating, three phase load R = 100 ohms, L = 20mH, each source Vdc = 5V, switching frequency 5KHz. Phase leg voltages have been calculated and drawn for PSC-SH PWM Method in Fig.27, 28, 29. Phase leg voltages have been calculated and drawn for CSFMC-SFO PWM Method in Fig.30, 31, 32.


Fig.27: PSC-SH PWM signal generation Fig.28: PSC-SH PWM output voltage

Fig.29: PSC-SH PWM Harmonic spectrum


Fig.30: PSC-SFO PWM signal generation

Fig.31: PSC-SFO PWM output voltage

Fig.32: PSC-SFO PWM Harmonic spectrum


B. Hardware Results

A hardware setup of three phase five level cascaded inverter has been built to validate the theoretical analysis. The hardware parameters for CSFMC PWM are as following, 5KW rating, three phase load R = 100 ohms, L = 20mH, each source Vdc = 5V, fundamental frequency 50HZ, switching frequency 2KHZ and Xilinix Spartan – DSP controller (FPGA). The three phase output voltage waveform for CSFMC-SH PWM method shown in fig.33 and CSFMC-SFO PWM method shown in fig.34.

Fig 33: CSFMC-SH PWM output voltage

Fig 34: CSFMC-SFO PWM output voltage The hardware parameters for VSFMC PWM are as

following, 5KW rating, three phase load R = 100 ohms, L = 20mH, each source Vdc = 5V, fundamental frequency 50HZ, switching frequency 2KHZ, 4Khz and Xilinix Spartan – DSP controller (FPGA). The three phase output voltage waveform for VSFMC-SH PWM method shown in fig.35 and VSFMC-SFO PWM method shown in fig.36.

Fig 35: VSFMC-SH PWM Phase voltage


Fig 36: VSFMC-SFO PWM Phase voltage.

The hardware parameters for PSC PWM are as

following, 5KW rating, three phase load R = 100 ohms, L = 20mH, each source Vdc = 5V, fundamental frequency 50HZ, switching frequency

2KHZ and Xilinix Spartan – DSP controller (FPGA). The three phase output voltage waveform for PSC-SH PWM method shown in fig.37 and PSC-SFO PWM method shown in fig.38.

Fig.37: PSC PWM output voltage


Fig.38: PSC-SFO PWM output voltage

Fig.39: Hardware setup of three phase cascaded multilevel inverterV. CONCLUSION

In this paper, three new schemes adopting the constant switching frequency multicarrier, variable switching frequency multicarrier and phase shifted carrier pulse width modulation concepts are proposed. The subharmonic pulse width modulation strategy reduces the THD and switching frequency optimal pulse width modulation strategies enhances the fundamental output voltage. The multilevel inverter improves output voltage, reduces output total harmonic distortion and voltage stress on semiconductors switches. These schemes are confirmed by simulation results and experimental results.

REFERENCES[1] K.A Corzine, and Y.L Familiant, “A New Cascaded Multi-level H-Bridge Drive,” IEEE Trans. Power.Electron., vol.17, no.1, pp.125-131. Jan 2002.

[2] R.Teodorescu, F.Blaabjerg, J.K.Pedersen, E.Cengelci, and P.N.Enjeti, “Multilevel Inverter by cascading industrial VSI,” IEEE Trans. Ind. Electron., vol.49, no.4, pp.832-838. Aug.2002.

[ [ 3] J.S.Lai, and F.Z.Peng “Multilevel converters – A new bread of converters,”IEEE Trans. Ind.Appli., vol.32, no.3, pp.509-517. May/Jun.1996.


[4] T.A.Maynard, M.Fadel and N.Aouda, “Modelling of multilevel converter,” IEEE Trans. Ind.Electron., vol.44, pp.356-364. Jun.1997.

[5] P.Bhagwat, and V.R.Stefanovic, “Generallized structure of a multilevel PWM Inverter,” IEEE .

Trans. Ind. Appln.., vol.1A-19, no.6, pp.1057-1069. Nov./Dec..1983.

[6] J.Rodriguez, Jih-sheng Lai, and F Zheng peng, “Multilevel Inverters; A Survey of Topologies, Controls, and Applications,” IEEE Trans.Ind.Electron., vol.49 , no4., pp.724-738. Aug.2002.

[7] G..Carrara, S.Gardella, M.Marchesoni, R.salutari,and G.sciutto, “A New Multilevel PWM Method; A theoretical analysis,” IEEE Trans. Power.Electron., vol.7, no.3, pp.497-505. Jul.1992.

[8] L.M.Tolber, T.G.Habetler, “Novel Multilevel Inverter Carrier based PWM Method,” IEEE Ind.Appli., vol.35. pp.1098-1107. Sep/Oct 1999.

[9] B.P.McGrath and Holmes, “Multicarrier PWM strategies for multilevel inverter,” IEEE Trans.Ind.Electron., vol.49, no.4, pp.858-867. Aug.2002.

[10] Samir koaro, PabloLezana, Mauricio Anguio, and Jose Rodriguez, “Multicarrier PWM DC-Link ripple forward compensation for multilevel inverters,” IEEE Trans. Power.Electron., vol.123, no.1, pp.52-56.Jan 2008.

[11] P.Palanivel and Subhransu Sekhar Dash “A FPGA based variable switching frequency multicarrier pulse width modulation for three phase cascaded multilevel inverter,” Proc. INCACEC-2009. conf , pp.811-815. Kongu Engineering College, Erode, India, June 2009.

[12] B.P.McGrath, Holmes, and T.Meynard, “Reduced PWM Harmonic distortion for multilevel inverter operating over a wide modulation range, “ IEEE Trans. Power.Electron., vol.21, no.4, pp.941-949. Jul.2006.

[13] S.Sirisukpraserl, J.S.Lai, and T.H.Liu, , “Optimum harmonic reduction with a wide range of modulation indices for multilevel converters,” IEEE Trans. Ind. Electron., vol.49, no.4, pp.875-881. Aug..2002.

[14] Roozbeh Naderi, and Abdolreza rahmati, “Phase-shifted carrier PWM technique for general cascaded inverters,” IEEE Trans. Power.Electron., vol.23, no.3, pp.1257-1269. May.2008.

[15] Samir koaro, PabloLezana, Mauricio Anguio, and Jose Rodriguez, “Multicarrier PWM DC-link ripple forward compensation for multilevel inverters,” IEEE Trans. Power.Electron., vol.23, no.1, pp.52-56. Jan 2008.

[16] P.Palanivel and Subhransu Sekhar Dash, “Multi carrier pulse width modulation based three phase cascaded multilevel inverter including over modulation and low modulation indices,” RI Pub. Int.

Journ. Eng. Studies. Vol.1, no.2, pp.71-82. June 2009. Biographical notes:

P.Palanivel received M.E degree in Electrical Engineering from Anna University, Chennai, India in 2004. He is currently pursuing the Ph.D in Electrical Engineering at the SRM University Chennai, India. His research interests are in Power Quality improvements in Inverters, Multilevel inverters & Resonant Inverters.

Subhransu Sekher Dash received the M.E degree in Electrical Engineering from UCE Burla , Orissa, India and Ph.D degree in Electrical Engineering from Anna University in 1996 and 2006 respectively. He is presently working as Professor in SRM University Chennai, India. His area of interest includes Power Quality, Inverters, Multilevel Inverters, Power System Operation, Control & Stability and Intelligent controlling Techniques.


COMPARISON OF CELL ADHESION CHARACTERISTICS OF ANODIZED AND HYDROXYAPATITE COATED TI6AL4V IMPLANT

MATERIAL BASED ON WET ABILITY AND IN-VITRO STUDIESK.K.SAJU1, S.VIDYANAND 3, JAYADAS .N.H1, M.K.JAYARAJ2,

JACKSON James3

1. Division of Mechanical Engineering –Cochin University of Science & Technology-Kochi-India.

2. Department of Physics –Cochin University of Science & Technology-Kochi-India.

3. Neuro- Stem Cell Biology Laboratory, Neurobiology Division, Rajiv Gandhi Center for Biotechnology, Trivandrum, India

91-484-2345344.91-484-2556187. 91-484-2556187. [email protected]

ABSTRACT:

In this study Ti6Al4V (Ti) samples were surface modified by anodization and hydroxyapatite coating and their properties compared to evaluate the adhesion characteristics of protein adhesion molecules which favor osteoblast attachment during implantation. The physical properties of the surfaces like surface roughness, porosity and contact angle were measured for the modified samples and compared with control Ti sample .The in –vitro cell viability was found out using an MTT (3-(4,5-dimethylthiazole-2-yl)-2,5-diphenyltetrazolium bromide) assay (MTT assay) as well as confocal imaging. It was found that the hydroxyapatite coated surface gave a more favorable protein adhesion characteristics based on its physical properties which was confirmed by in-vitro cell viability tests. Among the anodized Ti samples the one anodized to a higher degree showed comparable physical properties to HA coated sample for protein adhesion and the fact was established in the cell viability study. The study showed that even though HA coated Ti samples give appreciable cell adhesion properties, samples anodized to a higher level of anodization can give comparative cell adhesion to HA coated surfaces.Physical characteristics of the surfaces like roughness,porosities etc plays a more active role in cell adhesion compared to its wetability. The above results can be made use of while selecting surface modification techniques for orthopedic implants.

Keywords:

Anodization, pulsed laser deposition, hydroxyapatite, contact angle, average roughness. Ti6Al4V-Ti, Hydroxyapatite-HA, Pulsed laser Deposition-PLD.

1. INTRODUCTION

Biocompatibility of orthopedic implants depends on

the ability of surface tissues to bond directly on to the

implant material. Cellular attachment is dependent on

many parameters including surface roughness,

wetability, film thickness and other surface

properties. It has been established that the adhesion

of cells on to the surfaces of orthopedic implants

depends on the ability of the surfaces to

accommodate protein molecules . Protein adhesions

are found to be effective in cases of higher surface

roughness, macro porosity and hydrophobic nature of

implant materials [1-5].Various surface alteration

techniques have been applied to Ti to make it more

biocompatible. Studies have been made on the

adhesion


characteristics of modified Ti implants based on their surface properties like roughness, film thickness, porosity etc [6-8].Work on the wetability nature of modified surfaces to predict cell adhesion are still in progress. It has been established that theoretically molecular proteins favor a more hydrophobic surface in relation to other surface properties [9].This study conducted contact angle and wet-ability observations of anodized and HA coated Ti samples along with other surface properties to assess the biocompatibility of these surfaces and counterchecked the results by in-vitro cell viability studies.

2. MATERIAL AND METHODSTi discs of 15mm diameter cut from cylindrical rods were used as substrate for PLD and for anodization. Before coating, the substrate surface was mechanically polished through No. 600 grit, cleaned ultrasonically and sand blasted using 100-150 pm Silica sand. Targets used for the laser deposition of HA were prepared from pressing and sintering of HA powder prepared by the wet method [10]. The powder was subjected to cold isostatic pressing under a pressure of 120MPa and sintered at 1200°C for 2h in air.

2.1 Pulsed Laser Deposition of HA films The PLD system used in this study is shown

schematically in Figure 1 [11-17].

The ablation target and substrate, which was 4 cm from the target and parallel to the target, were placed in a stainless steel vacuum chamber. While the substrate was stationary, the target was rotating during the entire deposition process at a speed of 10 rpm. The NdYaG laser (Model X spectra physics USA) with a wavelength of 355nm and pulse duration of 9 ns was guided into the chamber using an anti-reflection coated convex lens through a quartz window.Stabilized output energy density of 1.6 J/cm2

per pulse with a repetition rate of 10 Hz was used. The laser beam was made to fall on to the target at an angle of 450. Prior to deposition, the chamber was evacuated to a base pressure of 3 x 10-6 m bar by a combined rotary and diffusion pumping system and oxygen was bled in to the chamber to maintain an operating pressure of 2 x 10-3 m bar. Substrates were heated with inbuilt heater and coatings were done on substrates at the temperature range 150 to 5000 C. Coating duration for all the films was uniform at 2 hrs. The as deposited samples were amorphous as observed by X Ray Diffractometer (XRD).These samples were subjected to a hydrothermal annealing. The samples were heated in the presence of 20 ml of water in a sealed hydrothermal parr bomb (100 ml) for two hours at a temperature of 1000 C. Since the coating at 2000C followed by hydrothermal treatment gave the best characteristics for near stoichiometric Ca/P ratio and coating thickness, it was used for our study.

Fig 1: Set up for pulsed Laser Deposition


2.2 Anodization

The set up for anodization is shown in fig 2 [18]. Ti

was anodized in 200 g/L sulfuric acid, 5% trisodium

phosphate, and 5% sodium bicarbonate. The

electrolyte was contained in a suitable chemical

resistant tank with Fume extraction. A D.C. electrical

supply with voltage regulation from 2 to 100 volts

was provided. Cathodes of lead were used. The parts

to be treated was immersed in the processing solution

and connected as the anode to the electrical D.C.

source.

The temperature of the bath was maintained in the range of 16° to 26°C throughout the duration of treatment. The cell voltage was varied between 50-75 volts. The time of treatment was for 15 minutes. Immediately after removal from the anodizing bath, parts were washed thoroughly in clean running water, rinsed in clean hot water and allowed to dry. Three samples each were developed at 50 volts, 60 volts and 75 volts yielding yellow, pink and blue colored surfaces and labeled as yellow, pink and blue samples.


2.3.1 Film ThicknessThe thickness and average roughness of the coated and anodized films were measured using a VEECO Dektak 6M stylus profilometer.

2.3.2 X-Ray diffraction

A X-ray diffractometer (Rigaku, D-max), with Cu-K? (1.5418A ) radiation which was operated at 30 kV and 20mAwith a scanning speed of 50/ min, was used to identify phases of the deposited HA film.

2.3.3 Surface Morphology

Surface morphology of the modified surfaces was examined using Scanning Electron Microscope (SEM) (Hitachi S-2500, Hitachi, and Tokyo, Japan).

2.3.4 Contact angle measurement

The contact angles of the samples with three different fluids were measured using a NRL contact angle goniometer using the sessile drop method [19] in three well characterized liquids, water, formamide and di-iodomethane as per previous studies [20].

2.4 Cell Viability studies

2.4.1 Cell Culture

Osteosarcoma cell line KHOS-NP (R-970-5) [NCCS] were grown in culture medium (DMEM+10%FBS+1mM NEAA) in a T-25 flask and incubated at 37oC for 2 days in a 5% C02 incubator (Thermo). ~ 5x105cells were plated on to three samples each of anodized, HA coated and plain polished control samples of Ti in 12 well plates. The culture was incubated for 72 hours at 37oC in 5% CO2

incubator. The samples with attached cells were used for MTT assay and confocal microscopy. 2.4.2 MTT assay

For MTT assay, all the samples were transferred to a fresh plate and 800 mL of MTT reagent was added to each well and incubated for 2 hours at 37OC. MTT transformed to dark blue formazan by mitochondrial dehydrogenises enabling cell viability to be assessed. 800 mL of lysis buffer (20%Sodium Dodecyl Sulphate 50% Dimethyl Formamide 30% Distilled water) was added to each well, mixed and incubated at 37oC for 4 hours. 200 mL of each sample was transferred to a fresh 96 well plate and the optical density of the solution was measured at 570nm in an ELISA microplate reader (Biorad USA). Analysis of optical variance was used to evaluate difference in cell viability between the groups [21].

2.4.3 Confocal Imaging

Osteosarcoma cells were grown on another set of anodized ,HA coated and control Ti samples in culture medium for 48 hours. After which the cells were fixed with 4% Paraformaldehyde ( PFA), washed twice in 1X Phosphate Buffered Saline (pH 7.4) (PBS) and incubated in DAPI (4’, 6-Diamidino-2-phenyindole) (1:1000) for 10 minutes at room temperature (280C). Cells were imaged for nuclear visualization thereafter using a confocal microscope (Leica TCS SPE Germany).

3 .RESULTS AND DISCUSSION:The measurements of Average Surface Roughness, Film Thickness and contact angle measurements of the various surface modified Ti substrates are shown in table 1.The Average surface roughness of the anodized film varied from 0.29 ?m for the yellow sample to 0.474 ?m for the blue sample. The surface roughness for HA coated film was the highest at 0.68 ?m compared to 0.032 ?m for the control Ti. The thickness of the anodized films coated at 50-75 volts varied as 0.89, 0.98 and 1.1?m.


Table 1: Contact angle and roughness measurements of surface modified Ti samples and control sample

The contact angle in water and Di-iodomethane was

the highest for the sample anodized at a voltage of 75

volts whereas for that in formamide the contact angle

was maximum for HA coated sample. In all cases the

sample coated with HA was either having the highest

contact angle or the next best of all the samples.

Since formamide has a ph value almost close to that

of blood (7.3), the contact angle measured in

formamide can be safely assumed to be the nature

that the sample will be exhibiting with blood. In

previous studies, mechanism of cell adhesion is seen

to be due to the adhesion of protein molecules present

on the surface of the

cells known as adhesion molecules [1-6]. A larger

contact angle theoretically increases the protein

adhesion chances by exposing surfaces with lesser

wetting. The adhesion nature is depicted in fig 3

wherein protein molecules are seen to get direct

adhesion sites on the surface compared to a more

hydrophilic surface. Protein adhesion molecules from

the extra cellular matrix (ECM) will get increased

chances of adhesion to a HA coated surface and act

as a ligand site for binding of cell adhesion molecules

present the osteoblasts thereby enhancing

osseointegration

.

Fig-3. Depiction of adhesion characteristics of protein molecules on to a hydrophobic and hydrophilic substrate. a) Protein adhesion hindered by adsorbed water molecules on a Hydrophilic substrate b) Direct

contact of protein molecules on a hydrophobic substrate.


Since higher roughness within a range of 0.22-0 .78

increases cell adhesion [7] ,HA coated Ti samples

will show a better cell adhesion characteristic

compared to other surface modified Ti samples. The

blue Ti also shows a higher roughness and contact

angle compared to other Ti samples. The contact

angle of control Ti is higher than that of the yellow

and pink samples but the roughness values of the

anodized surfaces are much higher and also they

exhibit micro porosity which will aid in cell adhesion

by mechanical interlocking.

in fig 4.The marked groupings shows HA with the

highest number of scores within a sample space of

roughness greater than 0.5 ?m and contact angle in

formamide and water respectively being greater than

or equal to 70 degrees and 80 degrees respectively.

The blue anodized sample exhibited comparable

groupings to the HA coated sample. The pink and

yellow samples did not have scores in this region,

this is more so due to its higher wetabilty nature even

though the roughness is close to this range.


Fig 4: Plot of contact angles and roughness for different surface modified Ti samples 1. Ti anodized at 75 volts (Blue), 2. Ti anodized at 60 volts (Pink), 3. Ti anodized at 55 volts (Yellow),

4. Ti Plain Polished (control), 6. Ti HA coated

The SEM micrographs of all the samples are shown

in figs 5. Broad face morphology of the HA coated

film after the hydrothermal treatment showed that the

HA coating comprised of numerous spheroidal and

needle like aggregates of different sizes. The surface

topography of the anodized samples showed micro

pores of the order of 24-32?m.The blue anodized

sample showed largest porosity of 32 ?m. The micro

pores help in cell adhesion by

providinginterlockingsites.

Fig 5: SEM micrographs of anodized and etched samples a. Ti anodized at 75 volts, b.Ti anodized at 60 volts, c.Ti anodized at 55 volts, d.HF etched Ti, e.Ti HA

coated


The cell viability plots based on the analysis of

optical variance method after MTT assay is shown in

fig 6.The maximum cell viability percentage is

shown by the HA coated Ti..The

blue anodized sample exhibits comparable cell

viability. The cell viability of the pink and yellow

sample is seen to be much greater than the control Ti.


Fig 6 : Cell viability by the optical variance method.

The nuclear visualization thro confocal imaging of all the samples are shown in fig 7.The HA coated sample shows highest nuclear density shown by the number of spots followed by the blue, pink and yellow samples. The control Ti shows the least nuclear visibility. This shows that the HA coated Ti which returned all physical attributes required for protein adhesion is giving the highest cell viability perecentage in the cell viability tests. However the blue anodized sample also returned comparable

physical properties required for protein adhesion and gives a comparable cell viability result to the HA sample. Even though the contact angle measurements of the pink and yellow sample is only comparable to the control Ti sample they returned a much higher cell viability result than the control Ti. The other physical properties of roughness and microporosities of the yellow and pink surfaces play an important role in this result compared to the wetability of the surfaces.

Fig 7: Confocal images of nuclear visualization a) HA Coated, b) Blue, c) Pink, d) Yellow, e) Control Ti


c d

e

5 CONCLUSIONS: Ti surfaces coated with Pulsed laser deposited HA shows better surface properties like roughness and more hydrophobicity required for protein adhesion compared to anodized Ti samples and will give good osseointegration. This is confirmed by the in-vitro cell viability analysis. However the anodized blue

surface also exhibited comparable surface properties to HA coated implants and showed good cell viability in the in-vitro studies .The other anodized pink and yellow surfaces also showed appreciable cell viability in comaparison with control Ti surfaces even though their wetability was on the higher side. This shows that surfaces anodized to the correct level can return comparable


osseointegration to HA coated surfaces. The surface roughness and microporosities of the anodized surfaces plays a greater role in the cell adhesion

compared to wetability. Further in vitro studies are required

to assess the actual cell adhesion parameters of the

modified implants.

ACKNOWLEDGEMENTS

The authors place their gratitude to the staff and research scholars of the Department of Physics, Cochin University of Science & Technology, and Kochi, India for their kind assistance in this study. The authors also thank Dr K.V.Menon and Dr H.K.Verma for their valuable advises and guidance. The staff of Sophisticated Tests and Instrument center Kochi is acknowledged for their support in doing various laboratory studies. We also acknowledge the assistance of Ms Geetha Devi Research scholar IISc Bangalore India in conducting the contact angle measurements.

REFERENCES

1. Clare M.Isacke, Michael A.Horton. The adhesion Molecule-Facts book, Academic Press .ISBN 0-12-356505-7

2. Joon .B.Park, Joseph D Bronzino. Biomaterials Principles and Applications, CRC Press ISBN 0-8493-1491-7

3. D.E.Pacham.Handbook of Adhesion second edition, John Wiley and Sons,Ltd ISBN 0-471-80874-1(HB)

4. Kevin Kendall Molecular Adhesion and its Applications, Kluwer academic publications ISBN 0-306-46520-5

5. Michael J Parnham Adhesion Molecules Functions and Inhibitions, Birkhauser Verlag ISBN 978-3-7643-7974-2

6. Jo-Young Suh, Bong-Vheol Jang, Xiaolong Zhu, Joo L.Ong, Kyohan Kim .Effect of hydrothermally treated anodic oxide films on osteoblast attachment and proliferation. Biomaterials 24 (2003) 347-355

7. Despina D Deligianni,Nikoleta D Katasala, Petros G Koutsoukos,Yiannis F Misirlis Effect of surface

roughness of hydroxyapatite on human bone marrow cell adhesion, proliferation, delamination and detachment strength. Biomaterials 22 (2001) 87-96

8. Young-Taeg Sul, Carina B Johansson, Kerstin Roser, Tomas Albrektsson. Qualitative and quantitative observations of bone tissue reactions to anodized implants. Biomaterials 23 (2002) 1809-1817.

9. G.Legeay,F Poncin-Epaillard .Surface Engineering by Coating of Hydrophilic Layers.Adhesion Current Research and Applications .Wiley VCH Verlag Gmbh & Co KGaA.ISBN 3-527-31263-3.

10. Hideki Aoki. Medical Applications of Hydroxyapatite, Takayama Press, 1994.

11. A.Bigi,B.Bracci,F.Cuisinier,R.Elkaim,M.Fini,I.Mayer,I.N.Mihailescu,G.Socol,L.Sturba,P.Torricelli. Human Osteoblast response to pulsed laser deposited calcium phosphate coatings. Biomaterials 26 (2005) 2381-2389.

12. Quanhe BAO, Chuanzhong Chen, Diangang Wang, Qianmao Ji, Tingquan Lei. Pulsed laser deposition and its current research status in preparing hydroxyapaptite thin films. Applied Surface Science 252 (2005) 1538-1544.

13. Yoshiaki Suda,Hiroharu Kawasaki,Tamiko Ohshima,Shouta Nakashima,Syuichi Kawazoe,Tetsuya Toma. Hydroxyapatite coatings on titanium dioxide thin films prepared by pulsed laser deposition method.Thin Solid Films 506-507 (2006) 1115-119.

14. J.M.Fernandez-Pradas, L.Cleries, E.Martinez, G.Sardin, J.Esteve, J.L.Morenza.Influence of thickness on the properties of hydroxyapatite coatings deposited by KrF laser ablation. Biomaterials 22 (2001) 2171-2175

15. Won-Jun Lee, Sang-Wook Lee, Hye-Lee Kim, Dae-Joon Kim.Characteristics of calcium phosphate films prepared by pulsed laser deposition under various vapor pressures. Journal of the Korean Physical Society vol 47 July 2005, 152-156.

16. W.Mroz, M.Jedynski, J.Hoffman, M.Jelinek, B.Major, A.Prokopiuk, Z.Szymanski.Effect of reactive atmosphere on pulsed laser deposition of


hydroxyapatite thin films. Journal of Physics. Conference Series 59(2007) 720-723.

17. Daniel Ferro, Segey M.Barinov, Julietta V.Rau, Roberto Teghil, Alessandro Latini.Calcium phosphate and fluorinated calcium phosphate coatings on titanium deposited by ND: YAG laser at a high fluence.

18. John A Disegi, Anodizing Treatments for Ttanium Implants .0-7803-3867-3/97 10.00 1997 IEEE.

19. Kash Mittal .Adhesion Aspects of Thin Films. VSP. ISBN 90-6764-421-8.

20. A.A. Thorpe, Thomas G. Nevell, Simon A. Young, John Tsibouklis.Surface energy characteristics of

poly_methylpropenoxyfluoroalkylsiloxane/ film structures. Applied Surface Science 136 _1998. 99–104.

K.Bordji, J.Y.Jouzeau, D.Mainard, E.Payan, P.netter, K.T.Rie, T.Stucky and M.Hage-Ali. Cytocompatibility of Ti-6Al-4v and Ti-5Al-2.5Fe alloys according to three surface treatments, using human fibroblasts and osteoblasts. Bioamaterials 17 (1996)929-940.


OPTIMIZATION OF ENERGY FOR AN AIR CONDITIONING SYSTEM USING PLC BASED DCV APPROACH

S.P. VENDAN

Assistant Professor

Department of Mechanical Engineering, P.S.G. College of Technology

Coimbatore, India, [email protected]

P.RAVIKUMAR

Professor

Department of Mechanical Engineering Jayararam College of Engineering and Technology, Tiruchirappalli, India

ABSTRACT:CO2 based DCV can be used in applications where the sources of indoor air contaminants are dominated by occupants of a building. A recent interpretation of ANSI/ASHRAE standard 62-1989 interpretation IC 62-1989-27 has affirmed that CO2 based DCV systems can use CO2 as an occupancy indicator to modulate ventilation below the maximum total outdoor air intake rate while still maintaining the required rate per person ,provided that certain conditions are met.The reduction of ventilation air will reduce the energy consumption in the compressor, blower and condenser fan. This technique is more suitable for movie theatres and conference halls. Where occupants level vary often below the design occupancy level. This paper presents a dynamic method for detecting the actual occupancy in indoor space by measuring CO2 concentration of return air and outdoor air flow rate for energy conservation in a movie theatre. In addition, variation in the occupancy level is estimated using a steady state algorithm.

Keywords:

DCV, Occupancy detection algorithm, Comfort condition,CO2 concentration, PLC programming.

1 INTRODUCTION

Ventilation is one method to maintain good indoor air quality. The more fresh air is brought in to the indoor environment, the better the indoor air quality can be achieved if the fresh air comes from non-polluted ambient source. However, conditioning fresh air can consume lot of energy, for typical office building around 30% or more of the annual heating and cooling cost is spent in handling the fresh air. Over ventilation may lead to significant waste of energy. Therefore, an operationally cost effective ventilation

system is very important in buildings (Chao, 2004).The work has been done in a movie theater that is fully air-conditioned.

The Demand Controlled Ventilation (DCV) technique describes the proper procedure for using CO2 control to modulate ventilation based on actual occupancy. For spaces subject to variable or intermittent occupancy, over ventilation can be avoided by reducing total ventilation below design


ventilation rates. This approach allows us to vary the ventilation air depends upon the number of occupants in design space. By this way the energy can be

conserved and ventilation for acceptable Indoor Air Quality (IAQ) can also be maintained in accordance with ANSI/ASHRAE standard.

It is generally accepted that considerable energy savings can be made if the ventilation supplied to buildings is no more than is necessary to maintain healthy environment for the occupants (Warren, 1991). Several methods of relating the ventilation rate to the occupation level have been proposed and are currently being used in certain buildings, which in effect attempt to adjust the ventilation rate in direct proportion to the number of people. Usually the fresh air requirements are stipulated as cubic feet per minute (cfm) per person or minimum air changes per hour. In movie theater the fresh air requirement is 15 cfm (7.5L/S) (Mike Schell, 2001). The number of researchers already had a research work in various applications using the same DCV technique and they were calculated the percentage of energy savings by applying the DCV technique. Alalwi (2001) presented a HVAC laboratory study comparing the effect of different feedback control strategies on the CO2 levels and HVAC equipment energy consumption .Warren and Harper (1991) compared simulation of energy savings for two buildings in U.K.using the data derived from a live project carried out by ECD partnership. Which shows that energy savings using CO2 concentration based ventilation control may be considerable, up to 53% of the energy consumption of the original 100% fresh air system.

2 METHODOLOGY

The energy conservation is achieved by applying the DCV technique to the air-conditioned movie theater. The energy savings is calculated by load reduction at compressor corresponding to the different occupant’s level over a particular period of time. Over ventilation is avoided by supplying the ventilation air depending upon the number of occupants inside the design space. The indoor air quality is maintained by supplying the base line ventilation air (10 % of the design ventilation air) at non occupant’s period [Sowa, 2003]. The base line ventilation air will remove the contaminants coming from the non-occupant’s sources. The occupancy level inside the movie theater can be determined by measuring the CO2 concentration inside the space by using the CO2 sensor. The output of the CO2 sensor is send to the Programmable Logic Control (PLC) unit, where the occupancy based detection algorithm is programmed and is compiled by using the software STEP 7.0 micro win. The position of the damper is controlled by PLC unit through the actuator according to the occupancy level in the movie theater.

3 PROBLEM DEFINITION

The Air-conditioned movie theater has a capacity of 600 occupants at the design level. But the occupation level

varies depending upon the period and nature of the movie. At the design level the design engineer calculates the ventilation air amount corresponding to the fulloccupant level. But ventilation air may be varied depending upon the number of occupants inside the theater. In actual practice they are supplying the ventilation air corresponding to the design condition i.e. 100% occupant level irrespective of the number of occupants in the theater, this is higher than what they need in actual condition, this will increase the energy consumption and create over ventilation. By using DCV technique the energy conservation at various occupant levels is calculated. Calculation is done by measuring CO2 concentration inside the theater, because the CO2 level is directly proportional to the occupants inside the room. This shows that lot of energy savings is possible in the movie theater and the over ventilation can also be avoided. The HVAC system design for new and renovated buildings is based upon compliance with applicable codes and standards. To assure high indoor air quality in buildings careful attention is given to the contaminant source control, proper ventilation, humidity management, and adequate filtration. The non occupancy pollutants are removed by supplying the minimum amount of 10% fresh air of the design occupancy level when there is no occupant in the movie theater [Sowa, 2003], the ventilation and humidity is maintained by supplying the ventilation air per person defined by the ASHRAE standards.

3.1 Rates of CO2 generation for various activities

Carbon dioxide is a byproduct of respiration. The rate at which we produce it varies with diet and health, as well as with the duration and intensity of physical activity. The more exertion an activity entails, as measured in metabolic equivalent task (MET) units, the more carbon dioxide we produce. The physical activity level for occupants seated at rest is 1.2 MET and corresponding CO2 generation rate is 0.3L/min.It is depicted in the figure 1.

4 DATA COLLECTION

The main objective of the work is to reduce the energy conservation in a large air-conditioned system where the occupant level is varying very frequently and the same way to achieve the comfort conditions like temperature, velocity, humidity and contaminants inside the building. For this purposes the movie theater, which is fully, air-conditioned is taken for the study. The various data’s are collected in the air conditioning movie theater for particular period of time. The movie theater has the dimension of 36mx12m with a ceiling height of 7.5m.The existing air intake damper and air handling unit which is used in the movie theater is shown in the figure 2.


4.1 Actual Data of Air-Conditioned Movie theater Located in TirunelveliTotal Number of Occupants = 600 personsInside Dry bulb temperature = 250cInside Relative Humidity = 50%DBT of the ventilation air = 350cWBT of the ventilation air = 220cTotal load = 55 tonsCOP of the system = 2.5Data are collected based on previous one year tickets sales.Theater operates ¾th of full occupant = 3 hours /day (average) Theater operates ½ndof full occupant = 5 hours/day (average)Theater operates ¼th of full occupant = 3 hours /day (average)

4.2 Theoretical CalculationsThermostat is used to control the variable load on the compressor due to the occupant variation, by measuring the temperature inside the movie theater. But the only other way to reduce the ventilation air load is by DCV technique. The total load distribution shows that 22 tons are due to occupation load and 15 tons are due to ventilation air load as shown in the figure 3. From this distribution, it is inferred that the ventilation air load of 27% and occupant’s load of 41% takes the 68% of the total load as shown in the figure 4.4.2.1 Condition of Air at Full Occupancy Level or Design ConditionSpecific humidity, enthalpy and temperature are calculated based per kg of air under circulation. Out of per kg 30 % of air constituted by the ventilation air (m1) and 70% by the re-circulation air (m2) as per standards.

Specific humidity W1=0.3 (11.8) + 0.7(9.5) =10.19 grams/kg of dry air Enthalpy h1=0.3 (63) + 0.7(48) = 52.5 kJ /kg of dry air Temperature T1=0.3 (35) +0.7 (25)

=28 0c

Total heat load to be removed Q1 = ma1 (“H) kW. Mass flow Rate of air required at full load ma1 =12.42 Kg/sFigure 6 shows that the enthalpy, temperature and specific humidity of the air entering into the compressor at different occupant level by applying the DCV technique. It is concluded that the temperature, enthalpy and specific humidity of the mixed air entering in to the air handling unit is reduced as mass flow rate of ventilation air is reduced depending upon the number of occupants inside the theater.4.3 Loads at Different Occupancy Conditions

As the mass flow rate of ventilation air varies depending upon occupancy level in the air conditioned space, the load in ton of refrigeration also changes. They are found out at different load conditions.

4.3.1 Load at Full occupant level Air Damper for ventilation is opened fully for full occupation. Q = r x V x (h1 – h2)

Q = 1.12 x 3.1296 x (63 –48)

Load due to ventilation air at full occupant level =15 tonsSimilarly the ventilation air loads for different occupant levels are calculated and illustrated in the figure 7.4.4 ENERGY CONSERVATIONBy implementing the DCV technique in a movie theater, the total load is reduced according to the occupancy level in the system.

4.4.1 Load saved during ¾ of the total occupant levelSaving of load due to the reduction of ventilation air = Total ventilation load- Ventilation load at ¾th of total occupant. Savings on ventilation air load =2.29 tonsPower saved on the compressor = 8.05/2.5=3.22 kW. Based on collected data the theater operates ¾th of total occupancy at 3 hours daily. So savings by optimal usage of the compressor is calculated as below.


The energy savings per day = 3.22 x 3

= 9.66 kW-hr/day

The energy savings per year = 289.8 x 12

= 3477.6 kW-hr /year

The cost per one unit = INR 5

Cost savings per year = INR 17388 /year

Similarly the load saved for other occupant levels are calculated. Figure 8 depicts that the savings of load increases as the ventilation air decreased from the design occupancy level.

5. THE CO2 BASED DCV CONTROL METHOD

The CO2 based DCV system may be installed in the air conditioned movie theater hall to ensure sufficient quantity of fresh air to cater for the variation of occupancy in the theater hall. CO2 sensor can be installed in the air stream of the return duct to monitor the concentration of the CO2 in the air and regulate the position of the damper opening according to the number of occupants inside the hall. The programmable logic controller (PLC) controls the position of the damper based on the signal coming out from the CO2 sensor. The figure 9 illustrates schematic diagram of the setup. This method will provide sufficient quantity of air stream to the occupants inside the hall, it will ensure better Indoor air quality (IAQ) and conserve energy in the air conditioning system because it reduces load on the compressor unit.

5.1 Occupancy Based Detection Algorithm

Assuming that people are the only source of carbon dioxide for the single room ventilated only with outdoor air the balance can be described by equation (1):(1)where M is the mass of air in the room, CR is the CO2 concentration in the occupied zone,e0 is the average CO2 generation rate of an occupant, a is the occupant activity level in the space, p is the number of persons, COA is the CO2 concentration in the supply air, VOA is the rate of outdoor air used to ventilate the

room, åAc is the air change effectiveness. The simplest occupation algorithm is based on the assumption that CO2 concentration reached the state of equilibirium.steady state algorithm was developed assuming that derivativ = 0.

Using the measurement data of out door air flow rate and both outdoor and indoor CO2 concentrations, the actual occupancy of a space in moment i can be estimated using

Equation (2):

(2)

Knowing the number of occupants, estimated in the above algorithm, the mass stream of outdoor air for next step i+1 can be adjusted according to equation (3):

mi+1OA = MIN ( MAX(Pi . Rp ; V MIN ); V MAX ) (3)

This algorithm can be used in the PLC unit to calculate the number of occupants for controlling the ventilation rate. The control system allows changing the ventilation rate by adjusting the damper position according to demands .on the other hand, the control system itself creates lower limitations for ventilation rate (V MIN of 10 % of the maximum ventilation rate) according to the standard ASHRAE 62-2001 when there is no occupant inside the hall. The algorithm can be programmed in the PLC unit by using the software Step 7.0 micro win .The simulation of the PLC program is shown in the figure 10 and 11.

6 RESULTS AND DISCUSSIONS

The data is measured over a particular period of time in the movie theater located in thirunelveli. By using this data the amount of energy saved and also the amount of cost saved per year is calculated. The investment required to achieve this savings is taken from the suppliers of CO2 sensor and motorized damper, the cost requirement of CO2 sensor with motorized damper is INR 20,000 /unit. The theater need 2 units of CO2 sensor with motorized damper .The total saving is calculated by summing the savings at different occupant level studied over a particular period of time.


7 CONCLUSIONS

In this paper, the data’s collected over particular period of time in the movie theater and the DCV methods for energy conservation are reported. Based on the collected data’s the amount of energy conservation is calculated and presented.

CO2 based DCV approach can offer significant energy savings in the air-conditioned movie theater, this also rectified the over ventilation problem. With the implementation of DCV technique around 40.8% of the total energy utilized in the air conditioning equipment was saved. The pay back period for the investment was also less than 4 months. In this method air intake damper, often subjected to maladjustment or arbitrary adjustment over time can be controlled automatically avoiding accidental and costly over or under ventilation. It is envisaged that lot of energy conservation is possible in the large air-conditioned system where the occupant’s levels varying widely. The future work will be carried out to find the DCV strategy using CO2 and non occupant related indoor air pollutant level as control signal and to find the other alternative approach to conserve the air conditioning system while maintaining better indoor air quality.

REFERENCES

1. Mohsin A. Alalawi and Moncef Krarti , “Experimental Evaluation of CO2-Based Demand –Controlled Ventilation Strategies”, ASHRAE Journal ,pp 307-316

2. Michael Kintner –Meyer and Martin Burns, “Energy –Related Information Services”, ASHRAE Journal, August 2001,pp 49 – 56.

3. Mike Schell and Dan in-Hout, “Demand Control Ventilation Using CO2” ASHRAE Journal, Feb 2001, pp 18-29.

4. K.A.Papakonstantinou,C.T.Kiranoudis et.al.,Numerical simulation of CO2 dispersion in an auditorium, energy and Buildings,2002,pp 245-250

5. J. Sowa, “Comparison of occupancy detection algorithms, methods of signal filtration and types of requirements expression for CO2 –based DCV systems”, Healthy Buildings, 2003, pp 574-579.

6. KWD Cheong, E Djunaedy et al., “Thermal Comfort study of a lecture theater environment-a case study approach”, Building and Environment, 2003, Vol 38, pp 63-73.

7. M.Mysen and J.P.Rydock, “Demand controlled ventilation for office cubicles-can it be profitable?” , Energy and Buildings, 2003, pp 657-662.

8. Thomas M.Lawrence ,James E. Braun , “Evaluation of simplified models for predicting CO2 concentrations in small commercial buildings”, Building and Environment ,Elsevier ,2005, pp:35-45.

9. C.Y.H.Chao and J.S.Hu, “Development of dual–mode Demand control ventilation strategy for indoor air Quality control and energy saving”, Building and Environment , Elsevier ,2004,pp 385-397. 10.

Moshen soleimani-moheseni,” A study of demand – controlled ventilation (DCV)and constant air volume (CAV) systems”, Healthy Buildings,2003,pp 392-397.

11. John Murphy ,”A breath of fresh air, Using CO2 for Demand–Controlled Ventilation”, Engineers Newsletter 2002,Volume 31,No.3,pp 1 -12

12. A.K.Persily and W.S. Dols , “Air Change effectiveness measurements in two modern office buildings”, Indoor Air,1994,Vol 4,pp 40-55

13. Shengwei Wang and Xinhua Xu ,”A robust control strategy for combining DCV control with economizer control”, Energy conservation and Management,2002,pp 2569-88

14. Mike B.Schell et.al, “Application of CO2 based demand controlled ventilation using ASHRAE standard 62: Optimizing energy use and ventilation”, ASHRAE Transactions, pp1213-1220.

15. Vitalijus Pavlovas, “Demand controlled ventilation a case study for existing Swedish multifamily buildings”, Building and Environment, 2004, pp 1029-1034.

16. B.F.Warren and N.C.Harper, “Demand controlled ventilation by room CO2 Concentration: a comparison of simulated energy savings in an auditorium space”, Energy and Buildings, 1990, pp 87-96.

17. William J.Fisk, Anibal, T .De Almedia, “Sensor–based Demand controlled Ventilation: Review”, Building and Environment, 1998, pp 35-45.

18. Warren BF, “Energy saving in buildings by control of ventilation as a function of indoor carbon dioxide concentration.” Building services Engineering Research & Technology ,1982,pp 4-12

19. ASHRAE,standard 62-1989, “ventilation for acceptable indoor air quality” American Society of Heating Refrigerating and Air-Conditioning Engineers ,Atlanta ,1989.

NOMENCLATURE


T1 = Dry bulb temperature of the ventilation air (o C)

T2 = Dry bulb temperature of the re circulated air (o C)

T3 = Dry bulb temperature of the mixed air (o C)

m1 = mass flow rate of ventilation air (Kg/s)

m2 = mass flow rate of re circulated air (Kg/s)

m3 = mass flow rate of mixed air (Kg/s) =m1+m2

W1 = Specific humidity of ventilation air (gram of water vapour /Kg of dry air)

W2 = Specific humidity of re circulated air (gram water vapour /Kg of dry air) W3 = Specific humidity of mixture air gram water vapour /Kg of dry air

h1 = enthalpy of fresh air KJ/Kg of dry air

h2 = enthalpy of re circulated air KJ/Kg of dry air h3 = enthalpy of mixture air KJ/Kg of dry a

Q = load due to ventilation air at full occupant condition (tons)

r = Density of air (Kg/m3)

v =Volume flow rate of ventilation air at full opening of damper (3.2196m3/s)

Figure 1 Rates of CO2 generation for various activities

(a)

(b)


Figure 2 (a) Existing damper in the movie theater (b) Air Handling Unit inside the movie theater

(AHU)

Figure 3 Total load distributions in tons

Figure 4 The distribution of total load for different factors in percentage.

Ventilation air

W1,h1,T1

Re circulated air

W2, h2, T2

Conditioned air

Mixed air

W3, h3, T3

AHU

Theater

Figure 5 Mixing of Ventilation air and re circulated air

Figure 6 The condition of air entering in to the Air handling unit

Figure 7 Ventilation air load at different occupant’s level.

1: full occupants, 2: ¾th of full Occupants, 3: ½nd of full occupants,

4:¼th of full occupants

Table 1 Savings of energy and cost by applying DCV technique


Figure 8 Savings of load by reducing the ventilation air rate at differen Occupant’s load

Figure 9 Schematic Diagram of the setup


Figure 10 CO2 concentration Vs Number of persons

Figure 11 CO2 concentration Vs Angle of rotation

RDF - A TECHNOLOGY FOR DESCRIBING WEB RESOURCES

C.SUBHASHINI 1 , S.PREMALATHA2 , A.NAGARATHINAM3 1. Lecturer / MCA, PES School of Engineering, Bangalore- 560 100, Email:[email protected]

2. Lecturer/MCA,A.V.C College of Engineering, Mayiladuthurai, Email: [email protected]

3. Senior Lecturer/MCA, A.V.C College of Engineering, Mayiladuthurai.Email:[email protected]

ABSTRACT:

The Resource Description Framework (RDF) is an infrastructure that enables the encoding, exchange and reuse of structured metadata. RDF is a family of World Wide Web Consortium (W3C) specifications for describing Web resources, such as the title, author, modification date, content, and copyright information of a Web page. It provides a universal way to describe information so that it can be read and understood by computer applications. RDF sets rules that will enable humans and machines to make and understand infinite number of statements whose subjects and objects are resources. RDF documents are written in XML. The XML language used by RDF is called RDF/XML.

Key words: XML , web services ,RDF..

1. INTRODUCTION:RDF was designed in 1999 as a standard on top of XML for encoding metadata—exactly, data about data. Metadata is, obviously, things like who authored a web page, what date a modification entry

was published, etc. Since then, and perhaps especially after the updated RDF specification in 2004, the scope of RDF has really evolved into something greater. The most important uses of RDF is not only encoding information about web resources, but also gives information about and


relations between things in the real world: people, places, concepts, etc.RDF is about the Semantic Web:(The Semantic Web is an evolving extension of the World Wide Web in which web content can be expressed not only in natural language, but also in a format that can be read and used by software agents, thus permitting them to find, share and integrate information more easily). RDF is a general solution for making statements that all machines and all humans can understand. RDF’s strength is that it can be understood everywhere, because RDF statements are interoperable – they are understood in the same way by all processors. RDF uses Web identifiers (URIs) to identify resources. RDF describes resources with properties and property values.

1.1. RDF Resource, Property, and Property Value

RDF identifies resources using Web identifiers (URIs), and describes resources with properties and property values.

· A Resource is anything that can have a URI, such as “http://www.w3.org/TR/2000/CR-rdf-schema-20000327 [RDFS] “

· A Property is a Resource that has a name, such as “author” or “homepage”

· A Property value is the value of a Property, such as “http://www.w3.org”.

The following RDF document could describe the resource, “http://www.w3.org/TR/2000/CR-rdf-schema - 20000327 [RDFS] “

<? xml version=”1.0"?><RDF> <Description about=”http://www.w3.org/TR/2000/ CR-rdf-schema - 20000327 [RDFS] “> <author>Tim Berners - Lee</author><homepage>http://www.w3.org</homepage> </Description></RDF>

1.2. RDF Statements

The RDF Statement is often called a triple because it has three parts namely a Resource, a Property, and a Property value. These three things are known as subject, predicate and object of a statement.For example, consider the following Statement:

Statement: “The homepage of “http://www.w3.org/TR/2000/CR-rdf-schema - 20000327 [RDFS] “is “http://www.w3.org”. Here,

· The subject of the statement above is: http://www.w3.org/TR/2000/CR-rdf-schema - 20000327 [RDFS] “

· The predicate is: homepage

· The object is: “http://www.w3.org”.

1.3. Core SpecificationRDF builds on two companion

specifications.

The model and syntax specification defines the triple in which RDF statements are made; the schema specification describes how to use RDF to build RDF vocabularies (collections of resources that can be used as predicates – the verbs in RDF statements).

2. The RDF Data ModelRDF provides a model for describing resources. Resources have properties (attributes or characteristics). RDF defines a resource as any object that is uniquely identifiable by an Uniform Resource Identifier (URI). The properties associated with resources are identified by property-types, and property-types have corresponding values. Property-types express the relationships of values associated with resources. In RDF, values may be atomic in nature (text strings, numbers, etc.) or other resources, which in turn may have their own properties. A collection of these properties that refers to the same resource is called a description. At the core of RDF is a syntax-independent model for representing resources and their corresponding descriptions.The following graphic (Figure 1) illustrates a generic RDF description.


Figure 1

The application and use of the RDF data model can be illustrated by concrete examples. Consider the following statements:

1. “The author of Document 1 is John Smith”

2. “John Smith is the author of Document 1”

To humans, these statements convey the same meaning (that is, John Smith is the author of a

particular document). To a machine, however, these are completely different strings. Whereas humans are extremely adept at extracting meaning from differing syntactic constructs, machines remain grossly inept. Using a triadic model of resources, property-types andcorresponding values, RDF attempts to provide an unambiguous method of expressing semantics in a machine-readable encoding.

RDF provides a mechanism for associating properties with resources. So, before anything about Document 1 can be said, the data model requires the declaration of a resource representing Document 1. Thus, the data model corresponding to the statement “the author of Document 1 is John Smith” has a single resource Document 1, a property-type of author and a corresponding value of John Smith. To distinguish characteristics of the data model, the RDF Model and Syntax specification represents the relationships among resources, property-types, and values in a directed labeled graph. In this case, resources are identified as nodes, property-types are defined as directed label arcs, and string values are quoted. Given this representation, the data model corresponding to the statement is graphically expressed as (Figure 2):

If additional descriptive information regarding the author were desired, e.g., the author’s email address and affiliation, an elaboration on the previous example would be required. In this case, descriptive information about John Smith is desired. As was discussed in the first example, before descriptive properties can be expressed about the person John Smith, there needs to be a unique identifiable resource representing him. Given the directed label graph notation in the previous example, the data model corresponding to this description is graphically represented as (Figure 3):


Figure 3

In this case, “John Smith” the string is replaced by a uniquely identified resource denoted by Author_001 with the associated property-types of name, email and affiliation. The use of unique identifiers for resources allows for the unambiguous association of properties. This is an important point, as the person John Smith may be the value of several different property-types. John Smith may be the author of Document 1, but also may be the value of a particular company describing the set of current employees. The unambiguous identification of resources provides for the reuse of explicit, descriptive informationIn the previous example the unique identifiable resource for the author was created, but not for the author’s name,email or affiliation. The RDF model allows for the creation of resources at multiple levels. Concerning the representation of personal names, for example, the creation of a resource representing the author’s name could have additionally been described using “firstname”, “middlename” and “surname” property-types.

3. THE RDF SCHEMARDF Schemas are used to declare vocabularies, the sets of semantics property-types defined by a particular community. RDF schemas define the valid properties in a given RDF description, as well as any characteristics or restrictions of the property-type

values themselves. The XML namespace mechanism serves to identify RDF Schemas.

4. CONCLUSIONS

The World Wide Web affords unprecedented access to distributed information. Metadata improves access to this information and RDF is a W3C proposed standard for defining the architecture necessary for supporting web metadata. RDF is an application of XML that imposes needed structural constraints to provide unambiguous methods of expressing semantics for the consistent encoding, exchange, and machine processing of metadata. RDF additionally, provides means for publishing both a

.

human-readable and a machine-processable vocabularies designed to encourage the exchange, use and extension of metadata semantics among disparate information communities.

REFERENCES:1. Frank P.Coyle , “XML,Web services ,and

the data revolution “ –Pearson Education ,2006

2. Mikael HillBrog ,”Wireless XML – Developer’s guide”, - TataMcGrawHill ,2002.

3. Dave Mercer , “ XML – A Beginner’s Guide”, - TataMcGrawHill.

4. “Preofessional XML “ ,Wrox Publications.

5. James McGovern, Sameer Tyagi , Michae E.Stevens , Sunil Mathew , “ Java Web Services Architecture “- Morgan Kaufmann publishers.


6. Alexander Nakshimovsky , Tom Myers ,” Professional Java XML Programming with Servlets and JSP “ – Wrox Publications.

7. Ramesh Nagappan , Robert Skocylas and Rima Patel Sriganesh, “ Developing Java Web services “, Wiley Publishers Inc., 2004.

8. Sandeep Chatterjee , James Webber , “ Developing Enterprise Web Services “, Pearson Education,2004.

9. www.w3.org

10. www.w3schools.com

11. http://en.wikipedia.org/wiki

IMPACT STRENGTH AND WORKABILTY BEHAVIOUR OF GLASS

FIBRE CONCRETE

CHANDRAMOULI 1 DR. SESHADRI SEKHAR. T2 DR.SRINIVASA RAO P3 DR.P.SRAVANA4


1.Principal , PRIYADARSHINI College of Engineering and Technology for Women. Guntur, Andhra Pradesh. E-Mail: [email protected]

2. Principal , Dr .S.G.I.T , MARKAPURN, Andhra Pradesh E .Mail : [email protected] Res Addr : Flat No 101, Bhavana Towers ,Haripuri colony Road No.3, Besides Alkapuri Darga, Hyderabad, India. Mobile No :

96526536663.4. Professor Dept of Civil Engineering J.N.T.U College of Engineering. Hyderabad. A.P. India. Mail:

[email protected] ABSTRACT:Cement based materials are quasi-brittle and are known to exhibit a highly stress-rate sensitive behaviour . In structures that are subjected to impact forces , this causes concern in two ways: first the brittleness may result in catastopic failure without warning and second, the properties of concrete during such events may be very different from those measured in standardized quasistatic tests. Unfortunately, there are no standardised tests available for testing concrete under impact loading and there is significant confusion as to what constitutes an appropriate test.This paper concentrates mainly on studying the properties like Impact strength of glass fibre reinforced concrete using Alkali-Resistant Glass Fibres in various proportions and comparing the same with the ordinary concrete mixes of grade M 20 and M 50.Key Words : impact strength , Alkaline Glass fibres .

INTRODUCTION

The present day world is witnessing the construction of very challenging and difficult Civil Engineering structures. Quite often, concrete being the most important and widely used material is called upon to possess very high strength and sufficient workability properties. Efforts are being made in the field of Concrete Technology to develop such concretes with special characteristics. Researchers all over the world are attempting to develop high performance concretes by using fibres and other admixtures in concrete upto certain properties.

In the view of the global sustainable developments, it is imperative that fibres like Glass, carbon, Poly propylene and aramid fibers provide improvements in Tensile strength, Fatigue characteristics, durability, shrinkage characteristics, impact, cavitation, erosion resistance and serviceability of concrete.

LITERATURE REVIEW:Bindiganavile et.al 1 examined two major issues related to impact loading on plain and fibre reinforced concrete.They observed that for cement based materials , the measured impact response is highly dependent on the characteristics of the drop weight impact machine used for testing.. The pulse duration was found to depend upon the drop height, with greater drop heights leading to shorter pulses. Results appear to be far less sensitive to the mass of the hammer than to the drop height. Yi Chen et.al 2

presented accelerated aging test results of a durability study on fibre reinforced polymer reinforcing bars for concrete structures. The results showed that when exposed to simulated environment significant strength loss resulted from the accelerated exposure of both bare and embedded. Glass Fibre reinforced polymer bars, including bond strength, especially for solutions at 600c. In contrast carbon fibre reinforced polymer bars displays excellent durability performance. Rajan.et.al 3 discussed some of the issues

relating the application of fibre reinforced polymers for repairing corrosion damaged structures located in tidal water. The experience indicated that the new technology is promising and may be competitive in applications where corrosion damage can patch repaired so that there is no need for reforming the cross section. Biswarup saikia et.al 4 studied the enhancement in ductility of the beam in which fibres

are added. They observed that the beams reinforced with GFRP bars indicate comparable levels of ductility and structural performance. Dr.P.Srinivasa Rao et.al 5 discussed about strength and durability properties of M 20 and M 30 grade of glass fibre reinforced concrete. They concluded that improved behaviour of glass fibre reinforced concretes over ordinary concrete. Dr.P.Srinivasa Rao et.al 6


discussed about the effect of glass fibres on durability properties of glass fibre reinforced concrete.

EXPERIMENTAL PROGRAMME

The objectives of the experimental study that was conducted are given below.

v To study the Impact resistance of ordinary Concrete and Glass Fibre Concrete at 28, 56, 90 and 180 days for M 20 to M 50 Grades of Concrete.

MATERIALS

Cement

Ordinary Portland cement of 53 grades available in local market is used in the investigation. The Cement used has been tested for various proportions as per IS 4031 – 1988 and found to be confirmed to various specifications of are 12269 – 1987. The specific gravity was 2.96 and fineness was 3200cm2/gm.

Coarse Aggregate

Crushed angular granite metal of 20 mm size from a local source was used as coarse aggregate. The specific gravity of 2.71 and fineness modulus 7.13 was used.

Fine Aggregate

River sand was used as fine aggregate. The specific gravity of 2.60 and fineness modulus 3.25 was used in the investigation.

Glass Fibres

The glass fibres are of Cem-FIL Anti – Crack HD with Modulus of Elasticity 72 GPA, Filament diameter 14 microns, Specific Gravity 2.68, length 12mm and having the aspect ratio of 857.1. The number of fibres per 1 kg is 212 million.

PREPARATION OF SPECIMEN

The gauge length of 40 cms has been accurately marked on each specimen of size 100 x100x 500 mm . The specimen has been placed in impact testing machine and it has been fixed to machine by using G-clamps. The distance which measures a swing of 100

has been marked from the centre on the angle section.

TESTING PROCEDURE

The impact hammer was positioned at the mark for swung upto the mark which measured 100 swing by a right angle hook and was released for applying the impact load on the specimen. The blows were repeated in above manner till the 1 st crack. The crack propagation for each blow after 1 st crack was marked on the specimen for clarity.

The number of blows required for the crack propagation from one edge to other on the tension face was noted. The experiment was continued till the spalling of mortar occurring on the compression face of the specimen. The number of blows required for spalling of mortar for the full width on compression face was noted and the propagation was marked on the specimen for clarity. The details are tabulated in table 1.0

DISCUSSION OF TEST RESULTS

Effect of Glass fibre on Bleeding

On the basis of the experimental study it was concluded that addition of Glass Fibres in concrete gives a reduction in bleeding. A reduction in bleeding improves the surface integrity of concrete, improves its homogeneity, and reduces the probability of cracks occurring where there is some restraint to settlement.

Impact strength of ordinary concrete and glass fibre concrete mixes

Table 2.0 gives the impact strength of various grades of ordinary concrete mixes of M 20, M 30, M 40 and M 50. These values are observed to be varied from 259 to 356 N/mm2 for 28 days, 284 to 392

N/mm2 for 56 days, 297 to 421 N/mm2 for 90 days, 305 to 432 N/mm2 for 180 days.

Table 2.0 gives the impact strength of various grades of Glass fibre concrete mixes of M 20, M 30, M 40 and M 50. These values are observed to be varied from 295 to 407 N/mm2 for 28 days, 321 to


445 N/mm2 for 56 days, 345 to 488 N/mm2 for 90 days, and 352 to 492 N/mm2 for 180 days.

The variation of impact strength of both ordinary concrete and glass fibre concrete mixes are given in Fig 1.0.

Variation of impact strength of the ordinary concrete and glass fibre concrete mixes compared with 28 days impact strength

Table 3.0 gives the impact strength of various grades of ordinary concrete mixes of M 20, M 30, M 40 and M 50 compared with 28 days impact strength. These values are observed to be 15 to 18 % for 90 days and 18 to 22 % for 180 days.

Table 3.0 gives the impact strength of various grades of glass fibre concrete mixes of M 20, M 30, M 40 and M 50 compared with 28 days impact strength . These values are observed to be 15 to 20 % for 90 days and 16 to 22 % for 180 days.

Variation of impact strength of glass fibre ordinary concrete in comparison with ordinary concrete mixes

Table 4.0.gives the increase in impact strength of various grades of glass fibre ordinary concrete mixes of M 20, M 30, M 40 and M 50 in comparison with ordinary concrete mixes at 28, 56, 90 and 180 days. These values are observed to be 12 % to 17 % .

The variation of impact strength of glass fibre concrete mixes over ordinary concrete mixes are given in Fig 2..0.It is Observed from the failure pattern of both ordinary concrete and glass fibre concrete specimens that all the cracks are brittle failure cracks .

CONCLUSIONS

1. A reduction in bleeding is observed by addition of glass fibre in the glass fibre ordinary concrete mixes.2 .A reduction in bleeding improves the surface integrity of concrete, improves its homogeneity, and reduces the probability of cracks.

3 The increase in impact strength of glass fibre ordinary concrete mixes is observed to be 15 to 20 % for 90 days and 18 to 22 % for 180 days compared with 28 days impact strength.4.The increase in impact strength of glass fibre ordinary concrete mixes at 28, 56, 90, 180 days is observed to be 12 % to 17 % when compared with ordinary concrete mixes .5. All the cracks observed in glass fibre ordinary concrete mixes on impact specimens are brittle Failure cracks.6. The increase in impact strength of ordinary concrete mixes is observed to be 15 to 18 % for 90 days and 18 to 22 % for 180 days compared with 28 days impact strength.7. All the cracks observed in ordinary concrete mixes on impact specimens are brittle failure cracks.

REFERRENCES

1.. Bindiganavile.V and Banthia (2002) “Some studies on the Impact Response of Fibre Reinforced Concrete”, ICI journal ,Oct –Dec , PP 23-28.2. Yi Chen, Julio F. Davalos, Indrajit Ray and Hyeong Yeol Kim “Accelarated aging tests for evaluations of durability performance of FRP reinforcing bars for concrete Structures” Elsevier Journal on composite Structures, 2005, pp 1-7.3. Rajan Sen and Gray Mullins “Developments in Underwater FRP Repair”, Proceedings of ICFRC International conference on Fibre Composites, High Performance Concrete and Smart Materials, 8-10 January 2004, Chennai, India pp 387-395.

4. Biswarup Saikia, Phanindra Kumar, Job Thomas, Nanjunda Rao. K.S. and Anatha Swamy “A study on Flexural Performance of Beams Reinforced with GFRP Bars”, Proceedings of ICFRC International conference on Fibre Composites, High Performance Concretes and Smart Materials, 8-10 January 2004, Chennai, India pp 465-474.

5. Dr. Srinivasa Rao. P and Seshadri Sekhar. T “Strength and Durability properties of glass fibre reinforced concrete” Proceedings of International Conference on Recent Advances in Concrete and Construction Technology, 07-09 Dec 2005, SRMIST, India pp P 43-50


6. Dr. Srinivasa Rao. P and Seshadri Sekhar. T “Strength and Durability properties of glass fibre reinforced concrete” Proceedings of International

Conference ACECON2005, 22-25 Sept 2005, ICI-Asian Conference Mumbai, India pp 67-72.

Table 3.0 Percentage increase in Impact strength of Ordinary Concrete and Glass Fibre Concrete Mixes In Comparison with 28 days Strength


Table No 4.0 Percentage increase in Impact Strength of Glass Fibre concrete in Comparison with ordinary concrete Mixes


Biography of the Author:Dr P. Srinivasa Rao

B.Tech., M.Tech., Ph.D., M.I.E ., M.I.S.T.E.,Professor and HeadSpecialized in structural engineering. Research interests are Concrete Technology, Structural Design, High Performance Concrete, Prefabricating Structures, Special Concretes and use of Micro Silica, Fly Ash in Building Materials. He has been associated with a number of Design projects, for number of organizations and involved as a key person in Quality control and Mix Designs. Has 22 years of academic, research and industrial experience published over 100 research papers. He guided 75 M.Tech projects. Guiding 15 Ph.D students delivered invited lecturers in other organizations and institutions. Member of ISTE, Member of ICI and Member of Institute of Engineers.

Dr. Seshadri Sekhar.T

B.Tech., M.Tech. M.S.,Ph.D ., M.I.S.T.E., F.I.E.T.E ., F.I.EPrincipal , Dr SGIET ,MarkapurRes Addr : Flat No 101, Bhavana Towers , Haripuri colony Road No.3, Besides Alkapuri Darga, Hyderabad, Andhra Pradesh , India .

Mobile No : 9490052301.Specialized in structural engineering. Research interests are Concrete Technology, High Performance Concrete, , Special Concretes and Fly Ash in Building Materials.. Has 20 years of academic, research and industrial experience published over 65 research papers. He associated with 25 M.Tech projects.. Member of ISTE, Fellow Member of Institute of Engineers and Fellow Member of Institute of Electronics and Telecommunication Engineers

Dr . P.Sravana Associate Professor

B.Tech, M.Tech., Ph.D., M.I.S.T.ESpecialized in transportation engineering . Research interests are Special concretes , Pavement Design. She has been associated with a number of Design projects, for number of organizations and involved as a key person in Quality control , Mix Designs and Bitumen Emalution Tests . Has 12 years of academic, research and industrial experience published over 30 research papers.Chandra Mouli principalB.Tech, M.Tech., M.I.S.T.ESpecialized in Structural engineering . Research interests are Special concretes. He has been associated with a number of Design projectsfor number of organizations . Has 22 years of academic, research and industrial experience published over 06 research papers.


INTEGRATION OF AGILE SOFTWARE DEVELOPMENT WITH THE EXISTING ORGANIZATIONAL PRACTICES AND THEREBY

IMPROVING THE SOFTWARE PROCESS

N.GANESH1, S.THANGASAMY2

1 Research Scholar, Department of Computer Science and Engineering, Anna University, Coimbatore – 641 006, India

2 Professor and Dean, Department of Computer Science and Engineering, Kumaraguru College of Technology, Coimbatore – 641 006, India

[email protected]:

The software process improvement (SPI) has been widely studied with respect to traditional software development. The Agile software development has challenged the traditional ways of delivering software as it provides a very different approach to software development. As organizations have started adopting agile software development methodologies to be used along with the traditional methodologies, new challenges and opportunities for the Software Process Improvement are also emerging. The traditional Software Process Improvement methods often emphasize the continuous improvement of organizational software development processes, whereas the principles of agile software development focus on iterative adaptation and improvement of the activities of individual software development teams to increase effectiveness. This paper focuses on integrating agile software development and continuous improvement of existing organizational practices and the comparison of SPI Elements in Plan-Driven and Agile Software Development.

Keywords:Software process improvement, agile software development, iterative improvement process.

INTRODUCTION

A software process can be defined as .the sequence of steps required to develop or maintain software aiming at providing the technical and management framework for applying methods, tools, and people to the software task [10]. In traditional SPI methods and approaches [10], the organizational improvement played a vital role, due to the fact that the planning and control of the SPI initiatives are managed by the organizational stakeholders. The positive aspects of SPI methods and approaches include reducing time to-market, risks and costs, and increasing the productivity and quality in software development organizations [1]. However, various negative effects have also been encountered [22], e.g. regarding the cost-effectiveness of SPI initiatives, their actual effectiveness in improving the software development practices of organizations, the volume of the effort needed to implement SPI initiatives and the low speed at which visible and concrete results are achieved [2].

Nearly two-thirds of SPI traditional initiatives fail to achieve their intended goals [3].From the late-1990s onwards, agile software development principles and methodologies have been increasingly challenging the traditional view of software development. The principles of the agile manifesto [25] identify the central elements of agility and the methods that claim to be agile. One of the twelve principles states that agile teams should regularly reflect on their work in order to become more effective, while another principle addresses the self-organization of teams. The fundamentals of agile software development address the improvement and management of software development practices within individual teams. Agile software development provides a highly untraditional approach to SPI, in which the process improvement knowledge of software developers and software development teams is acknowledged and valued.


Two perspectives can be identified at the project level and at organizational level. The project level SPI examines the regular process adaptation activities in individual agile software development teams, emphasized in the principles of agile software development.

The ProblemThe main objective in this paper is to implement the Software Process Improvement in an individual agile software development teams and how it could be integrated with the traditional continuous organizational Software Process Improvement activities.

PLAN-DRIVEN APPROACH OF TRADITIONAL METHODS

As the names suggest, a common feature for the plan-driven process models is their emphasis on defining the scope, schedule, and costs of the project upfront including, for example, an early fixing stage and extensive documentation of the end product requirements. One common characteristic could also be the recurrence of the software development phases only once during the development process.The two-step process model of code-and-fix, used in the early days of software development, resulted in difficulties that necessitated explicit sequencing of the phases of software development. The need to design prior to coding, to define requirements prior to design and the need for early preparation for testing and modification were recognized. This model evolved from the problems caused by the increasing size of software programs, which could not be handled by a single programmer.Another aspect of the plan-driven approach to software development is the concept of holding the certainty. The prime motive of requirements definition and locking of the project scope, leads to contracts and decisions based on estimations of costs, time and resources. The success of software projects is often measured against these estimates from the viewpoint of both an acquirer and supplier, to agree fixed costs, scope and schedule for the project up-front. However, the quest for certainty definitely affects the quality of the end product.The plan-driven models of software development should be applied in a dynamic way by repeating the phases or even the entire process, if necessary.

The purpose of the software process models was not to welcome changes during the development, but rather to try to fix factors, such as scope, time and money, up-front in order to eliminate change which was considered a risk factor.

Iterative Change-Driven modelIterative development refers to the overall lifecycle model in which the software is built in several iterations in sequence [5]. A single iteration can be considered as a mini-project in which the activities of requirements analysis, design, implementation and testing are conducted as given in [24] in order to produce a subset of the final system, often resulting in internal iteration release. An iteration release has been defined as .a stable, integrated and tested partially complete system [5]. Incremental development involves adding functionality to a system over several releases, i.e., a repeated delivery of a system into the market or production. Thus, one incremental delivery may be composed of several iterations. A development approach where the system is developed in several iterations is called iterative and incremental development [5]. The spiral model consists of four iteratively repeatable steps: 1) determining the objectives, alternatives, and constraints, 2) evaluating alternatives, and identifying and resolving risks, 3) development and verification, and 4) planning the next phase. [4] Defined the spiral model as a risk-driven approach for software development.

Agile Software Development and its present scenario

The emergence of agile methodologies has begun in the mid-1990s, when software methodologies and techniques such as eXtreme Programming (XP) [6], Scrum [7], eXtreme testing [8], Crystal Family of Methodologies [9], Dynamic Systems Development Method (DSDM) [11], Adaptive Software Development (ASD) [12], and Feature-Driven Development (FDD) [13] began to emerge.The core features of agile method have been further specified as an iterative development of several cycles, incremental development, ability and permitting the teams to self-organize and determine the management of work, and emergence of processes, principles, and work structures during the project [14]. In agile software development, the uncertainty of schedule, scope and budget of any software development project can be considered as a baseline assumption.


Currently, the agile methodologies are adapted to organizations with established and mature plan-driven processes [14].

Ways to Improve the Software Process for the Traditional elements

The aim of improving the software process in organizations is to provide a return on investment. The return on investment has been reported for improved efficiency of the development process and reduction of total software costs, increased quality of the end product, higher predictability of cost and schedule, and increased level of reuse.The focus on quality is not only to get better quality but also lower cost and improvement of competitive position. [15].One of the characteristics of improving the software process is its emphasis on the continuous improvement of organizational software development processes in terms of performance, stability, compliance, and capability.Both the traditional method and the agile method are classified based on six different elements for improving the software process. They are the organizational models, standard process models, tailoring the process, deploying the process, measuring process methods.

Organizational Models

The organizational framework uses continuous, Top-down approach such as Quality Improvement Paradigm (QIP). The QIP provides process-focused mechanisms to improve the quality and productivity of software. The software process conducts the corporate level activities of improvement include the definition of current status, setting of goals and scheme of improvement, and analysis and storing of experiences and feedback resulting from the project learning cycle.

Tailoring the Process

The traditional approach for process tailoring is the definition of the project specific software process. It has been suggested that it is done either up-front, while included as a part of a project plan, or conducted as dynamic tailoring when needed during the software development project [16].

Process Deployment

The deployment may include such activities as piloting the processes, methods, and tools that are identified as potential solutions for the existing goals and problems, and evaluating their effect on the software development. An organization may adopt a big bang or an evolutionary approach for deploying new processes, practices and tools. The evolutionary requires repetitive activation of different deployment-related mechanisms and also for creating resentment among practitioners due to the constant changes they are requested to implement. The big bang approach has been suggested in cases where the business is no longer capable of achieving its purpose, where the external circumstances are changing rapidly in which the motivation and activity of its members need to be improved.

Measuring Process Methods

Various measurement mechanisms provide quantitative support from understanding the current status to planning improvements and assessing the rate and level of learning to ensure that gains. The software processes as well as their outputs have measurable attributes which can be observed to describe the quality, quantity, cost, and timelines of the results produced. Software metrics are used for measuring specific attributes of a software product or a software development process. This is done in order to enhance decision-making by drawing up estimates, tracking the progress, and evaluating the state of quality. The measurements also serve for analyzing defects, and validating the best practices for development.The GQM (Goal-Question-Metric) method provides a goal-based approach for defining metrics that, in return, constitute answers to the underlying questions and goals. Statistical Process Control (SPC) aims at stable processes with predictable results using statistical software process management. Automated tools such as SLIM - metrics and PROM (PRO Metrics) have been proposed for collecting and analyzing metrics data.

Ways to Improve the Software Process for the Agile Software Development

When considering the relationship between agile software development and SPI, there are three principles, in particular, of the agile manifesto that deserve attention:


the valuing of individuals and interactions over processes and tools, the principle that encourages regular reflection by software development teams in order to become more effective, and the self-organization of software development teams. Taking regular improvement within project teams as one of the twelve principles of agile software development highlights the importance of continuous improvement also in the agile software development context. In order to welcome changes throughout the agile software development project, whether they concern product requirements or technical aspects, the software process with its practices, methods, and tools must be able to adapt to the specific context while also to respond to the changes when needed.

Organizational ModelsIn [14] have addressed the difficulty of scaling up and integrating agile methodologies into traditional, top down systems of development organizations has been identified as a major challenge in implementing the agile processes in traditional development organizations. The difficulty of adopting agile methods in the CMM [23] based Companies is that they lack the visibility on how to take advantage of the existing best practices of an organization in transition towards agile methods, that has been identified as one of the major obstacles in the co-existence of the agile and traditional approaches [17].

Standard Process Models and their assessments

The problem is to determine the ways to merge the agile processes with standard industrial processes. The pitfalls of some of the agile methodologies, such as XP or Scrum, have been reported, along with a lack of practices to support the commitment of management to the defined software development process, and also regarding the setting up and staffing of an independent quality assurance group has been studied by [18]. In addition, the degree of documentation and the infrastructure required by current process standards for lower-level certification are issues of concern [14].The agile assessment method provides a lightweight approach for assessment to identify and adopt the most suitable agile methods amongst the existing organizational practices [19].

Tailoring the Process

Agile specific methods are needed to tailor agile software development practices within individual projects and within the entire organization.

The tailoring activity is broadly classified in to two groups. 1. Static tailoring - at the beginning of the project, 2. Dynamic tailoring – which deals with the Continuous process adaptation, which takes place throughout the life-cycle of an individual project.Three different strategies for process tailoring in the context of agile software development were suggested by [20]. using a comprehensive pattern based process framework for selecting appropriate elements at the beginning of each project which is a static tailoring approach, 2) defining a set of processes, as in Crystal, and selecting the best match for the project at hand with possible fine tuning which is a static tailoring approach, and 3) defining a tailored process for the project by blending ideas and techniques from best practices and local experience, which is a dynamic tailoring approach.An agile approach to tailoring will yield the following: 1) factoring a core set of process artifacts suitable for a wide range of developments, 2) identifying candidate patterns for core types of development as well as for contextual application of activities, artifacts, and guidelines, and this pattern-based approach for agile process tailoring also suggests that the reflection of project teams on the experiences of applying selected patterns should contribute to the organizational pattern catalogue.The objective of the agile project management is to define the approach the project team will use to deliver a product. [21]. The agile project management has the process and practice tailoring approach that includes both static tailoring at the beginning of the project anddynamic tailoring throughout the project, to be conducted even iteratively. One of the agile principles that has the regular team reflections of software developers in order to become more effective relate directly to the continuous and dynamic project-specific tailoring activity, whereby the organizational base process is iteratively tailored throughout the project by the software development team.The problem of traditional methods is that it lacks the mechanism to ensure subsequent utilization of the lessons learned from finished projects. The existing process adaptation methods do not specifically speak about the definition of goals in the process adaptation. In project-centered process adaptation, the link between the business goals with the project level software process improvement actions may not be of immediate concern from the viewpoint of an individual project team.

Deploying the Process


The deployment process consists of six steps: 1) assess the current state, 2) set (or revise) goals, 3) Identify risks,

4) plan process implementation, 5) execute process implementation, and 6) evaluate process implementation. The deployment steps identified in the model include decision-making on behalf of agile methods, selecting an approach for deployment, selecting methods to deploy, initial deployment, iterative adaptation of process, and ensuring the use of process knowledge in the future.

Measuring Process Methods

In eXtreme Programming it has been said that the process adaptation within software development teams requires mechanisms for assessing the effects of the changes [6]. Traditionally, metrics have been the main tool of such evaluation. But in the scrum approach for project management, if the correct things were measured the correct improvements can be made.

Similarities and Differences between Traditional SPI and the Agile SPI

The traditional SPI provides high predictability and stability using quantitatively monitored software development processes. On the contrary, the agile principles highlight the need for the software process to be flexible and will rapidly respond to the constant changes and context specific needs of software development. In Traditional SPI the goals are defined by planning, managing, and controlling the SPI initiatives. In the agile SPI, it is based on the experience and knowledge of software developers and by improving their daily working practices. Moreover, in the agile approach, the role of management is to organize and co-ordinate rather than to plan and execute. In Traditional SPI, the process tailoring is an activity in which the organizational software processes are modified for a specific project at the beginning of a project. On the contrary, In the agile process tailoring it has been characterized as on ongoing dynamic process in which the experience and face-to-face learning is the main source of adaptation. In traditional SPI the process measurement and metrics have been used to identify weaknesses in the organizational software development and verify the effects of the stakeholders. On the contrary, in the agile SPI, no quantitative mechanisms have been suggested.

CONCLUSION

In the recent days, the organizations are running both the traditional and agile software development projects simultaneously.

Hence, both the Traditional SPI and the agile SPI has to be balanced together in order to benefit from the strengths and thereby overcoming the weaknessesHigher customer satisfaction and a higher quality of end products can be obtained through continuous collaboration with the customer and incremental delivery of working software. Face-to-face communication and self-organization of teams will yield better software process improvement methods.

REFERENCES

[1]. Krasner, H. 1999. “The Payoff for Software Process Improvement: What it is and How to Get it. In: Elements of Software Process Assessment & Improvemen”t. El Emam, K. & Madhavji, N. H. (ed.). IEEE Computer Society: Los Alamitos, California. Pp. 113 - 130.

[2]. Dyba, T. 2000. “An Instrument for Measuring the Key Factors of Success in Software Process Improvement”. Empirical Software Engineering, Vol. 5, pp. 357 - 390.

[3]. Debou, C. 1999. “Goal-Based Software Process Improvement Planning. In: Better software practice for business benefit: Principles and experience”. Messnarz, R. & Tully, C. (ed.) IEEE Computer Society: Los Alamitos, Pp. 107 - 150

[4]. Boehm, B. 1988. “A Spiral Model of Software Development and Enhancement”, Vol. 21, 5 (5), May 1988, pp. 61 - 72.

[5]. Larman, C. 2004. “Agile and Iterative Development: A Manager’s Guide”. Pearson Education, Inc. Boston

[6]. Beck, K. 1999. “Embracing Change with Extreme Programming”. IEEE Computer, Vol. 32, 10 (10), pp. 70 -77.

[7]. Schwaber, C. & Fichera, R. 2005. “Corporate IT Leads The Second Wave of Agile Adoption”. Forrester Research, Inc. November 30, 2005

[8]. Jeffries, R. E. 1999. “eXtreme Testing: Why Aggressive Software Development Calls for Radical Testing Efforts”. Software Testing & Quality Engineering, Vol. March/April, pp. 23 -26.

[9]. Cockburn, A. 2005. “Crystal Clear: a Human-Powered Methodology for Small Teams”. Addison-Wesley

[10]. Humphrey, W. S. 1995. “A Discipline for Software Engineering”. Addison Wesley Longman, Inc.


[11]. Stapleton, J. 2003. “DSDM: Business Focused Development”. Second Edition. Addison Wesley. London

[12]. Highsmith, J. & Cockburn, A., 2001. “Agile Software Development: The Business of Innovation. Computer”, Vol. 34, 9 (9), September

[13]. Coad, P., LeFebvre, E. & De Luca, J. 1999. “Java Modeling In Color With UML: Enterprise Components and Process”. Prentice Hall

[14]. Boehm, B. & Turner, R. 2005. “Management Challenges to Implementing Agile

Processes in Traditional Development Organizations” IEEE Software, Vol. 22(5), September/October, pp. 30 - 39

[15]. Deming, W. E. 1990. “Out of the Crisis”, Massachussetts Institute of Technology, Center for Advanced Engineering Study

[16]. Fitzgerald, B., Russo, N. L. & O.Kane, T. 2004. “Software Development Method Tailoring” at Motorola Communications of the ACM, Vol. 46, 4 (4), April, pp. 64 - 70.

[17]. Reifer, D. J. 2003. “XP and the CMM”. IEEE Software, Vol. 20, 3 (3), May/June, pp. 14 - 15.

[18]. Vriens, C. 2003. “Certifying for CMM Level and ISO 9001 with XP and Scrum”. The proceedings of the Agile Development Conference (ADC.03). September, 2003. IEEE Computer Society. Pp. 120 - 124

[19]. Pikkarainen, M., Salo, O. & Still, J. “Deploying Agile Practices in Organizations: A Case Study”. EuroSPI 2005, European Software

Process Improvement and Innovation Conference, Budapest, Hungary, November, 2005

[20]. Keenan, F. 2004. “Agile Process Tailoring and problem analysis (APTLY)”. The proceedings of the 26th International Conference on Software Engineering (ICSE 2004). 23 - 28 May, 2004. Edinburgh, Scotland. IEEE. Pp. 45 - 47.

[21]. Highsmith, J. 2004. “Agile Project Management”. Addison-Wesley

[22]. Marjanovic 2009. “Inside Agile Processes: A Practitioner’s Perspective”, 42nd Hawaii International Conference on system Sciences, 2009 (HICSS ’09). 05 – 08 Jan 2009, Sydney, Australia, IEEE, Pp. 01 – 10.

[23]. Philip Miller. 2007. “An SEI Process Improvement Path to Software Quality”. Sixth International Conference on the Quality of Information and Communications Technology. Nov 2007. IEEE, Pp. 12 – 18

[24]. Susan D. Shaye. 2008. “Transitioning a Team to Agile Test Methods”, Agile 2008 Conference, Aug 2008, IEEE, Pp. 470 – 477.[25]. www.agilemanifesto.org/principles.html

N.GANESH is a research scholar in the faculty of Computer Science and Engineering, Anna University, Coimbatore, India. He is doing his research under the guidance of Dr. S. Thangasamy. He has obtained his B.E.[CSE] from Bangalore University, M.E.[CSE] from Anna University Chennai, M.B.A.[HR] from IGNOU New Delhi. He has to his credit a rich experience of Teaching and Training from both Educational institutions and from the Corporate World.

S.THANGASAMY is currently working as Professor and Dean, in the Department of Computer Science and Engineering, Kumaraguru College of Technology, Coimbatore, India. He holds his Engineering Degree in the field of Electrical Engineering from A.C. College of Technology, Karaikudi, passed out in 1971. He then holds the Ph.D Degree in Control and Computers from Indian Institute of Technology, Bombay, passed out in 1983. He has a rich professional experience on varied fields for more than 30 years.


WAVELET TRANSFORMS BASED ARTIFICIAL INTELLIGENT APPROACH TO THE DETECTION OF BRAIN INJURIES AND EPILEPSIES

PROF. H. N. SURESH1, SHARANABASAVESHWAR G HIREMATH DR .ARULARASU 2,

1Bangalore Institute of Technology, Dept.of Instrumentation Technology, VV puram, Bangalore–04,Karnataka, India.

2 Sharanabasaveshwar G Hiremath SCTIT Bangalore3 Govt.college of Technology, Dept.of Production Engg. coimbotore, Tamilnadu.

E-Mail ID : [email protected],

ABSTRACT : A technique proposed for the automatic detection of spikes in electroencephalograms (EEG). The important features of the raw EEG data are extracted using two methods.i.e, Wavelet Transform and Energy Estimation. This data is normalized and given as input to the Neural Network, which is trained using Back propagation algorithm. Energy Estimation is used as an amplitude threshold parameter. The Wavelet transform (WT) is a powerful tool for multi-resolution analysis of non-stationary signal as well as for signal compression, recognition and restoration, which uses Daubechies 4 as the mother wavelet. The details of the Wavelet decomposition level, 1,2,3 and energy estimation parameters are given as input to the neural network in order to detect spikes. The codes are written in C and implemented on the Texas Instruments TMS320C5410-100 processor board, and a Del Mar PWA EEG Amplifier. The waveforms are observed on MATLAB. The effectiveness of the proposed technique was confirmed with and EEG layouts. Key words: Spike, Wavelet Transform, Energy estimation, Back propagation,

1. INTRODUCTION:

The EEG is an important clinical tool for diagnosing, monitoring and managing neurological disorder related to epilepsy. This disorder is characterized by sudden recurrent and transient disturbances of mental function and / or movement of the body that results from excessive discharge of groups of brain cells. The presence of Epileptiform activity in the EEG confirms the diagnosis of epilepsy, which some times can be confused with other disorders producing similar seizures like activity. During seizures the scalp EEG of patients with epilepsy is characterized by high amplitude, synchronized periodic EEG waveforms, reflecting abnormal discharge of large group of neurons.

Between seizures, epileptiform transient waveform which include spikes and sharp waves are typically observed on the scalp EEG of such patients. Detecting and classifying sharp transient waveforms by visual screening of the EEG record is a complex and time consuming operation. Also such EEG records requires highly trained professional who are in generally short supply. Hence a requirement for automatic detection of EEG spikes and seizures. In addition the use of EEG monitoring, which produces 24 hours or longer continuous EEG recording, is becoming more common thus further increasing the need for automated detection methods. In the past many methods have been investigated to detect the EEG spikes.



Mimetic techniques have been widely used to detect spikes, but difficulties arises with artifacts. These problems increase the number of false detection’s, which commonly plague all automatic systems. In recent year an artificial intelligence approach using expert system methods have been introduced to solve these problems. Although fairly successful this approach becomes increasingly difficult due to the proliferation of the rules and the need for computers with large memories and large processing power. In addition, EEGers cannot agree on a complete set of rules acceptable to all, limiting the success of this method. If we used Fourier Transform (F.T) to detect spikes, but that gives only frequency information of the signal. The Short time Fourier Transform gives time and frequency information simultaneously, but it suffers from resolution problems. In this research work, the features of the raw EEG data are extracted using Wavelet Transform and Energy Estimation. Energy Estimation is used as an amplitude threshold parameter. This data is normalized and given as input to the Neural Network which is trained using Back propagation algorithm.

2.METHODOLOGY:

Spike detection in EEG is an important task for the diagnosis of epilepsy. The shape and size of epileptic spikes essentially change from one patient to the other. They appear in the EEG as isolated events, as well as quasi periodic oscillations of spike-and-wave. Epileptic spike detection is a very difficult task, since normal brain activity, non pathological events that resemble pathological ones, noise and instrumental artifacts can be misinterpreted as epileptic spikes. our approach to spike detection relies on the observation that the impulse-like shape of spike would result in a broad-band signal, displaying large energy at all frequencies. Indeed when analyzed with a filter bank like the one provided by the wavelet multi-resolution decomposition, a spike generates events in all the

sub-bands. On the contrary, normal brain activity and non-pathological events likely have low frequency contents and appear only in low resolution sub-bands. In the presence of broadband noise, on the other hand, the mid-range frequency sub-bands have a large spike signal-to-noise-ratio, thus allowing for an easier detection. Our scheme does not decimate the EEG sub-bands, as in non-redundant representations, avoiding the problems arising from the shift-variant property of the wavelet transform. The energy of the input signal is used as an amplitude threshold and the Wavelet Transform is used to retain the time and frequency information.2.1 Energy estimation: In signal processing techniques, the word “spike” means a) localized high frequency and b) increase in instantaneous energy. The quantitative descriptions of the amplitude and spectrum of spikes vary from signal to signal, subject to subject; it even varies from time to time for the same subject. As the spike base width increases, energy is concentrated more in the low-frequency band where the energy of the background signal is also located and detection becomes more difficult in the frequency domain. We, therefore, estimate the instantaneous energy of the output. Since spike by definition has high energy, we can implement this to detect the spikes. The energy is estimated by the formula:

E(n)= x2(n)

where E(n) is the output energy of the input, x(n) is the raw EEG input.

Fig 2.1: EEG signal and its energy

This will emphasize the spike and de-emphasize the unimportant features of EEG. The resulting signal consists mainly of high amplitude spikes. The wavelet representation is a powerful technique that has been successfully exploited in the analysis of non stationary signals, like biomedical signal processing [M.Unser & A.Aldroubi, I.Clark.& M. Echeverria.].


Unlike classical Fourier analysis, the wavelet representation allows for trading frequency resolution and time resolution. In its discrete implementation, the wavelet transform can be

viewed as a filter bank which provides a multi-resolution decomposition of the signal [M. Vetterli and J. Kovacevic]. The signal is decomposed into a series of sub-bands, each relative to a peculiar spectral region, whose bandwidth linearly increases with frequency [S.G. Mallat].

The simplest approaches that can be devised for spike detection in a multi-resolution analysis framework consists of Energy Estimation is used as an amplitude threshold parameter. [Attellis, S. I. Isaacson]. Indeed, although very fast, a single-resolution approach like that in [Sartoretto and M. Ermani] has some limitations. In [Mukhopadhyay and G. C. Ray] a nonlinear energy operator (SNEO) is proposed for the direct analysis of the EEG signal. We show that multi-resolution analysis combined with Energy Estimation give some advantages and provide a useful tool for EEG analysis[E. P. Simoncelli and E. H. Adelson.].

2.2 Subband Decompostion Principles:

In this section we briefly review the discrete-time wavelet transform and its relations with subband decomposition.

y00ýHo(z) ’! “!2 ’! ’! ‘!2 ’! Fo(z

x(n) y1

0 ý H1(z) ’! “!2 ’! ’! ‘!2 ’! F1(z)

Figure 2.2: Two- channel subband system.

Consider the two-channel filter bank Fig. 2.1. The input signal x(n) is decomposed into two sub-bands by filtering with the low-pass filter Ho(z) and the high pass filter H1(z). The output of the filters is

decimated by a factor two. It is well know that it is possible to design the analysis filter Ho(z), H1 (z) and the synthesis filter pair Fo(z),F1(z) in order to have

perfect reconstruction of x(n) at the output of the synthesis stage. One possible way to achieve perfect reconstruction is to design the analysis filter impulse response ho(n) such that its z-transform satisfies.

Ho (z)Ho (z-1) +Ho (-z) Ho (-z

-1) = 2, (1)

and choose fo (n) = ho(-n), f1(n) = h1(-n), h1 (n) = (-1)1-

n ho(1-n). Note that the above equations imply that the filter impulse response ho(n) is orthogonal to its

even-translates, namely

<ho, n, ho, n+2k > = Ó ho (n)ho (n+2k) = ä(k), n

and that < h1, n, ho.n+2k > = 0, for all k. It is easy to

see that the synthesis filters satisfy similar orthogonality conditions. If we explicitly write the synthesis stage output as a function of the sub-band signal y0

0 (n) y10(n), we have for an orthogonal perfect

reconstruction system,

x(n) = Ó yo0 (k)fo(n-2k) + Ó y1

o (k)f1(n-2k) (2)k k

Thus equation (2) can be interpreted as the series expansion of the input over the orthogonal family of function {fo(n-2k), kªz}. In an octave filter bank, or discrete time wavelet transform, the low-pass signal yo

0 (n) is further split

by low-pass filtering and sub-sampling with the analysis filter. Fig. 2 shows the equivalent scheme for a two-stage sub-band scheme, where yo

0(n) is split into yo

1 (n) and y11 (n), and Ho,o(z = Ho(z), Ho (z2),

Ho,1(z) = H 0 (z) H1(z2). The equivalent scheme is


+

obtained by applying the Noble Identities, which allow to exchange the role of decimators and filters in the iterated sub-band scheme (3).

Note that, for an analysis filter Ho(z) with

approximate bandwidth [0,fc/4], the equivalent filters Ho,o (z), Ho,1(z), and H1 (z), have bandwidth [0, Fc/8], [fc/8, Fc/4], [Fc/4, Fc/2] respectively, where Fc

is the input signal sampling frequency. Thus, the sub-bands y1

j(n)provide a multi resolution

representation of the input, each relative to a different frequency band. In particular, yo

1 (n) is a decimated smooth version of x(n), while y1

1(n) and y10 (n) are

detailed signals to be added in the synthesis stage. Note that the decimators in Fig. 2.3 give rise to a ship-variant analysis stage. This is not a desirable feature when our goal is performing time localization of events, rather that providing a compact representation of the signal. To perform spike detection, we consider the signals z1

j (n) before

decimation in Fig. 2.3, where j denotes the multi resolution level, and i ª{0,1}.

H0,0

(z) z0

1(n) “!4 y01(n)

x(n) H0,1

(z) z11(n) “!4 y

11(n)

H1(z) z

10(n) “!2 y

10(n)

Fig. 2.3 Equivalent scheme for two levels of multi-resolution analysis.

2.3 The SNEO Operator In The Frame Work Of Multi-resolution Analysis

The smoothed Nonlinear Energy Operator (SNEO) has been proposed in [7] for the analysis of EEG signals. SNEO is a smoothed version of the nonlinear energy operator.

ψ[x(n)] = x2 (n) – x (n+1) x(n-1) (3)

smoothing is achieved by low-pass filtering ψ[x(n)], in order to obtain an estimate SNEO [x(n)] of the

expectation E[ψ(n)]. Indeed, taking the expectation of (3), for a stationary zero mean process x(n) we obtain.

E[y[x(n)]] = rx (0) – rx (2) = ò2p Rx(ejw) (1 – cos2w)

dw/2p (4) 0Where rx(k) = E [x(n)x(n+k)] is the input process autocorrelation function and Rx(e

jw) is the

spectral density of x(n). From equation (4) one can see that SNEO [x(n)] is an approximation of the power of a band pass filtered version of the input process. For non-stationary process, a similar interpretation can be given in terms of the evolutionary spectrum [Miroslaw latka & Ziemowit]. More-over, if the smoothing low-pass filter has a short compact support, the information provided by SNEO [x(n)] is relative to the local characteristics of x(n) around time n.Beside its good properties for spike detection, the SNEO operator has dome disadvantages, pointed out in the sequel, with respect to interference immunity, which our multi-resolution approach should overcome. Assume first that a constant value K is added to the EEG signal x(n), during a given time interval. Such a phenomenon, is produced, as an example, by patient movements, which produce an offset in the EEG measurement. We have ψ[x(n) + K] = ψ[x(n)] + K (2x(n) – x(n – 1) – x (n + 1)).Although low pass filtering attenuates the interference term, it is apparent that SNEO [x(n)] depends on the local DC value of the signal, and this is not a desirable effect in spike detection.Our scheme exploits the SNEO operator in the framework of multi-resolution analysis[ H.N.Suresh,Dr.V.Udaya shankara]. The signal is analyzed using a three level discrete-time wavelet decomposition. The 5-tap almost orthogonal linear phase filters of [8] are used in the experiments. The detail signals z1

0(n), z11(n)

and z12(n) are then processed using the SNEO

operator. Note that, when the EEG signal is sampled by an Fs Hz frequency, the three details signals pertain to the Frequency bands [Fs/4, Fs/2] Hz, [Fs/8, Fs/4] Hz and [Fs/16, Fs/8] Hz, respectively. An

impulse-like signal, as a spike, generates a significant output in all the three sub-bands. On the other hand,


sinusoidal, band pass and low pass interference is present in some or none of the sub-bands.

Our idea is to devise a spike detector based upon the values SNEO[zI

j(n)], j = 0, 1, 2, i = 1. Given a

specific threshold on each of the three levels, we say that a spike is detected at time n what at that time SNEO [zI

j(n)], is above the level threshold, for all j =

0, 1, 2, i = 1. A specific threshold value is used in each subband, to take into account the peculiar subband amplitudes corresponding to a spike.

3. DATA SELECTION :

The EEG data required for the detection of spikes was obtained from the National Institute of Mental Health and Neurosciences (NIHMANS), Bangalore. The data acquired was from both normal and epileptic patients. The data was recorded using a “10-20” system with bipolar montages.

4.EXPERIMENTAL RESULTS:

4.1 Energy Estimation

The spikes are always associated with high energy. We can obtain the instantaneous energy of spikes by using squaring the input EEG signal. Fig 4.1& 4.2 consists of 256 and 512 data samples. On executing the C code an ENERGY.TXT data file is generated which has the normalized energy values. The values from this file are plotted using MATLAB. The raw EEG data is squared and the energy output is obtained.

Fig 4.1: Energy output Data file1

Fig 4.2: Energy Output Data file24.2 Wavelet transform

A feature extraction scheme using the wavelet transform (WT) has been applied. Through wavelet decomposition of the EEG records, transient features are accurately captured and localized in both time and frequency context. The mother wavelet is chosen from the Daubechies family for its easy implementation in this application. Daub4 characterized by 4 vanishing moments is implemented here. Here both scale2 and scale3 coefficients are given as input to Artificial Neural Networks for training and testing and the results obtained from these 2 are compared. The differential of this output is given as the input to the Artificial Neural Network for training. Fig 4.3 shows the decomposition principle of Wavelet transform.


Fig 4.3 Wavelet transform decomposition

The Data files have shown Fig.4.4, 4.5, & 4..6 have 512 samples. The raw EEG data is decomposed into 3 levels. This is differentiated and this result is stored in the output file. On executing the C code, 2 data files are generated, COEFFS.

TXT which has the details of the specific level and DIFF.TXT which has the differential of the details. The values from these files are plotted using MATLAB. Fig.4.4, 4.5, & 4..6 represent the detailed & differentiated results of Level 1, Level 2, & Level 3 of wavelet transform.

Normalization:

The operation on a digital computer system limits us on the size of the input number. So, we must restrict the values of input between 0 and 1. Since this does not affect the resolution of the input data, the input values are normalized between the values 0 and 1. This makes the learning process simpler

Fig 4.4: Level 1 Details (d1) and d/dt of d1 for Data file1



5. IMPLEMENTATION USING BACK PROPAGATION ALGORITHM:

The functional diagram gives a description of the method adapted for spike detection. Here, the features of the raw EEG data are extracted using Wavelet Transform and Energy Estimation technique.

Fig. 5.1 functional diagram


This data is normalized and given as input to the Neural Network that is trained using Back propagation algorithm. We, hence, obtain the spike output. The steps for this procedure and the output waveforms are included in this chapter. Fig.5.1 shows the functional diagram of the neural network. The Input Data files shown fig.5.1 have 512 samples. The 2 extracted features are given as inputs to the Back propagation Algorithm and the output is obtained. On executing the C code, 2 files are generated, NETOUT.TXT which contains the output of the Neural Network and PULSE.TXT which contains 1’s at the points of spike occurrence. The values from this file are plotted using in MATLAB.

Fig 5. 2: Network outputs for data file 1

Fig 5.3: Pulse outputs for Data file1

6. SPIKE OUTPUT

The pulse train output obtained from the Backpropagation Algorithm is logically ANDed with the original signal i.e., raw EEG data to retain only spikes as the output.

The Input Data files have 512 samples. The pulse train is given as input which is logically ANDed and the spike output is obtained. On executing the C code, an output file SPIKE.TXT is generated which contains only spikes. The values from this file are plotted using MATLAB.

Fig 6.1: Spike outputs for Data file 1

Feature extraction is necessary to enhance the spike information and suppress the unwanted background activity. We have chosen two features, energy of the EEG signal and the wavelet decomposed signal of the raw EEG data. These are given as inputs to the neural network.The neural network uses Backpropagation algorithm for training. The efficiency of the Backpropagation algorithm depends on the choice of learning rate and the momentum. Variations in the input thus have little effect on the output. Also, if the values are very small, the learning process is slow. After executing the code several times we have chosen the value of learning rate and momentum to


be 0.35 and 0.15 respectively. Then the network was trained to give an output pulse train in the presence of spike activity. The pulse train was used to window the original signal to retain only the spikes.In our research, we have compared the outputs obtained from three different levels of wavelet decomposition along with the energy of the raw EEG signal was given as one of the inputs to the neural network. It was observed that the implementation with level 3 produced better results compared to d1 and d2.

Fig 6.2: Spike outputs for Data file2

CONCLUSION: The use of Wavelet Transformation and Neural Network base approaches have proven its usefulness in achieving safety related to head injuries and Epilepsy. Results found that, the spikes are successfully detected using Neural network based on wavelet transform & energy estimation as a preprocessor. The details of the Wavelet decomposition level, 1,2,3 and energy estimation parameters are given as input to the neural network in order to detect spikes. The codes are written in C and successfully implemented on the Texas Instruments TMS320C5410-100 processor board. The waveforms are observed on MATLAB. The effectiveness of the proposed technique was confirmed with and EEG layouts. REFERENCES: [1] M.Unser and A. Aldroubi, “A review wavelets in biomedical applications,” Proceedings of the IEEE, vol.84.pp. 626-638. April 1996.

[2] I. Clark. R. Biscay. M. Echeverria. And T. Virues, “Multiresolution decomposition of non-stationary EEG signals: a preliminary study”, Comput. Biol. Med., vol.25, no.4 pp. 373-382, 1995.

[3] M. Vetterli and J. Kovacevic, Wavelets and Subband Coding. Englewood Cliffs NJ: Prentice-Hall, 1995.

[4] S.G. Mallat, “A theory for multiresolution signal decomposition: The wavelet representation,” IEEE Trans. Pattern Anal. Mach. Intel., vol.11, pp. 674-693, July 1989.

[5] C. E. D’ Attellis, S. I. Isaacson, and R. O. Sirne, “Detection of epileptic events in electroencephalograms using wavelet analysis”, Annals of Biomed. Engng, Vol 25, pp. 286-293, 1997.

[6] F. Sartoretto and M. Ermani, “Automatic detection of epileptiform activity by single-level wavelet analysis”, Clinical Neurophysiology, vol.110, pp. 239-249, 1999.

[7] S. Mukhopadhyay and G. C. Ray. “A new interpretation of nonlinear energy operator and its efficacy in spike detection”, IEEE Trans. On Biomed. Engng, vol.45, no.2, 1998.

[8] E. P. Simoncelli and E. H. Adelson. “Subband transforms” in subband Image Coding (J. W. Woods. ed.) pp. 143-192, Kluwer Academic Publisher. 1991.

[9] Miroslaw latka & Ziemowit “Wavelet analysis of epileptic Spikes”. Wroclaw university of Technology, Poland, Dec. 22, 2002.


[10]. H.N.Suresh,Dr.V.Udaya shankara,”A new Technique for Multi resolution Characterization of epileptic spikes in EEG”. International Journal of Information Technology Volume 2 Number 1 2005 ISSN 1305-2403,.pp.42-45.

PITCH DETECTION ALGORITHM USING HARMONIC PATTERN MATCHING IN FOURIER OF FOURIER TRANSFORM DOMAIN

AKANT K. [1], PANDE R. [2], LIMAYE S.S [3]

[1] Department of Electronics & Communication Engineering, Manoharbhai Patel Institute Of Engineering & Technology, Gondia, Maharashtra, INDIA

[2] Department of Electronics Engineering, Shri Ramdeobaba Kamla Nehru Engineering College, Nagpur Maharashtra, INDIA

[3] Department of Electronics Engineering, Jhulelal Institute of Technology, Nagpur, Maharashtra, INDIA

[email protected][1], [email protected][2],[email protected][3]

ABSTRACT:

An algorithm is proposed to detect the pitch of a singing voice for Query By Humming applications. Investigation of harmonic pattern of the sound in frequency domain gives us fundamental frequency. The periodicity of the Fourier transform is detected by again taking its Fourier transform to obtain the “Fourier of Fourier transform” (FFT2). In this method, harmonic pattern matching is done in Fourier of Fourier Transform domain to obtain pitch of the sound. Cross correlation function of spectra of singing voice with the ideal harmonic pattern in FFT2 domain, which contains ones at the positions of harmonic frequency components is calculated. Fundamental frequency of that sound is determined by knowing ideal harmonic pattern for which maximum correlation is obtained. The resolution of frequency measurement is improved much beyond the FFT bin size by employing Parabolic Interpolation. Advantages of FFT2 over Fourier Transform (FT) in context to estimation of fundamental frequency are discussed.

Key words: Pitch, Query By Humming, Fourier of Fourier transform, harmonic pattern matching, Parabolic Interpolation.

1.INTRODUCTIONIf the input signal is a musical note, then its spectrum should consist of a series of peaks, corresponding to fundamental frequency with harmonic components at integer multiples of the fundamental frequency. The frequency that human ear interprets as the pitch is this fundamental frequency. Hence pitch estimation is the problem of determining the fundamental

frequency present in the signal. It is inherently related to the detection and estimation of sinusoids. Many methods to extract pitch of the music signal have been proposed. Some use time domain methods, some use frequency-domain methods, and others use a combination of both (Zhao and Brown, 2003). In time domain methods fundamental frequency is obtained by measuring periodicities in time domain whereas in frequency domain methods this is


achieved by harmonic pattern matching in frequency domain (Bapat and Rao, 2005).Frequency domain algorithms are more robust than time domain algorithms (Cuadra, Master and Sapp, 2001). In harmonic matching pitch detection algorithms (PDAs) measured spectrum is matched with an ideal harmonic spectrum. In many pitch extractors, magnitude of the short-time Fourier transform (STFT) is evaluated for the initial processing. Few examples of harmonic pattern matching PDAs are: pattern recognition (PR) PDA (Brown1992), the two-

way mismatch (TWM) PDA (Maher 1994), Harmonic Product Spectrum(HPS) PDA (Cuadra, Master and Sapp), and Maximum Likelihood(ML) PDA (Cuadra, Master and Sapp). In TWM PDA, the pitch of a signal is estimated by choosing the fundamental frequency which minimizes the discrepancy between the measured spectrum and the spectrum predicted for that fundamental frequency. The PR PDA exploits the fact that for a logarithmic frequency scale, corresponding to musical intervals, a harmonic structure always takes on the same pattern regardless of the value of its F0.

Consequently, a pattern recognition algorithm is applied to detect such patterns in the measured spectrum by correlation with ideal spectral patterns for different trial F0 values in the expected pitch range. In HPS PDA fundamental frequency is obtained by compressing the spectrum of music signal a number of times (downsampling), and comparing it with the original spectrum. We can see that the strongest harmonic peaks line up. The first peak in the original spectrum coincides with the second peak in the spectrum compressed by a factor of two, which coincides with the third peak in the spectrum compressed by a factor of three. Hence, when the various spectrums are multiplied together, the result will form clear peak at the fundamental frequency. The ML algorithm searches though a set of possible ideal spectra and chooses the one which best matches the shape of the input spectrum. Present work addresses the problem of reliable pitch tracking of monophonic sound to be applied for query by humming and microtone research. Pitch tracking is achieved through harmonic pattern matching in FFT2 domain. Multimedia applications are rapidly developed in last two decades. Due to which people are enjoying many entertainment activities such as music songs in their life. This led to the development of many music information retrieval (MIR) systems. Broadly there are two types of MIR systems. One based on text retrieval models by entering music names, genre, artists etc., while the other based on content based retrieval models by humming the melody or singing the lyrics (Fu, Xue, 2005). Songs are usually recalled by its melody rather than its title or composer. Hence, MIR based on humming a melody is much more convenient for music retrieval (Liu, Wu, Li, 2003 and Pollastri, 2002). Query by humming (QBH) is a music retrieval system that takes an input query as user hummed, sung or whistled query and compares it to an existing database. The system then returns a ranked list of music closest to the input query. (Birmingham,

Dannenberg, and Pardo, 2006). Monophonic pitch tracking is required for QBH applications. Important acoustic features of the query are those related to the tune or melodic pitch contour. Note labelling process mainly depends on note onset detection. It deals with music content of audio signal. For each note event, an estimation of pitch measured in Hertz must be converted into music note labels; this process should be carried out with respect to a musical scale. It is critical because intonation is never absolute and singers always introduce errors during a real performance. Also, note segmentation is the most difficult problem and QBH applications perform well when accurate segmentation is done. Many QBH systems have restricted the user query syllable to unvoiced phone like /ta/ to facilitate accurate note segmentation (Kumar, Joshi, Hariharan, Dutta-Roy, Rao, 2007). Most research into music information retrieval thus far has only examined music from Western tradition using equi-tempered scale. In this scale, an octave is divided into 12 notes whose frequencies are in geometrical progression. In other words, these notes are logarithmically equidistant in frequency. This division has been accepted as a compromise for ease of manufacturing musical instruments. However, purists argue that for ensuring perfect harmony, the note frequencies should be related by simple ratios rather than be in a geometric progression. This gives rise to the “Just tuned” scale. The notes in this scale slightly deviate from the equi-tempered scale notes. These minute variations are called “microtones”. Very little work has taken place in the area of applying techniques from computational musicology and artificial intelligence to the realm of Indian Classical Music (ICM) (Pandey, 2003). “Raag is a melodic abstraction around which almost all ICM is organized. A raag is most easily explained as a collection of melodic gestures and a technique for developing them. The gestures are sequences of notes that are often inflected with various micro-pitch alterations and


articulated with an expressive sense of timing. Longer phrases are built by joining these melodic atoms together (Chordia , 2007)”. In other words, Raag is a typical musical pattern observed in ICM. Raag identification from the audio sample is nothing but musical pattern recognition, which would be another application of our method. (Sinith , Rajeev, 2007) propose a scheme for the recognition of such predefined musical patterns in a monophonic environment in the context of South Indian Classical Music. (Sridhar, Geetha 2009) used Raag

identification as a basis for music information retrieval. Music of other origins often conforms to different tuning systems. Therefore there are problems both in representing this music as well as finding matches to queries from these diverse tuning systems (Suyoto, Uitdenbogerd, 2004). In the context to ICM microtones are termed as Shruti. There are 22 Shrutis in an octave (Dr. Oke, 2008). This is debatable point but many musicologists do agree that there are 22 intervals per octave in ICM, while others disagree with such claims (Krishnaswamy, 2003).

In order to find the positions of these microtones, we need to analyze the sound samples from the well known maestros. For a QBH application we should be able to estimate the pitch to the nearest semitone. The semitones are separated by a factor 2(1/12) i.e. roughly 6%. For microtone research, the expected accuracy is 1%. Performance of proposed method in terms of accuracy is very much beyond expectation. QBH is also real time and there is need for computationally efficient algorithm. In view of these applications accuracy and computational efficiency are the mainly required features, which are fulfilled by the proposed method. We have employed combination of pattern recognition (Brown, 1992) method and Fourier of Fourier transform method (Marchand, 2001) for the estimation of pitch in view of QBH. The method suggested by Bay and Beauchamp, (2006) can be thought of in FFT2 domain for pitch tracking in polyphonic environment. This paper is organized as follows. Section 2 presents the details of Fourier of Fourier transform. In section 3 how frequency resolution is improved is discussed. Section 4 presents the proposed pitch detection algorithm. Section 5 presents experimental results and in section 6 the conclusion is given.

2. FOURIER OF FOURIER TRANSFORM

In our analysis we have used two Fourier transforms in sequence referred as Fourier of Fourier Transform (FFT2). Our method works very well in the case harmonic sounds, i.e. sounds rich in harmonics. It is not suited for pure sinusoids. Fourier transform, FT (first Fourier transform of the signal) of a typical musical sound has a series of peaks in its magnitude spectrum corresponding to the harmonics of the sound, at frequencies close to multiples of the fundamental frequency F. The peak showing fundamental frequency may not always be dominant.

Hence single Fourier transform is inefficient to identify correct peak. Fourier of Fourier Transform is of great interest in locating this peak, which helps to overcome the possibility of octave error. To find out Fourier of Fourier Transform, we compute magnitude spectrum of the Fourier transform of singing voice. Magnitude spectrum of the Fourier transform of the above magnitude spectrum is then computed. Note that this transform is not the same as the well-known “Cepstrum”, which is the (inverse) Fourier transform of the logarithm of the spectrum resulting from the Fourier transform. Figure 1 shows the FT of piano C# of 5th octave. This FT has a series of uniformly spaced peaks as shown in Figure 1, corresponding to the harmonics of fundamental frequency.

Figure 1: Fourier Transform of piano C# of 5th

octave

We can clearly see that, peak corresponding to fundamental frequency is not dominant. If fundamental frequency is F, the distance between two consecutive peaks corresponds to a period of “1 bins where:

FsFN11 =∆

(1)


N1 : size of the first Fourier transform

Fs : sampling frequency

The first peak is at bin 0 and it corresponds to the DC level. The difference between second peak (shown by an arrow in Figure 1) and the first peak is “1 bins. Figure 2 shows the spectrum of Fourier of Fourier Transform of piano C# of 5th octave. In this spectrum of FFT2, there are series of peaks. Here also, the first peak is at bin 0 and it corresponds to the DC level. The second peak is shown by an arrow in Figure 2. The distance between two consecutive peaks corresponds to a period of “2 bins where:

1

22 ∆

=∆ N

(2)

N2: size of the second Fourier transform

From Equation (1) and Equation (2), we get

=∆

FsFN

N)( 1

22

(3)

Figure 2: Fourier of Fourier Transform of piano C# of 5th octave

If size of first and second Fourier transform is same (N2 = N1), Fundamental frequency F is given by,

2∆= FsF

(4)

3. IMPROVING FREQUENCY RESOLUTION

The maximum frequency resolution is limited to bin size. In our study, the sampling frequency is 44100 Hz, window size is 2048 and size of first as well as second Fourier transform is 4096. This yields a bin size of Fs/N, which is equal to 10.76 Hz. This resolution is inadequate for micro pitch estimation, which requires accuracy of at the most half a semitone or preferably less than that. For QBH applications, accuracy of at least a semitone is required; hence resolution of 10.76 Hz. is not sufficient for QBH applications as well. Frequency resolution can be improved by increasing window size but we will lose time resolution and computational effort will also increase. In our method frequency resolution is reduced to 1 Hz without changing bin size, using parabolic interpolation strategy given by Smith and Serra (1987). This ensures accurate estimation of pitch without increasing computational complexity.

4. PROPOSED PITCH DETECTION ALGORITHM

Step 1. To determine pitch at a certain time t, temporal frame centered at t is considered. The first step is to multiply temporal frame with Hann analysis window. For good harmonic resolution, at least four periods of the harmonic signal under the window are required. Window size was 2048. This corresponds to 46.4 ms duration for sampling frequency of 44100 Hz.

Intentional variations of pitch may occur as fast as 480 times per minute. i.e. 8 times per second. Thus minimum note duration is 1/8 seconds i.e. 125 ms. Thus 46.4 ms overlapping frames have sufficient time resolution. Frame overlapping is done by selecting hop size such that there is 75% overlap of the frames.

Step 2. Fourier of Fourier Transform (FFT2) of this windowed signal is computed as discussed in section 3. Size of first and second FFT was 4096. This is chosen double the frame size to increase frequency resolution.

Step 3. Ideal FFT2 patterns for four octaves in audio range, with 48 frequencies in an octave are generated. 48 points per octave gives expected resolution of quarter of a semitone. Ideal FFT2 pattern is considered as series of equidistant peaks, whose amplitudes are taken as unity as shown in Figure 3. The numbers of peaks are chosen to be equal to 5. These peaks are separated by “ bins, where “=Sampling Frequency/Trial Frequency. Sampling frequency is 44100 Hz. Trial frequency values are


considered with four octaves in expected frequency range at a resolution of 48 points per octave.

Figure 3: Ideal FFT2 patternStep 4.Correlation of all these patterns with FFT2 in step 2 is performed. Maximum will occur at frequency close to the fundamental frequency.

Step 5.” in bins for the frequency in step 4 is calculated as “=Fs/Frequency

Step 6.Fine tuning is done by finding the exact location of peak in FFT2 of signal in the vicinity of “ in step 5 and applying parabolic interpolation to this peak.

Step 7.Once correct “ is found fundamental frequency will be Fs divided by “.

Step 8.Now, frame is advanced by one hop and whole procedure is repeated to find out fundamental frequency for new frame.

Step 9.This is continued till end sample of the signal to get time vs. pitch graph.

Figure 4: FFT2 spectrum of note Ab in 2nd

octave showing peaks corresponding to fundamental and its double.

In Figure 4, peak at bin 0, peak 2, 4, 6, 8 are corresponding to fundamental, whereas peak at bin 0, peak 1, 2, 3, 4 are corresponding to double the fundamental. We get correlation peaks at fundamental and its double.

We found that, the peak at double the fundamental is sometimeslarger. This would lead to octave errors in pitch estimation. This is avoided by finding out the peak corresponding to the fundamental (peak 2) and its double (peak 1) in the FFT2 spectrum. Then test range in step 3 is varied frame to frame such that the highest frequency is considered to be this fundamental. In original PR PDA (Brown1992), ideal spectrum was considered as logarithmically spaced impulse train, which is nothing but a Fourier Transform in log frequency domain. Impulse train was having 20 harmonics with equal amplitude. In our case, we are doing correlation in FFT2 domain, where peaks are linearly spaced. We consider only 5 peaks instead of 20. This reduces computational load to great extent. Another advantage of FFT2 is that the peaks in FFT2 are more widely spaced as illustrated in the Table 1. Here 12 notes in the 4th octave are analyzed with sampling frequency of 44100 Hz and FFT size as 4096 and the bin index numbers in FT and FFT2 algorithms are tabulated. (Note that due to slight mistuning of the Piano, A is having a frequency of 442.8272Hz rather than 440 Hz.)


Table1. : Indices of harmonics in terms of bins in FT and FFT2 for notes in 4th octave.

We observe from above table that, in FT there is only one or two bins difference

for a semitone, while in FFT2 there is

Table2. : Differences between J. C. Brown PR PDA and our proposed method

five to nine bins difference. Also, as we move to the lower octaves, index in FT goes on reducing while index in FFT2 goes on increasing. For some of the two consecutive semitones in third octave, the indices in FT are same but in actual their frequencies are different. Estimation of fundamental frequency without parabolic interpolation would give same value for these semitones. Hence it is parabolic interpolation which plays important role in finding correct frequency of such semitones. The singers in Indian Classical Music mostly cover frequency range from 100 Hz to 600 Hz (A2 to D5) while singing. Hence for lower octaves, estimation of fundamental frequency becomes more critical as the peaks are closely spaced in FT domain. This problem is overcome in our case as, we have used FFT2 instead of FT (in original method), where harmonics are placed at quite larger locations. Common problem of octave error is obviated in our method by finding out correct peak through FFT2. Another feature of FFT2 is its ability to detect peaks of harmonics corresponding to multiple pitches.

5. RESULTS

5.1Ability of FFT2 to detect multiple pitches

Figure 5: FFT2 spectrum when A flat and C sharp of 4th octave played together.

In FFT2 domain the spectral peaks are not as closely placed as in FT domain, so it becomes easier for peak detector to locate the peaks without any ambiguity. Figure 5 shows the FFT2 spectrum when A flat and C sharp of 4th octave played together. This ability of FFT2 is of great interest in music segregation in polyphonic environment.


Figure 6: Fundamental frequency tracking results for piano from F3 to E4

Figure 7: Fundamental frequency tracking results for piano from F4 to F5

5.2 Accuracy

The accuracy of the algorithm was tested by finding out % Error for mathematically generated harmonic signals ranging from 90 to 1000 Hz in step of 1 Hz. Harmonic signal was considered as fundamental (with amplitude 0.75) and its first harmonic (with amplitude 0.25) with sampling frequency 44100 Hz. Mathematical signal is generated by the Equation (5) given below.

y = 0.75 sin(2 pi Fref t)+0.25 sin(2 pi (2Fref) t) (5)

Figure 8: % Error vs. Fundamental frequency for the proposed method.

Size of first and second Fourier transforms (N1 and N2) was 4096. Number of samples per analysis frame (window size) was 2048.% Error is calculated as follows.

100% ×

−=

FrefFrefFcal

Error (6)

Where, Fref be the exact fundamental frequency which is put in mathematical equation for generating a signal and Fcal its measured value. Our method is found to be very accurate in view of microtone research and QBH applications. Method proposed by Marchand (2001) shows 1.5% to 6% Error for 150 to 1000 Hz, where as our method shows 0.5% to 0.01% Error for the same range, which is very much appreciable in context to concerned applications. % Error at lower frequencies is more as numbers of cycles accommodated in analysis window are less. Hence it is obvious that the error goes on reducing for higher frequencies.

5.3 Robustness

Frequency resolution in Autocorrelation method is very good but it sometimes gives octave errors such as higher harmonic or sub-harmonic detection. Results were compared with autocorrelation methods Praat (Boersma, Weenink,) for various vocal samples.One of them is song Dnyaniyancha Raja Guru Maharav for 10 sec. to 13 sec. available at www.dhingana.com. It was analysed to get pitch graph by Praat method. It gives octave error as shown in Figure 9. For the same audio sample, pitch graph was obtained using our method with no octave error (Figure 10). Proposed algorithm obviates the octave error as the crux of this algorithm is the ability of detecting correct peak corresponding to the fundamental frequency.


Figure 9: Pitch graph for song Dnyaniyancha Raja Guru Maharav for 10 sec. to 13 sec. available at www.dhingana.com by Praat method showing octave error.

Figure 10:Pitch graph for the same sample in Figure 9 above by proposed method showing no octave error.

5.4 Performance

This method requires computation of two Fourier transforms and harmonic pattern matching by correlation. It is much faster than well known autocorrelation method. Also FFT2 is similar in principle to Cepstrum ( Noll, 1967) but computationally more efficient as logarithm is not required. Pitch graph for Happy Birthday To You song in humming voice for all four lines is shown in Figure 11, Figure 12, Figure 13, Figure 14. This algorithm is tested for monophonic male vocal, monophonic female vocal as well as monophonic instrumental audio samples and proved to be accurate and octave error free.

Figure 11:Pitch graph : Happy Birth Day, Line 1, Singer 1 (Female)

Figure 12: Pitch graph : Happy Birth Day, Line 2, Singer 1 (Female)


Figure 13: Pitch graph : Happy Birth Day, Line 3, Singer 1 (Female)

Figure 23: Pitch graph : Twinkle Twinkle, Line 1, Singer 4 (Male)

6.CONCLUSION

There are two major difficulties, namely, octave errors and pitch estimation accuracy, which most pitch detection algorithms have to deal with (Dziubinski, Kostek, 2004). In this article we have proposed a method of pitch tracking which gives accurate estimation of pitch and results with no more octave errors. The computational complexity in the algorithm is not very large as it is in the other pitch detection methods. For applications such as QBH, microtone research, musical pattern recognition and musical genre classification, the first step is to identify musical note through pitch tracking. High accuracy in pitch estimation is the main requirement for such cases. This method is suitable in the realm of Indian Classical Music, where, there is concept of Shruti (microtones). Our algorithm is limited to frame level analysis. Its input is a frame and outputs are frequency and average amplitude. In future, the amplitude value of the signal will be used by the higher layer to detect silence zones, note onsets and note duration. The higher layer will convert the frequency to a MIDI note number and quantized note duration. Still higher

layer will operate on the note sequence and do higher level tasks like HMM (Rabiner, 1989) genre identification, database search etc. Obtaining pitch contour of the melody from polyphonic audio recordings is important from musicological perspective. It has several applications such as automatic transcription of music, indexing of music for music retrieval system, music classification, sound segregation in music signals, singer identification etc. Human ear perception is subjective. It varies from person to person. Hence it is difficult job for a novice singer to identify whether he is singing in correct pitch. It is where our algorithm can help them immensely by giving correct pitch recognition and thereby helping them improve their singing ability. Getting pitch contour in polyphonic environment would be next step in further developing this algorithm, as reliable pitch estimation in polyphonic music is still a challenging problem. We feel that there is lot of scope for using this method to estimate pitch of sound in percussive background and this might open avenues for future research.

REFERENCES


[1] Bapat, A.; Rao, P. 2005. Pitch tracking of voice in tabla background by the two-way mismatch method, Proc. of the 13th Int. Conf. on Advanced Computing and Communications, Coimbatore, India.

[2] Bay, M. and Beauchamp, J. W. 2006. Harmonic Source Separation Using Prestored Spectra, Independent Component Analysis and Blind Signal Separation , Springer Berlin / Heidelberg, Book Volume 3889.

[3] Birmingham, W.; Dannenberg, R. and Pardo, B. August 2006. Query by humming with the Vocal Search system, Communications of the ACM, Vol. 49, pp. 49-52

[4] Boersma, P.; Weenink, D. “Praat: doing phonetics by computer”: Institute of Phonetic Sciences, University of Amsterdam (www.praat.org)

[5] Brown, J. C. 1992. Musical fundamental frequency tracking using a pattern recognition method, Journal of the Acoustical Society of America, vol.92, no. 3

[6] Chordia, P.; Rae, A. 2007. Raag recognition using pitch-class and pitch-class dyad distributions, In Proceedings of International Conference on Music Information Retrieval

[7] Chordia, P.; Rae, A. 2007. Understanding Emotion In Raag : An Empirical Study of Listener Responses, , In Proceedings of the International Computer Music Conference, ICMC 2007

[8] Cuadra, P. de la; Master, A.; Sapp, C. 2001. Efficient Pitch Detection Techniques for Interactive Music, in the Proceedings of the International Computer Music Conference (ICMC) 2001, Havana, Cuba. pp. 403-406[9] Dr. Oke, V. 2008, http://www.22shrutiharmonium.com

[10] Dziubinski, M.; Kostek, B. 2004. High accuracy and octave error immune pitch detection algorithms, ARCHIVES OF ACOUSTICS, VOL 29, pp. 3-24

[11] Fu, L.; Xue, X. April 2005. A New Spectral Based Approach to Query by Humming for MP3 Songs Database, World Academy of Science, Engineering and Technology, Vol. 4, pp. 117-121

[12] Krishnaswamy, A. April 2003. Application of pitch tracking to South Indian classical music, Proceedings of IEEE, ICASSP, Hong Kong, China, pp. 49-52

[13] Kumar, P.; Joshi, M.; Hariharan; Dutta-Roy, S.; Rao, P. January, 2007.Sung Note Segmentation for a Query-by-Humming System, Proc. of Music-AI (International Workshop on Artificial Intelligence and Music) in IJCAI, 2007

[14] Liu, B.; Wu, Y.; Li, Y. 2003. Linear hidden Markov model for music information retrieval based on humming, Proceedings of ICASSP’03, IEEE International Conference on Acoustics, Speech, and Signal Processing, Vol.5, pp. 533-536

[15] Maher, R.C. and Beauchamp, J.W. 1994. Fundamental frequency estimation of musical signals using a Two-Way Mismatch procedure, J. Acoust. Soc. Am., vol. 95., no. 4, pp. 2254-2263

[16] Marchand, S. 2001. An efficient pitch-tracking algorithm using a combination of Fourier transforms, Proceedings of the COST G-6 Conference on Digital Audio Effects (DAFX-01), Limerick, Ireland, December 6-8

[17] Noll, M. February 1967. Cepstrum Pitch Determination, Journal of Acoustical Society of America, Volume 41, Issue 2, pp. 293-309

[18] Pandey, G. 2003. Tansen : A system for automatic raga identification, Indian International conference on Artificial Intelligence 2003, Hyderabad, India .

[19] Pollastri, E. 2002. A pitch tracking system dedicated to process singing voice for music retrieval, IEEE International Conference on Multimedia and Expo, 2002. ICME ’02. Vol.1, pp. 341-344.


[20] Rabiner, L. R. February 1989. A Tutorial on Hidden Markov Models and Selected Applications in Speech Recognition, Proc. IEEE, Vol. 77, No. 2, pp. 257-286

[21] Sinith, M.S.; Rajeev, K. Dec. 2007. Pattern Recognition in South Indian Classical Music Using a Hybrid of HMM and DTW, IEEE Conference on Computational Intelligence and Multimedia Applications, 2007, Vol. 2,pp. 339-343

[22] Smith, J. O. and Serra, X. 1987. PARSHL: An Analysis/Synthesis Program for Non-Harmonic Sounds Based on a Sinusoidal Representation, Proceedings of the 1987 International Computer Music Conference, International Computer Music Association, San Francisco, pp. 290 - 297

[23] Sridhar, R. and Geetha, T.V. May 2009. Raga Identification of Carnatic music for Music Information Retrieval, International Journal of Recent Trends in Engineering, Vol. 1

[24] Suyoto, I. S. H.; Uitdenbogerd, A. L. 2004. Exploring Microtonal Matching, ISMIR 2004, 5th International Conference on Music Information Retrieval, Barcelona, Spain, October 10-14, 2004, Proceedings

[25] Zhao, Z.; Brown, L. J. December 2003. Musical pitch tracking using internal model control based frequency cancellation, 42nd IEEE Conference on Decision and Control, Vol.5, pp. 5544- 5548.

DEVELOPMENT OF MATHEMATICAL MODEL FOR GROUP REPLACEMENT OF ELECTRONIC SYSTEM USING MARKOV CHAINS

1NAVEEN KILARI, 2DR.C. NADHAMUNI REDDY, 3DR.B.BALU NAIK

1Member Technical Staff E-III, Centre for Development of Advanced Computing (C-DAC),

Hyderabad, India

2Professor & Principal, S.V.P. College of Engineering and Technology, Andhra Pradesh,India

3Associate Professor , Department of Mechanical Engineering , JNTU College of Engineering, Hyderabad, Andhra Pradesh, India


ABSTRACT:

This paper, attempts to develop Markov process based mathematical model for group replacement of a block of Computer and Computer Based System. To make the model more realistic two intermediate states i.e. minor repair state and major repair state are introduced between working and failure states of the system. Markov process (chain), a stochastic process, is employed to compute the probabilities of transition from a given state to any future state.

Key words:

Markov Process, Group replacement, Spectral decomposition.

INTRODUCTIONWith the advent of Computers and the widespread of the internet and fibre optics network across the world, the huge population comprising good number of unemployed Indian youth provided a platform for MNCs to open up their ICT offices in India. Consequently the decisions on capital investment on Computer and Computer based system in these MNCs became important and a need for scientific approach for the replacement decisions is felt.

THE DECISION PROBLEMThe replacement decisions in MNCs in ICT area are predominantly with the computers and computer based system. The primary decision is generally whether to replace the existing computer based system consisting a large number of computers or use for some more period of time.

The replacement of the items arises in the following situations:

- The existing machinery or the system is less-efficient and demanding too much maintenance

- Breakdown of the existing machinery or the system due to accident or otherwise

- The existing machinery or the system wears out and is likely to fail shortly

- A new system with better design and capability is available in the marketTo act upon the above situations, the problem is to balance the cost of the new system against the maintenance cost to be incurred to keep the existing system in efficient condition or the cost due to the loss of efficiency. It is essential to determine the optimum duration over which the present system can be used. And it is also required to decide the optimal replacement policy that covers the time at which the replacement is most economical. This policy shall aim at minimizing the annual capital and maintenance cost. The replacement decisions are relatively complex to analyse and estimate the true cost for either a proposed new system or for an existing system. Some of the reasons for this:

- Various costs viz. capital cost, repairs & maintenance, system operator’s wage, resale price,

- cost of finance- associated with the present and proposed systems

- Repairs & maintenance involves a variety of repairs ranging from small to big which cannot be defined and computed exactly in specific

- The influence of various economic variables such as Inflation, value of money etc.

- Random element in breakdowns and subsequent repairs: two identical systems maintained in an apparently similar manner may have different repair cost.

- Change in government policies and other external factors may change the decision in unexpected manner.

Given these inherent complexities, it is unrealistic to expect the systems maintenance manager, a technical guy, or his advisors to make these calculations on regular basis. Although computer program can facilitate these calculations, the data requirements are a big problem and the results are open to misinterpretation. However, presently, the decisions being made on intuition would seem less than satisfactory.

PROBLEM DEFINITION


Accordingly the objective of this study reported here was to develop a more realistic mathematical model for group replacement of a block of computers. An attempt was made to bring the maintenance cost more realistic by introducing two intermediary repairs viz. minor repair and major repair. Markov process (chain),a stochastic process, is employed to compute the probabilities of transition from a given state a at time t-1 to any future state b at time t.

Replacement Policy: The maintenance cost increases with time. Then the replacement policy will be (i) replace if the next period’s cost is greater than the weighted average of previous costs (ii) do not replace if the next period’s cost is less than the weighted average of the previous costs.

Development of Group (Block) replacement Model using Markov Process (Chain)

A group replacement model is applicable to the items and the result is the group replacement age for the entire group of items irrespective of whether the items are functioning or not. However in the computers and computer based system it involves some states where they can be repaired and can be reusable with the same efficiency. Though it is difficult to identify the specific repaiable intermediate states, to make the model simplistic two intermediate states i.e.

minor repair state and major repair state are introduced. Non-working of USB ports, key board,

mouse, colour flickering, network card connections, LAN card etc. are considered as minor repairs where as Non-working of mother-board, monitor, XGA/SVGA/VGA cards, SMPS, memory modules etc. and processor overheating etc. are considered as major repairs.Markov process (chain), a stochastic process, is employed to compute the probabilities of transition from a given state a at time t-1 to any future state b at time t.

Mechanics of Markov process (chain)Markov process is a stochastic or random process, which has property that the probability of transition from a given state to any future state depends on the present state and not on the manner in which it was reached. The first order Markov Process is based upon the following assumptions:

(i) The set of possible outcomes is finite(ii) The probability of next state depends only on the

immediately preceding state(iii) The transition probabilities are constant over

time(iv) The second order Markov process assumes that

the probability of next state may depend on the two previous states. Likewise a third order Markov process assumes that the probability of next state can be calculated by obtaining and taking account of the past three states.Thus if t0 < t1 < . . . < tn represents the points on time scale then the family of random variables {X(tn ) } is said to be a Markov process provided it holds the Markovian property:

P{X(tn) = Xn | X(tn-1) = Xn-1 ,…, X(t0) = X0 } = P{X(tn) = Xn | X(tn-1) = Xn-1 }

for all X(t0), X(t1) ,…,X(tn)

The simplest of the Markov Process is discrete and constant over time. A system is said to be discrete in time if it is examined at regular intervals, e.g. daily, monthly or yearly.Mathematically the probability Px(n-

1),x(n) = P{X(tn) = Xn | X(tn-1) = Xn-1 } is called transition probability that represents the probability of moving from one state to another.

The transition probabilities can be arranged in a matrix form and such a matrix can be called as Transition Probability Matrix(TPM).

In this paper, a group replacement model for the items (N) that fail completely on usage, considering two intermediate states i.e. minor repair and major repair, is developed by using first order Markov chain. So as it consists 4 states, the Transistion Probability Matrix(TPM) can be formulated as

I II III IV

TPM = P = I P11 P12 P13 P14

II P21 P22 P23 P24


III P31 P32 P33 P34

IV P41 P42 P43 P44

where I, II, III and IV represents Working, Minor repair, Major repair, and Complete failure States. Spectral Decomposition Method can be used to calculate the values of Pi , where i = 1 to n

Spectral Decomposition Method:

It is a method of Matrix Decomposition based on eigen values and hence it is also called as Eigen Decomposition. It is applicable to square matrix. In this matrix will be decomposed into a product of three matrices, only one of which is diagonal. As a result, the decomposition of a matrix into matrices composed of its eigen values and eigen vectors is called Eigen or Spectral decomposition.An n x n matrix P always has ‘n’ eigen values, which can be ordered ( in more than one way) to form an n x n diagonal matrix D and a corresponding matrix of non zero columns V that satisfies the eigen value equation PV = VD.

Assume P has non-degenerate eigen values ë1 , ë2, ë3 , . . ., ëk and corresponding linearly independent eigen vectors X1 , X2, X3 , . . ., Xk which can be denoted

Define the matrices composed of eigenvectors

V =

=

and eigen values

giving the amazing decomposition of P into a similarity transformation involving Vand D,

P = V D V-1

The fact that this decomposition is always possible for a square matrix Pas long as Vis a square matrix is known in this work as the eigen decomposition theorem.

Furthermore, squaring both sides of above equation gives

P2 = (V D V-1 ) (V D V-1 )

= V D (V-1V) D V-1

= V D2 V-1

Mathematically, Spectral Decomposition can be represented as Pi = V Di V-1 where i= 1 to n

Higher order Transition Probability Matrix (TPM) of four state Markov Chain using Spectral Decomposition Method:

If P represents the four state TPM then the higher order transition probabilities are obtained by the following procedure.


i. Determine the eigen values of the Transition Probability Matrix ‘P’ by solving ÐP-ëIÐ= 0

ii. If all eigen values say ë1 , ë2, ë3 , . . ., ëk are distinct then obtain k-column vectors say X1 , X2, X3

, . . ., Xk corresponding to the Eigen values by solving PV = VD or PX = ëX where X ‘“ 0

iii. Denote these column vectors (eigen vectors by matrix V) where V= (X1 , X2, X3 , . . ., Xk ) and obtain V-1

iv. Compute D, a diagonal matrix formed from the eigen values of P.

v. Higher order Transition Probability Matrix (TPM) of four state Markov chain can be computed using the equation, Pi = V Di V-1 where i= 1 to n

Probability(X) of units during a specific period (i) for various states:

Let XiI = proportion of units under working

condition at the end of ith time period

XiII = proportion of units under minor repair at the end

of ith time period

XiIII = proportion of units under major repair at the

end of ith time period, and

XiIV = proportion of units under complete failure at

the end of ith time period

Probability(X) of units during a specific period (i) for various states can be calculated as

[ XiI Xi

II XiIII Xi

IV ] = [ X0I X0

II X0III X0

IV ] Pi , where i = 1 to n

In general, Xi = X0 x Pi

XiI Xi

II , XiIII and Xi

IV can be calculated using software MATLAB(proprietary software) or OCTAVE(Open source software).

OCTAVE Program for calculation of Xi :

clc

P=[ P11 P12 P13 P14 ; P21 P22 P23 P24 ; P31 P32 P33 P34 ; P41

P42 P43 P44 ]

X=[ X0I X0

II , X0II

, X0IV ]

eig(P)

[V,D]=eig(P)

IV=inv(V)

for i=1:15

Pi=V * D^n *IV

Xi=X*Piend

Replacement Decision:


Total cost for n time periods, TC: = NC4 + C1*Sai

+ C2 *S bi + C3*Sli , i = 1 to n andwhere C1 = Individual replacement cost per unitC2 = Minor repair costC3 = Major repair costC4 = Group replacement costWeighted average cost, W(t) = TC / nPolicy: sum of the individual replacement cost and repair cost in t+1 period should be greater than the average cost in tth period, to group replace in tth period.Model Development:Notations:N = Total number of items in the systemC1 = Individual replacement cost per unitC2 = Minor repair costC3 = Major repair costC4 = Group replacement cost X0

I = proportion of units under working condition initiallyX0

II = proportion of units under minor repair initiallyX0

III = proportion of units under major repair initiallyX0

IV = proportion of units under complete failure initiallyXi

I = proportion of units under working condition at the end of ith time periodXi

II = proportion of units under minor repair at the end of ith time periodXi

III = proportion of units under major repair at the end of ith time period, andXi

IV = proportion of units under complete failure at the end of ith time period

Pij = Probability of items switching over from i th

state to jth state in a periodW(t) = Weighted average cost per period in group replacement policy,As the repair costs of items falling in intermediate states are varying appreciably, depending on the severity of the break down, repairs are considered in two states like minor and major repairs. So as it consists 4 states, the TPM can be formulated asI II III IV

TPM = P = I P11 P12 P13 P14

II P21 P22 P23 P24

III P31 P32 P33 P34

IV P41 P42 P43 P44

where I, II, III and IV represents Working, Minor repair, Major repair, and Complete failure States.Probability(X) of units during a specific period(i) for various states can be calculated as[ Xi

I XiII Xi

III XiIV ] = [ X0

I X0II X0

III X0IV ] Pi

, where i = 1 to nIn general, Xi = X0 x Pi

The values of TPMs (Pi , where i = 1 to n) are calculated by using Spectral decomposition Method and using these TPMs, probabilities of items falling in different states Xi

I XiII , Xi

III and Xi

IV are calculated using OCTAVE, an open source software.Total Cost (TC) =Individual replacement cost + Minor repair cost + Major repair cost + Group replacement cost

Number of individual Replacements:

1st period, a1 = N X 1IV

2nd period, a2 = N X 2IV + a1 X 1

IV

3rd Period, a3 = N X 3IV + a1 X 2

IV + a2 X 1IV

4th Period, a4 = N X 4IV + a1 X 3

IV + a2 X 2IV + a3 X 1

IV

Number of Minor Repairs:

1st period, b1 = N X 1II

2nd period, b2 = N X 2II + a1 X 1

II

3rd Period, b3 = N X 3II + a1 X 2

II + a2 X 1II

4th Period, b4 = N X 4II + a1 X 3

II + a2 X 2II + a3 X 1

II

Number of Major Repairs

1st period, l1 = N X 1III

2nd period, l2 = N X 2III + a1 X 1

III


3rd Period, l3 = N X 3III + a1 X 2

III + a2 X 1III

4th Period, l4 = N X 4III + a1 X 3

III + a2 X 2III + a3 X 1

III

Total cost for n time periods, TC: = NC4 + C1*Sai + C2 *S bi + C3*Sli , where i = 1 to n

Weighted average cost, W(t) = TC / n

Policy: sum of the individual replacement cost and repair cost in t+1 period should be greater than the average cost in tth period, to group replace in tth

period.

Case Study:N = 1000 itemsC1 = Rs.40000; C2 = Rs.1000; C3 = Rs.8000; C4

= Rs.30000 ;

[ X0I X0

II X0III X0

IV ] = [0.810 0.110 0.070 0.010]P1 = TPM

= 0.88890 0.06790 0.03700 0.00620 0.63640 0.22730 0.09090 0.04550 0.57140 0.32860 0.07140 0.02860 0.00000 0.00000 0.00000 1.00000Probability(X) of units during a specific period(i) for various states can be calculated as [ Xi

I XiII Xi

III XiIV ] = [ X0

I X0II X0

III X0IV ] Pi ,

where i = 1 to nThe values of TPMs (Pi , where i = 1 to n) are calculated by using Spectral decomposition Method and using these TPMs, probabilities of items falling in different states Xi

I XiII , Xi

III and XiIV are calculated

using OCTAVE, an open source software. TPM’s and Xi s are shown in Appendix I).The calculations for the above data are tabulated in the table #1.

* Costs in the table are in lacs

From the Table #1 as the average annual cost is less during 10th year, it can be inferred that the existing system can be used for a period of 10 years and can be replaced with the new one at the end of 10th year.

CONCLUSIONA Markov chain based mathematical model for

group replacement model has been developed for a block of computers system. This paper considers four

discrete states- working condition, minor repair and major repair and break down- of a computer system to make the maintenance cost more realistic.

There are several other aspects, such as influence of macro economic variables viz. inflation, time value

of money etc. and obsolescence of present computer system due to the advent of new


technology, which are to be considered for further study of maintenance cost.

REFERENCES1. David L Epstein and Sharyn O’Halloran (2005):

Higher order Markov Models”, paper presented at the political methodology summer meetings, Tallahasse, Florida, July 21-23, 2005.

2. Zhenqing Li and and Weiming Wang (2005):”Computer aided solving the high-order transition probability matrix of the finite Markov Chain”, Elsevier journal of Applied Mathematics and Computation (Article in Press)

3. Bruce A Craig and Peter P Sendi(1998): “Estimating the Transition Matrix of A homogeneous Markov Chain”, Technical Report #98-12, Department of Statistics , Purude University, June 1998

4. Sutawanir Darwis and Kulsan(2008): “Spectral Decomposition of Transition Matrix”, Journal Matematika Dan Sains, September 2008, Vol 13, No. 3, pp 97-101.

5. Liana Cazacioc and Elena Corina Cipu (2004): “Evaluation of the Transition Probabilities for Daily precipitation Time series using a Markov chain Model”, Proceedings of 3rd International Colloquium “Mathematics in Engineering and Numerical Physics”, October 7-9, 2004, pp:82 to 92

6. Lawrence K Saul & Michael I Jordan(1999): “Mixed memory Markov Models: Decomposing Complex Stochastic Processes as Mixtures of Simpler Ones”, Machine learning 37,1998, pp:75-86

7. Andre Berchtold & Adrain Raftery(2002): “The Mixture Transition Distribution (MTD) model for High-order Markov Chains and Non Gaussian Time series”, Statistical Science, vol. 17, No.3, 2002, PP:328-356

8. Avik Ghosh Dastidar, Deepanwita Ghosh, S Dasgupta and UK De (2010): “Higher order Markov Chain Models for monsoon rainfall over West Bengal, India”, Indian Journal of Radio and Space Physics, Vol. 39, February 2010, pp 39-44.

9. Stelios H Zanakis and Martin W Maret (1980): “A Markov Chain Application to Manpower Supply Planning”, The Journal of the Operational Research Society, Vol. 31, No. 12, Dec. 1980, pp1095-1102

10 Jianming HU, Jingyan SONG, Guoqiang YU & Yi Zhang (2003): “A Novel Networked Traffic parameter Forecasting Method based on Markov Chain model”, IEEE Transactions, 2003, P 3595-3600.

11. P L Nuthall, KB Woodford and AC Beck (1983): “Tractor replacement Policies and cost Minimisation”, Discussion paper no.74, Agricultural Economics research unit, Lincoln College, New Zealand, Nov. 1983, ISSN 011-7720.

12. Yue Pan and Marlin U Thomas (2010): “Repair and Replacement Decisions for warranted products under Markov Deterioration”, IEEE transactions on reliability, Vol. 59, No.2, June 2010, PP:368-373

13. Thomas W Archibald & Rommert Dekkar (1996): “Modified Block replacement for Multiple-component systems”, IEEE transactions on reliability, Vol. 45, No.1, 1996, PP:75-83

14. Isha Bagai and Kanchan jain (1994): “Improvement, deterioration and Optimal replacement under age- replacement with

minimal repair”, IEEE transactions on reliability, Vol. 43, No.1, March 1994, PP:156-162

15. Ruey Huei Yeh, Gaung-Cheng Chen & Ming-Yuh Chen (2005): “Optimal age-replacement policy for Non-repairable products under

16. http://en.wikipedia.org/wiki/Markov_chain


renewing free-replacement warranty”, IEEE transactions on reliability, Vol. 45, No.1, March 2005, PP:92-97.

17.

http://en.wikipedia.org/wiki/Matrix_decomposition

18.

http://mathworld.wolfram.com/EigenDecomposition.html

19. http://mathworld.wolfram.com/Eigenvalue.html

COMPARATIVE STUDY OF FACE REPRESENTATION METHODS FOR EFFICIENT FACE RECOGNITION- SURVEY PAPER

T.SYED AKHEEL1,S.A.K JILANI 2, V.VENKAT RAMI REDDY 3

1,2 P.V.K.K Institute of Technology,Anantapur 3 JNTU,Engineering College ,Hyderabad


ABSTRACT:

Face is an important biometric feature for personal identification. Human beings easily detect and identify

faces in a scene but it is very challenging for an automated system to achieve such objectives. The need for

reliable identification of interacting users is obvious. The research on face recognition has been actively going

on in the recent years because face recognition spans numerous fields and disciplines such as access control,

surveillance and security, credit-card verification, criminal identification and digital library.Many methods

have been proposed for face recognition but each has its own drawback due to the variety of uncontrolled

scenarios such as illumination, pose variations and occlusions.In this paper we discuss past research on

biometric face feature extraction and recognition of static images. We will present implementation outline of

these methods along with their comparative measures and result analysis.

Key words :Automatic face recognition, Appearance based recognition, Principal component Feature extraction statistical approaches ,template based approaches),and feature based methods eigenface fisherface Fisher’s Linear Discriminant (FLD) ,singular values, fractional Singular value.

1. INTRODUCTION

The developing of face recognition system is quite difficult because human faces is quite complex, multidimensional and corresponding on environment changes. For that reason the human machine recognition of human faces is a challenching problem due the changes in the face identity and variation between images of the same due to illumination , pose variations and some natural effects.The issues are how the features adopted to represent a face under environmental changes and how we classify a new face image based on the chosen representation.Computers that recognize human faces systems have been applied in many applications such as security system, mug shot matching and model-based video coding.

The eigenfaces is well known method for face recognition. Sirovich and Kirby [1] had efficiently representing human faces using principle component analysis. M.A Turk and Alex P. Pentland [2] developed the near real-time eigenfaces systems for

face recognition using eigenfaces and Euclidean distance.Most effort in the literature have been focused mainly on developing feature extraction methods and employing powerful classifiers such as probabilistic hidden Markov models (HMMs) [4,36] neural networks (NNs) [3,5] and support vector machine (SVM) [4,37]. The main trend in feature extraction has been representing the data in a lower dimensional space computed through a linear or non-linear transformation satisfying certain properties.

Statistical techniques have been widely used for face recognition and in facial analysis to extract the abstract features of the face patterns. Principal component analysis (PCA) [1][7],[8] and linear discriminant analysis (LDA) [3][7] are two main techniques used for data reduction and feature extraction in the appearance-based approaches. Eigen-faces and fisher-faces [6] built based on these two techniques, have been proved to be very successful. LDA algorithm selects features that are most effective for class separability while PCA selects features important for class representation. A study in [10] demonstrated that PCA might outperform LDA when the number of samples per class is small and in the case of training set with a large number of samples, the LDA still outperform the PCA. Compared to the PCA method, the computation of the LDA is much higher [4] and PCA


is less sensitive to different training data sets. However, simulations reported in [4] demonstrated an improved performance using the LDA method compared to the PCA approach. When dimensionality of face images is high, LDA is not applicable To resolve this problem we combine the PCA and LDA methods, by applying PCA to preprocessed face images, we get low dimensionality images which are ready for applying LDA. Finally to decrease the error rate in spite of Euclidean distance criteria this was used in [4]. A system is implemented using neural network to classify face images based on its computed LDA features.Kirby and Sirovich [11] showed that any particular face can be (1) economically represented along the eigenpictures coordinate space, and (2) approximately reconstructed using just a small collection of eigenpictures and their corresponding projections (‘coefficients’).Turk and Pentland [12] applied PCA technique to face recognition, and proposed the well-known eigenfaces method. A recent major improvement on PCA is to directly manipulate on two-dimensional matrices (not one-dimensional vectors as in traditional PCA), e.g., two-dimensional PCA (2DPCA) [13], generalized low rank approximation of matrices [14], non-iterative generalized low rank approximation of matrices (NIGLRAM) [15] and so on. The advantages of manipulating on two-dimensional matrices rather than one-dimensional vectors are [13]: (1) it is simpler and straightforward to use for image feature extraction; (12) it is better in terms of classification performance; and (13) it is computationally more efficient.

Based on the viewpoint of minimizing reconstruction error, the above PCA-based methods [12,13–15] are unsupervised methods that do not take the class labels into consideration.Taking the class labels into consideration, LDA aims at projecting face samples to a subspace where the samples belonging to the same class are compact while those belonging to different classes are far away from each other. The major problem in applying LDA to face recognition is the so-called small sample size (SSS) problem (namely, the number of samples is far less than sample dimensionality), which leads to the singularities of the within-class and between-class scatter matrices. Recently, researchers have exerted great endeavor to deal with this problem. In [6,7], a

PCA procedure was applied prior to the LDA procedure, which led to the well known PCA+LDA or Fisher faces method.In [7,8], samples were first projected to the null space of the within-class scatter matrix and then LDA was applied in this null space to yield the optimal (infinite) value of the Fisher’s linear discriminant criterion, which led to the so-called discriminant common vectors (DCV) method. In [19, 20], LDA was applied in the range space of the between-class scatter matrix to deal with the SSS problem, which led to the LDA via QR decomposition (LDA/QR) method. In [3] a general and efficient design approach using a radial basis function (RBF) neural classifier to cope with small training sets of high dimension, which is a problem frequently encountered in face recognition, is presented. In order to avoid over-fitting and reduce the computational burden, face features are first extracted by the principal component analysis (PCA) method. Then, the resulting features are further processed by the Fisher’s linear discriminant (FLD) technique to acquire lower-dimensional discriminant patterns. These DR methods have been proven to effectively lower the dimensionality of Face Image. Furthermore, in face recognition, PCA and LDA have become de-facto baseline approaches.

However, despite of the achieved successes, these FR methods will inevitably lead to poor classification performance in case of great facial variations such as expression, lighting, occlusion and so on, due to the fact that the face images are very sensitive to these facial variations.

It is illustrated that the eigen value of an image are not necessarily be stable hence discrimination of images affected by illumination and other said factors is very difficult.

These DR methods have been proven to effectively lower the dimensionality of Face Image. Furthermore, in face recognition, PCA and LDA have become de-facto baseline approaches. However, despite of the achieved successes, these FR methods will inevitably lead to poor classification


performance in case of great facial variations such as expression, lighting, occlusion and so on, due to the fact that the face image A on which they manipulate is very sensitive to these facial variations. It is illustrated that the eigen value of an image are not necessarily be stable hence discrimination of images affected by illumination and other said factors is very difficult. A paradigm is proposed [2] called Singular value decomposition (SVD) based which uses the singular values(SVD consists of finding the eigenvalues and eigenvectors) for feature extraction which represent algebraic properties of an image and have good stability and good discrimination ability was obtained. But In [3] it is illustrated that singular values of an image are stable and represent the algebraic attributes of an image, being intrinsic but not necessarily visible. Moreover SVs are very sensitive to facial variations such as illumination, occlusions, thus it gives the good discrimination results only when the illumination effect is uniform. Based on the observations a new method is proposed [1] method in which the weights of the facial variation sensitive base images (SVs) are deflated by a parameter á called fractional order singular value decomposition representation (FSVDR) to alleviate facial variations for face recognition and gives the good classification result even in non-uniform effects.

2.FACE RECOGNITION SYSTEM DESIGN

Due to the complexity of the face recognition problem, a modular approach was taken whereby the system was separated into smaller individual stages. Each stage in the designed architecture performs a intermediate task before integrating the modules into a complete system.

The face recognition system developed performs three major tasks pre-processing of given face image, extracting the face feature for recognition, and performing classification for the given query sample. The system operates on two phase of operation namely training and testing phase.

The functional blocks of the proposed systems is as follows NN Classifier

Figure 2.1 General Face Recognition Systems2.1 Face Recognition Using Eigen Features

2.2.1 Overview

When designing a complex system, it is important to begin with strong foundations and reliable modules before optimizing the design to account for variations. Provided a perfectly aligned standardized database is available, the face recognition module is the most reliable stage in the system. Biggest challenge in face recognition still lies in the normalization and preprocessing of the face images so that they are suitable as input into the recognition module. Hence, the face recognition module was designed and implemented first.Eigenface approach is one of the earliest appearance-based face recognition methods, which was developed by M. Turk and A. Pentland [12] in 1991. This method utilizes the idea of the principal component analysis and decomposes face images into a small set of characteristic feature images called eigenfaces. The eigenface method simply evaluates the entire image as a whole. These properties make this method practical in real world implementations. The basic concept behind the


eigenface method is information reduction. When one evaluates even a small image, there is an incredible amount of information present. From all the possible things that could be represented in a given image, pictures of things that look like faces clearly represent a small portion of this image space. Because of this, we seek a method to break down pictures that will be better equipped to represent face images rather than images in general. To do this, we generate “base-faces” and then represent any image being analyzed by the system as a linear combination of these base faces [33].Once the base faces have been chosen we have essentially reduced the complexity of the problem from one of image analysis to a standard classification problem. Each face that wish to classify an be projected into face-space and then analyzed as a vector. A neural network approach is used for classification. The technique can be broken down into the following components:1) Generate the eigenfaces2) Project training data into face-space to be used with a predetermined classification method3) Evaluate a projected test element by projecting it into face space and comparing to training data

The idea of using eigenfaces was motivated by a technique for efficiently representing pictures of faces using principal component analysis. It is argued that a collection of face images can be approximately reconstructed by storing a small collection of weights for each face and a small set of standard pictures. Therefore, if a multitude of face images can be reconstructed by weighted sum of a small collection of characteristic images, then an efficient way to learn and recognize faces might be to build the characteristic features from known face images and to recognize particular faces by comparing the feature weights needed to (approximately) reconstruct them with the weights associated with the known individuals.

The eigenfaces approach for face recognition involves the following initialization operations:[33].

1. Acquire a set of training images.

2. Calculate the eigenfaces from the training set, keeping only the best M images with the highest eigen values. These M images define the “face space”. As new faces are experienced, the eigenfaces can be updated.

Calculate the corresponding distribution in M-dimensional weight space for each known individual

(training image), by projecting their face images onto the face space.

Having initialized the system, the following steps are used to recognize new face images:

1. Given an image to be recognized, calculate the eigen features of the M eigenfaces by projecting the it onto each of the eigenfaces.

2.

2. Determined features are further processed using pca so as reduce the dimension of the image so as to have more samples since more eigenfaces will always produce greater classification accuracy, since more information is available

3. However, the eigenface paradigm, [3] which uses principal component analysis (PCA), yields projection directions that maximize the total scatter across all classes, i.e., across all face images. In choosing the projection which maximizes the total scatter, the PCA retains unwanted variations caused by lighting, facial expression,and other factors [3]. Accordingly, the features produced are not necessarily good for discrimination among classes. In [3], [12],the face features are acquired by using the fisherface or discriminant eigenfeature paradigm. This paradigm aims at overcoming the drawback of the eigenface paradigm by integrating Fisher’s linear discriminant (FLD) criteria, while retaining the idea of the eigenface paradigm in projecting faces from a high-dimension image space to a significantly lower-dimensional feature space.

4. These features are classified by using RBF classifier as either a known person or as unknown. The goal of using neural networks is to develop a compact internal representation of faces,which is equivalent to feature extraction. Therefore, the number of hidden neurons is less than that in either input or output layers, which results in the network encoding inputs in a smaller dimension that retains most of the important information. smaller dimension that retains most of the important information. Then,


the hidden units of the neural network can serve as the input layer of another neural network to classify face images.

2.2.2 Principal Component Analysis

Until G. Bors and M. Gabbouj [4] applied the Karhunen-Loeve Transform to faces, face recognition systems utilised either feature-based techniques, template matching or neural networks to perform the recognition. The groundbreaking work of Kirby and Sirovich not only resulted in a technique that efficiently represents pictures of faces using Principal Component Analysis (PCA), but also laid the foundation for the development of the “eigenface” technique of Turk and Pentland [1], which has now become a de facto standard and a common performance benchmark in face recognition .Starting with a collection of original face images, PCA aims to determine a set of orthogonal vectors that optimally represent the distribution of the data. Any face images can then be theoretically reconstructed by projections onto the new coordinate system. In search of a technique that extracts the most relevant information in a face image to form the basis vectors, Turk and Pentland proposed the eigenface approach, which effectively captures the variations within an ensemble of face images.

2.2.2.1 Calculating Eigenfaces

Mathematically, the eigenface approach uses PCA to calculate the principal components and vectors that best account for the distribution of a set of faces within the entire image space. Considering an image as being a point in a very high dimensional space, these principal components are essentially the eigenvectors of the covariance matrix of this set of face images, which Turk and Pentland [12] termed the eigenface. Each individual face can then be represented exactly by a linear combination of eigenfaces, or approximately, by a subset of “best” eigenfaces – characterized by its eigenvalues, Let a

face image ( )yx,Γ be a two-dimensional N by N array of intensity values. An image may also be considered as a vector of dimension, so that a typical image of size 256 by 256 becomes a vector of dimension 65,536, or equivalently, a point in 65,536-dimensional space. An ensemble of images, then, maps to a collection of points in this huge space.face images, being similar in overall configuration, will not be randomly distributed in this huge image space and thus may be described by a relatively low dimensional subspace. The main idea of the principal

component analysis is to find the vector that best account for the distribution of face images within the entire image space.

These vectors define the subspace of face images,

which we call “face space”. Each vector is of length ,

describes an N by N image, and is a linear

combination of the original face images. Because

these vectors are the eigenvectors of the covariance

matrix corresponding to the original face images, and

they have face-like in appearance, hence referred as

“eigenfaces”.

Let the training set of face images be,, 1Γ 2Γ 3Γ MΓ , .

The average face of the set if defined by .

∑=

Γ=ΨM

nnM 1

1

(2.1)

Figure 2.1 Mean Image

Each face differs from the average by the

vector. An example training set is shown in

Figure 1a, with the average face shown in

Figure 1b. This set of very large vectors is then

subject to principal component analysis, which

seeks a set of M orthonormal vectors, , which

best describes the distribution of the data. The

kth vector, is chosen such that


∑=

Φ=M

nn

Tkk M 1

2)(1 µλ

(2.2)

is a maximum, subject to

=

=otherwise

klk

Tl ,0

,1µµ

(2.3)

The vectors and scalars are the eigenvectors and eigenvalues, respectively, of the covariance matrix

∑=

=ΦΦ=M

n

TTnn AA

MC

1

1

(2.4)

where the matrix . The matrix C, however, is by , and determining the eigenvectors and eigenvalues is an intractable task for typical image sizes. A computationally feasible method is needed to find these eigenvectors.

If the number of data points in the image space is less than the dimension of the space (), there will be only , rather than , meaningful eigenvectors (the remaining eigenvectors will have associated eigenvalues of zero). Fortunately, we can solve for the -dimensional eigenvectors in this case by first solving for the eigenvectors of and M by M matrix—e.g., solving a 16 x 16 matrix rather than a 16,384 x 16,384 matrix—and then taking appropriate linear combinations of the face images . Consider the eigenvectors of such that (2.5)

Premultiplying both sides by A, we have

nnnT AAAA νλν = (2.6)

from which we see that are the eigenvectors of .Following this analysis, we construct the M by M matrix , where , and find the M eigenvectors of L. These vectors determine linear combinations of the M training set face images to form the eigenfaces :

MnA n

M

kknkn ,......,1,

1==Φ= ∑

=ννµ

(2.7)

Figure 2.2 Eigen Faces

The associated eigenvalues allow us to rank the eigenvectors according to their usefulness in characterizeing the variation among the images.

2.2.3 Feature Extraction by FLD

The PCA paradigm [3] does not provide any information for class discrimination but dimension reduction. Accordingly, the FLD[3] is applied to the projection of the set of training samples in the eigenface space .

( ) nrnXXXXX ×ℜ⊂= ,......,, 321 The paradigm

finds an optimal subspace for classification in which the ratio of the between-class scatter and the within-class scatter is maximized.


Let the between class scatter matrix be defined

asT

ll

C

i

iB XXXXnS ))((

1−−= ∑

= (2.8)and the within-class scatter matrix be defined ass

Tll

C

i

iiW XXXXnxS ))((

1−−∈= ∑ ∑

= ( )∑

==

c

jiXnX

1/1

(2.9) Where is the

mean image of the ensemble, s∑

=

=in

j

iji

iX

nX

1

1

and is the mean image of the ith class and c is the number of classes. The optimal subspace by the FLD is determined as fallow

[ ]1,21 ......,arg max −== C

WT

BT

Eoptimal eee

RSE

ESEE

(2.10)

where (e1,e2,e3,…ec-1) is the set of generalized eigenvectors of and corresponding to the c-1 largest generalized eigen values λi=1,2,3,….c-1 i.e.

(2.11)

Thus the feature vectors P for any query face image Z in the most discriminant sense can be calculated as follows:

(2.12)

2.2.4 Face Image Classification

The eigenface images calculated from the eigenvectors of L span a basis set with which to describe face images. As mentioned before, the usefulness of eigenvectors varies according their associated eigenvalues. This suggests we pick up only the most meaningful eigenvectors and ignore the rest, in other words, the number of basis functions is further reduced from M to M’ (M’<M) and the computation is reduced as a consequence.

Experiments have shown that the RMS pixel-by-pixel errors in representing cropped versions of face images are about 2% with M=115 and M’=40.In practice, a smaller M’ is sufficient for identification, since accurate reconstruction of the image is not a requirement. In this framework, identification becomes a pattern recognition task. The eigenfaces span an M’ dimensional subspace of the original image space. The M’ most significant eigenvectors of the L matrix are chosen as those with the largest associated eigenvalues.A new face image Γ is transformed into its eigenface components (projected onto “face space”) by a simple operation for n=1,……,M’. This describes a set of point-by-point image multiplications and summations.

The weights form a vector that describes the contribution of each eigenface in representing the input face image, treating the eigenfaces as a basis set for face images. The vector may then be used in a standard pattern recognition algorithm to find which of a number of predefined face classes, if any, best describes the face.

Figure 2.2 Eigen Faces

The associated eigenvalues allow us to rank the eigenvectors according to their usefulness in characterizeing the variation among the images.

2.2.3 Feature Extraction by FLD

The PCA paradigm [3] does not provide any information for class discrimination but dimension reduction. Accordingly, the FLD[3] is applied to the projection of the set of training samples in the eigenface space .The paradigm finds an optimal subspace for classification in which the ratio of the between-class scatter and the within-class scatter is maximized.

Let the between class scatter matrix be defined

as

(T

ll

C

i

iB XXXXnS ))((

1−−= ∑

= 2.8)


and the within-class scatter matrix be defined ass

Tll

C

i

iiW XXXXnxS ))((

1−−∈= ∑ ∑

= (2.9) Where is the

( )∑=

=c

jiXnX

1/1

mean image of ∑

=

=in

j

iji

iX

nX

1

1

the ensemble, and is the mean image of the ith class and c is the number of classes. The optimal subspace by the FLD is determined as fallows

[ ]1,21 ......,arg max −== C

WT

BT

Eoptimal eee

RSE

ESEE

(2.10)

where (e1,e2,e3,…ec-1) is the set of generalized eigenvectors of and corresponding to the c-1 largest generalized eigen values λi=1,2,3,….c-1 i.e.

(2.11)

Thus the feature vectors P for any query face image Z in the most discriminant sense can be calculated as follows:

(2.12)

2.2.4 Face Image Classification

The eigenface images calculated from the eigenvectors of L span a basis set with which to describe face images. As mentioned before, the usefulness of eigenvectors varies according their associated eigenvalues. This suggests we pick up only the most meaningful eigenvectors and ignore the rest, in other words, the number of basis functions is further reduced from M to M’ (M’<M) and the computation is reduced as a consequence. Experiments have shown that the RMS pixel-by-pixel

errors in representing cropped versions of face images are about 2% with M=115 and M’=40.In practice, a smaller M’ is sufficient for identification, since accurate reconstruction of the image is not a requirement. In this framework, identification becomes a pattern recognition task. The eigenfaces span an M’ dimensional subspace of the original image space. The M’ most significant eigenvectors of the L matrix are chosen as those with the largest associated eigenvalues.

A new face image Γ is transformed into its eigenface components (projected onto “face space”) by a simple operation for n=1,……,M’. This describes a set of point-by-point image multiplications and summations.

The weights form a vector that describes the contribution of each eigenface in representing the input face image, treating the eigenfaces as a basis set for face images. The vector may then be used in a standard pattern recognition algorithm to find which of a number of predefined face classes, if any, best describes the face.The simplest method for determining which face class provides the best description of an input face image is to find the face class k that minimizes the Euclidian distance where is a vector describing the face class. The face classes are calculated by averaging the results of the eigenface representation over a small number of face images (as few as one) of each individual. A face is classified as “unknown”, and optionally used to create a new face class.Because creating the vector of weights is equivalent to projecting the original face image onto to low-dimensional face space, many images (most of them looking nothing like a face) will project onto a given pattern vector. This is not a problem for the system; however, since the distance between the image and the face space is simply the squared distance between the mean-adjusted input

image Ψ−Γ=Φ i ∑

==Φ

M

iiij

1µω

and, its projection

onto face space: . 22

fΦ−Φ=εThus there are four

possibilities for an input image and its pattern vector: (1) near face space and near a face class; (2) near face space but not near a known face class; (3) distant from face space and near a face class; (4) distant from face space and not near a known face class. In


the first case, an individual is recognized and identified. In the second case, an unknown individual is present. The last two cases indicate that the image is not a face image. Case three typically shows up as a false positive in most recognition systems; in this framework, however, the false recognition may be detected because of the significant distance between the image and the subspace of expected face images. So, the eigenfaces approach for face recognition could be summarized as follows:

Collect a set of characteristic face images of the known individuals. This set should include a number of images for each person, with some variation in expression and in the lighting (say four images of ten people, so M=40).

Calculate the (40 x 40) matrix L, find its eigenvectors and eigenvalues, and choose the M’ eigenvectors with the highest associated eigenvalues (let M’=10 in this example).Combine the normalized training set of images according to Eq. (6) to produce the (M’=10) eigenfaces

For each known individual, calculate the class vector by averaging the eigenface pattern vectors [from Eq.(2.8)] calculated from the original (four) images of the individual. Choose a threshold that defines the maximum allowable distance from any face class, and a threshold that defines the maximum allowable distance from face space [according to Eq. (2.9)].

For each new face image to be identified, calculate its pattern vectorthe distance to each known class, and the distance to face space. If the minimum distance and the distance, classify the input face as the individual associated with class vector.If the minimum distance but, then the image may be classified as “unknown”, and optionally used to begin a new face class. If the new image is classified as a known individual, this image may be added to the original set of familiar face images, and the eigenfaces may be recalculated (steps 1-4). This

gives the opportunity to modify the face space as the system encounters more instances of known faces.

1.2.5 Experimental Results

For the implementation of face recognition a well known face database called YALE face database [31] is used. YALE face database contains 165 gray level face images of 15 persons. There are 11 images per subject, and these 11 images are, respectively, under the following different facial expression or configuration: center-light, happy, left-light,glasses, normal, right-light, sad, sleepy, surprised, and wink. In this implementation, all images are sized to a size of 128 x 128. For classification or recognition purpose for the given training and test samples

1. For each given input, Mean Square Error (MSE) in calculated between network output and the target vector.MSE index should have either global minimum or maximum. To find these values Gradient decent is calculated.

Table 2.1 Result analysis for Eigen based approach


Figure 2.3 : Training and Test Performance2. Out of the 12 faces, 8 are correctly classified in the 1st match. Hence the total accuracy for the eigen feature based approach is 66.67% for illuminated affected Yale database. Thus in real time scenarios this method may be inappropriate for the illuminated affected database.

3. FACE RECOGNITION USING SINGULAR FEATURES

3.1 Introduction

The singular value decomposition is a outcome of linear algebra. It plays an interesting fundamental role in many different applications. On such application is in digital image processing.

SVD in digital applications provides a robust method of storing large images as smaller, more manageable square ones. This is accomplished by reproducing the original image with each succeeding nonzero singular value. Furthermore, to reduce storage size even further, images may approximated using fewer singular values.

3.2 Singular Value DecompositionThe singular value decomposition of a matrix A of m x n matrix is given in the form,

∑= TVUA

(3.1)Where U is an m x m orthogonal matrix; V an n x n orthogonal matrix, and å is an m x n matrix

containing the singular values of A and along its main diagonal. A similar technique, known as the eigenvalue decomposition (EVD), diagonalizes matrix A, but with this case, A must be a square matrix. The EVD diagonalizes A as

1−= VDVA(3.2)

Where D is a diagonal matrix comprised of the eigenvalues, and V is a matrix whose columns contain the corresponding eigenvectors. Where Eigen value decomposition may not be possible for all facial images SVD is the result.

3.2.1 SVD Working PrincipleLet A be an m x n matrix. The matrix ATA is symmetric and can be diagonalized. Working with the symmetric matrix ATA, two things are true:1. The eigenvalues of ATA will be real and nonnegative.2. The eigenvectors will be orthogonal.

To derive two orthogonal matrices U and V that diagonalizes a m x n matrix A,First, if it is

required to factor A as ∑= TVUA then the following must be true.

)()( TTTT VUVUAA ∑∑= TTTT VUUVAA ∑∑=

TTT VVAA ∑∑=

TT VVAA 2∑= ∑= TVUA (3.3)

this implies that Σ2 containes the eigenvalues of ATA and V contains the corresponding eigenvectors. To

find the V nrrr λλλλλλ ...... 2121 ==≥≥≥ ++ of the svd,rearrange the eigenvalues of ATA in order of decreasing magnitude and some eigenvalues are set equal to zero. Define the singular values of A as the square root of the corresponding eigenvalues of the matrix ATA; that is,

wherejj ,λσ =

nj ,......2,1= (3.4)


re-arranging the eigenvectors of ATA in the same order as their respective eigenvalues to produce the matrix

[ ]nrrr VVVVVVV ,......,,,......, 2121 ++= (3.5)

Let the rank of A be equal to r. Then r is also the rank of ATA, which is also equal to the number of nonzero eigenvalues.

jj λσ = [ ]rVVVV ,,........., 21= Let and be the set of eigenvectors associated with the non-zero eigenvalues and be the set of eigenvectors associated with zero eigenvalues. It follows that:

),......,,( 212 nrr AvAvAvAV ++=

)0,....0,0(2 =AV

0= (3.6)

where this zero is the zero matrixes.It is defined earlier that the matrix to be the diagonal matrix with the singular values of A along its main diagonal. From above equation, each zero eigenvalue will result in a singular value equal to zero. Let be a square r x r matrix containing the nonzero singular

values{ }rσσσ ,......,, 21 of A along its

main diagonal. Therefore matrix Σ may be represented by:

∑=∑

0001

(3.7)

Where the singular values along the diagonal are arranged in decreasing magnitude, and the zero singular values are placed at the end of the diagonal. This new matrix Σ, with the correct dimension m x n, is padded with m - r rows of zeros and n - r columns of zeros. To find the orthogonal matrix U. Looking at the equation it follows that:

[ ] [ ]

= ++

0

0,.., ..,. .,. .,,. .,

1

!111

⋯

⋯

rnrrnrr uuuuvvvvA

σ

σ

[ ] [ ]0...0,,...,,...,...,..., 1111 rrnrr uuAvAvAvAv σσ=+ (3.8)

Therefore, ri ,......1=

jjj uAv σ= (3.9)

Examining above equation is a scalar value and that and is column vectors and a matrix vector multiplication results in another vector. Therefore, the vector resulting from the multiplication of is equal to the vector multiplied by the scalar.It could be observed at the vector as lying in the direction of the unit vector uj with absolute length. can be calculated from previously found matrix V. Therefore, the unit vectoris a result of dividing the vector by its magnitude.. (3.10)

This Equation is restricted to the first r nonzero singular values. This allows to finding the column vectors of so long as there is no division by zero. Therefore this method allows finding only part of the matrix U.

To find the other part where the singular values of A are equal to zero. As seen before in the matrix V, the matrix U may be defined as:

(3.11)

Letting [ ]211 ,...,uuU =

[ ]mr uuU ,...,12 +=

[ ]rvvAAv ,......,11 =

[ ]rAvAv ,......,1=

[ ]rruu σσ ,......,11=


[ ]

=

r

ruuσ

σ⋯

1

1 ,......,

111 ∑= UAv

Then (3.12)

referring to the illustration of the four fundamental subspaces the null space of a matrix A, denotes the set of all nontrivial (non-zero) solutions to equation .Using above equation

where zero represents a zero matrix (3.13)

It follows that V2 forms a basis for the Also because

(3.14)where , and .The orthogonal complement to the is since the columns in the matrix V are orthogonal, the remaining vectors must lie in the subspace corresponding to the From above equation, we see that .This equation holds the valuable information that the column vectors of U, are in the column space of A. This is because the column vectors of U are linear combinations of the columns of A. or, in matrix notation where .It now follows that and are orthogonal complements. Since the matrix U is an orthogonal matrix and the first r column vectors of U have been assigned to lie in the , then must lie in the .

The vectors that lie in the are the vectors which form the matrix.Once matrix V is derived, the matrix Σ, and the matrix U, the singular value decomposition can be found for any matrix A. where actually does diagonalize and equal the matrix A.

[ ]

∑= T

T

VVUU

2

1121 00

0,

[ ]

∑=00

, 121

TVUU

0111 +∑= TVU

TVU 111 ∑=

TVAV 11=

A=

AVU T =∑ (3.15)

The rank r is equal to the number of nonzero eigenvalues referring to four fundamental subspaces,it observed that are the eigenvectors corresponding to the nonzero eigenvalues of A, and that the remaining column vectors of V correlate with the eigenvalues equal to zero. So there exists an r nonzero singular value. Thus r is equal to the number of nonzero eigenvalues termed as rank of the matrix.

These SVD features are used for facial feature decomposition to represent an image in dimensionality reduction (DR) factor. An SVD operation breaks down the matrix A into three separate matrices.

∑= TVUA

[ ]

=

Tn

T

n

nn

v

vuu

11

,......σ

σ⋮

[ ]

=Tnn

T

n

v

vuu

σ

σ⋮

11

1,......,

Tnnn

T vuvu σσ ++= ...111

Trrr

T vuvu σσ ++= ...111 because nr σσ ....!+ are equal to zero

(3.16)

Singular values at each iteration are obtained as follows.


1. At 1st iteration SV values the facial information provided is given by

TvuA 111σ=

2. After n=2 iteration the facial approximation is given by,

TT vuvuA 222111 σσ +=

3. After n=3 iteration the facial approximation is given by,

TTT vuvuvuA 333222111 σσσ ++=

From the above observations it could be observed that the facial information’s are though presented in high leading images such as the eye, mouth and nose regions they are less definitive to facial expressions. So a same face image with facial variation may not be predicted in such SV approach. To overcome this limitation the SVD based face recognition approach is modified to FSVD approach as presented below.

3.3 Fractional Order Singular Value Decomposition (FSVDR)

To alleviate the facial variations on face images, a novel FSVDR is suggested.

The main ideas of FSVD approach are that;

(2) The weights of base images corresponding to relatively small σi’s should be inflated, since they may be less sensitive to the facial variations within A.

(3) The order of the weights of the base images in formulating the new representation of SVD should be retained. More specifically, for each face image matrix A which has the SVD, its FSVD ‘B’ can be defined as,

(3.16)

Where U, Σ and V are the SV matrices, and in order to achieve the above underlying ideas, α is a fractional parameter that satisfies

It is seen that the rank of FSVDR ‘B’ is r, i.e., identical to the rank of A as the B matrix is fractional raised the values are inflated retaining the rank of the matrix constant.The form a set of which are similar to the base images for the SVD approach.It is observed that the intrinsic characteristic of A, the rank, is retained in the FSVD approach. In fact it has the same like base images as A, and considering the fact that these base images are the components to compose A and B, the information of A is effectively been passed to B. From the observation it could be observed that:

(1) The FSVD is still like human face under lower SV.

(2) The FSVD deflates the lighting condition in vision. Taking the two face images (c) and (d) under consideration, when α is set to 0.4 and 0.1, from the FSVD alone, it is difficult to tell whether the original face image matrix A is of left light on or right light on.

(1) The FSVD reveals some facial details. In the original face images (a) presented, neither the right eyeball of the left face image nor the left eyeball of the right face image is visible, however, when setting α to 0.8 and 0.1 in FSVD, the eyeballs become visible.

In the case of FSVD thus the fractional parameter and it’s optimal selection is an important criterion in making the face recognition process more accurate.

3.4 Fractional Parameter ‘α’

In FSVD, is a key parameter that should be adjusted. On a suitable selectivity of α parameter the recognition system can achieve superior performance to existing recognition performance. Further, in images (which are sensitive to facial variations) are deflated but meanwhile the discriminant information contained in the leading base images may be deflated. Some face images have great facial variations and are perhaps in favor of smaller ’s, while some face images have slight facial variations and might be in favor of larger’s. The learned from the training set is a tradeoff among all the training samples and thus is


only applicable to the unknown sample from the similar distribution. Each DR method has its specific application scope, which leads to the difficulty in designing a unique selection criterion for all the DR methods. As a result, the criterion for automatic choosing should be dependent on the training samples, the given testing sample and the specific DR method. To optimally choose the value minimum argument MSV criterion is used.Mean square variance (MSV) criterion state that,

(3.17)

where is the standard variance of the ith class defined as

(3.18) Where and respectively, denote the element of the d-dimensional samples and class mean ,C is the number of classes, and is the number of training samples contained in the class. For an optimal selection of value the MSV value must be optimally chosen. The smaller MSV value represents, compact the same class samples are, and on the contrary, the bigger MSV is, the looser the same class samples. When the same class samples are very loose, these samples will lead to biased estimation of the class mean, within class and between-class scatter matrices, while on the contrary, when the same class samples are compact, the estimation of the class mean, within-class and between-class variance matrices may be much more reliable.

When the same class samples are compact, it is more likely that these samples can nicely depict the Gaussian distribution from which they are generated and considering the fact it is essential for the same class samples to be compact, namely MSV to be small in the recognition methods. The MSV of the training samples is given as ,

( )αα

αMSVopt minarg=

(3.19)

3.5 Operational Evaluation

The FSVD based recognition approach is observed to be more effective in face recognition compared to the

existing approach due to the fact that the FSVD approach works on the simple principle of deflating the more dominant leading base image (i.e. the higher order ) and inflating the lower ’s. As these lower’s are content of low variations which are facial expressions in the given face image. As FSVD inflate these ’s the expression or illumination, which are completely neglected in previous, approach resulting in lower estimation accuracy are overcome in FSVD approach. The proposed FSVD based recognition approach is found to be very effective in case of intermediate feature extraction for face recognition. This feature extraction could then be used as a information for recognition systems such as PCA, LDA, NN etc. The FSVD based recognition approach is focused to overcome the effects of various real time factors in face image. Though this technique is found to inflate the low so as to reveal the expression affects this method is found to be of same computation time as compared to the existing recognition system.

When employing FSVDR as a recognition approach for face recognition method, the time complexities in training and testing are almost the same as the existing methods. The time complexity for training N samples with dimensionality is and the time complexity in testing any given unknown sample is , where C is the number of classes. For and FSVD based system it consumesin computing the FSVD for N samples where is usually smaller than N,

and thus the time complexity in training ( )dNT 2is still

as with the original recognition system, on the other

hand, for any unknown sample, it takes in ( )dNT 2

computing

FSVD, and thus the time complexity in testing is

also ( )dcT since is usually comparable to or less than C.

Fractional singular value approach for face recognition involves the following initialization operations:

1. Acquire a set of training images.


2. Calculate the singular values from the training set, keeping only the best M images with the highest values. These M images define the “face space”.

3. Calculate the corresponding distribution in M-dimensional weight space for each known individual (training image), by projecting their face images onto the face space. Having initialized the system, the following steps are used to recognize new face images:

4. Given an image to be recognized, calculate the singular features of the M faces by projecting the it onto each of the faces.

5. These SVs features are raised by a fractional parameter α hence called fractional order based method. On selectivity of this parameter this recognition system can achieve superior performance by deflating the more dominant leading base features i.e. higher order SVs and inflating the lower SVs. and are tested at α = [0.5,0.8,1]

6. The proposed system consist of five steps in addition to the four steps existing in the above two methods called fractional order singular value decomposition representation (FSVDR) which acts an intermediate representation between face images and data representation for face recognition.

7. Determined features are further processed using pca so as reduce the dimension of the image so as to have more samples since more number of samples will always produce greater classification accuracy, 8. However, principal component analysis (PCA), yields projection directions that maximize the total scatter across all classes, i.e., across all face images. In choosing the projection which maximizes the total scatter, the PCA

3. Corresponding FSV projection in face-space samples having right side light on at α = 1

Figure 3.6 FSVD Projections for 8 iterations

3 . Corresponding FSV projection in face-space samples having left side light on at α = 1

Figure 3.7 FSVD Projections for 8 iterations a

Table 3.2 Result analysis for FSVD approach

From above observation it is found that at α = 1 the projected in the face spaces looks like an original images. Hence the recognition accuracy is more as compared to previous two methods.


Figure 3.8 Performance Evaluations

3. Out of the 12 faces, 11are correctly classified in the 1st match at α = 1.Total accuracy for the fractional singular feature based approach is 91.96% for illuminated affected Yale database.

ConclusionThe objective of this paper is to analyze developed face recognition systems. Here, the effect of environmental and surrounding effect due to illumination effect is been focused and to minimize the effect a face recognition approach is suggested.For the elimination of the additional illuminitation effects due to external surrounding effect is been also focused and is processed using singular value decomposition approach. The variation of the lighting effect could be reduced by the normalization of eigen feature values. in further improving effort the SVD approach is further improved to Fractional –SVD approach is developed.For the nono-linearlity effect of face lighting and face expression effect. The variation of these factors are observed to be effective in face recognition approach. To evaluated the performance of the developed approach a evaluation is carried out for recognition accuracy over various face images with lighting effect and expression variation. For more accurate results these systems are to analyzed thoroughly and tested different databases with different effects.

Acknowledgement

We are thankful all the authors who made available and gave an opportunity to study and analyze their papers and helped in writing this survey paper.

REFERENCE

1. Jun Liu, Songcan Chen, Xiaoyang Tan “Fractional order singular value decomposition 1. representation for face recognition” ELESVER Journal 26 March 2007

2. Y. H. Wang, T. N. Tan and Y. Zhu “Face Verification Based on Singular Value Decomposition and Radial Basis Function Neural Network” National Laboratory of Pattern Recognition (NLPR) Institute of Automation, Chinese Academy of Sciences.

3. Meng Joo,Shiqian Wu ,Juwei Lu and Hock Lye Toh “Face Recognition With Radial Basis Function (RBF) Neural Networks” IEEE Transactions on Neural Networks Vol. 13, No. 3, May 2002 697.

4. A. Hossein Sahoolizadeh, B. Zargham Heidari, and C. Hamid Dehghani “A New Face Recognition Method using PCA, LDA and Neural Network” International Journal of Computer Science and Engineering 2:4 2008.

5. M. Er, S. Wu, J. Lu, L.H.Toh, “face recognition with radial basis function(RBF) neural networks”, IEEE Trans. Neural Networks, Vol. 13, No. 3, pp. 697-710.


6. P.N. Belhumeur, J.P. Hespanha, and D. J. Kriegman, “Eigen faces vs. Fisher faces: Recognition using class specific linear projection”, IEEE Trans. Pattern Anal. Machine Intel., vol. 19, PP. 711-720, may 1997.

7. W. Zhao, R. Chellappa, A, Krishnaswamy, “Discriminant analysis of principal component for face recognition”, IEEE Trans. Pattern Anal. Machine Intel., Vol 8, 1997.

8. M.J.Er, W.Chen, S.Wu, “High speed face recognition based on discrete cosine transform and RBF neural network”, IEEE Trans. On Neural Network, Vol. 16, No. 3, PP. 679,691, 2005.

9. D.L. Swets and J.J. Weng”Using Discriminant Eigen features for image retrieval”, IEEE Trans. Pattern Anal. Machine Intel, vol. 18, PP. 831-836, Aug. 1996.

10. J.J. Weng”using discriminant eigenfeatures for image retrieval”, IEEE Trans. Pattern Anal. Machine Intell. , Vol. 18, No. 8, pp. 831-836,1996.

11. M. Kirby, L. Sirovich, Application of the Karhunen–Loeve procedure for the characterization of human faces”,IEEE Trans. Pattern Anal. Mach. Intell. 12 (1990) 103–108.

12. M. Turk, A. Pentland, “Eigenfaces for recognition”, J. Cognitive Neurosci. 3 (1) (1991) 71–96.

13. J. Yang, D. Zhang, “Two-dimensional pca: a new approach to appearance-based face representation and recognition”, IEEE Trans. Pattern Anal. Mach. Intell. 26 (1) (2004) 131–137.

14. J.Ye, “Generalized low rank approximation of matrices”, International Conference on Machine Learning, pp. 2004, pp. 887–894.

15. J. Liu, S. Chen, “Non-iterative generalized low rank approximation of matrices”, Pattern Recognition Lett. 27 (9) (2006) 1002–1008.

16. D. Swets, J. Weng, “ Using discriminant eigenfeatures for image retrieval”,IEEE Trans. Pattern Anal. Mach. Intell. 18 (1996) 831–836.

17. H. Cevikalp,M.Neamtu,M.Wilkes, A. Barkana, “ Discriminative common vectors for face recognition” , IEEE Trans. Pattern Anal. Mach. Intell. 27 (1) (2005) 4–13.

18. J. Liu, S. Chen” Discriminant common vecotors versus neighbourhood components analysis and laplacianfaces: a comparative study in small sample size problem”, Image Vision Comput. 24 (3) (2006) 249–262.

19. J. Liu, S. Chen, “Resampling lda/qr and pca+lda for face recognition”, The 18th Australian Joint Conference on Artificial Intelligence, 2005, pp. 1221–1224.

20. J. Ye, Q. Li, “ A two-stage linear discriminant analysis via qrdecomposition”, IEEE Trans. Pattern Anal. Mach. Intell. 27 (6) (2005) 929–941.


21. Z. Hong,”Algebraic feature extraction of image for recognition”,Pattern Recognition 24 (1991) 211–219.

22. Y. Cheng, K. Liu, J. Yang, Y. Zhuang, N. Gu, “Human face recognition method based on the statistical model of small sample size”, Intelligent Robots and Computer Vision X: Algorithms and Techniques, 1991, pp. 85–95.

23. Y. Tian, T. Tan, Y. Wang, Y. Fang,”Do singular values contain adequate information for face recognition”, Pattern Recognition 36 (3) (2003) 649–655.

24. Berk GÄokberk,”Feature Based Pose Invariant Face Recognition”,Bogazici University, 1999

25. Hong,Z.,” Algebraic Feature Extraction of Image for Recognition”, Pattern Recognition, Vol. 24, pp. 211{219, 1991.

26. Daugman, J. G., “ Complete discrete 2D Gabor transforms by neural networks for image analysis and compression”, IEEE Transactions on Acoustics, Speech, andSignal Processing, Vol. 36, pp. 1169{1179, 1988.

27. Muhammad Tariq Mahmood “Face Detection by Image Discriminating”, Blekinge Institute of Technology Box 520 SE – 372 25 Ronneby Sweden

28. C.Gonzalez and Paul Wintz. “Digital Image Processing.” Addison-Wesleypublishing company, second edition, 1987.

29.S.Haykin “ Neural Networks, A Comprehensive

Foundation Network” Macmilla, 1994.

30. Mohammed Aleemuddin,”A Pose Invariant Face Recognition system using Subspace Techniques”, King Fahd University of Petroleum and Minerals Dhahran, Saudi Arabia November 2004

31. Min Luo, “Eigenfaces for Face Recognition”

32. Steven Tjoa, “Statistical Pattern Recognition:Face Recognition” Dept. of Electrical and Computer Engineering, University of Maryland December 12, 2007.


33.http://cvc.yale.edu/projects/yalefaces/yalefaces.html

34. www.dtreg.com

35.http://www.wiau.man.ac.uk/ ct.Kernel principal component analysis.

36.H. Othman, T. Aboulnasr, “ A separable low complexity 2D HMM with application to face recognition” IEEE Trans. Pattern. Anal. Machie Inell., Vol. 25, No. 10, PP. 1229-1238, 2003.

37.K. Lee, Y. Chung, H. Byun, “SVM based face verification with feature set of small size”, electronic letters, Vol. 38, No. 15, PP. 787-789, 2002.[12].

ANALYSIS OF PHISICALLY

INFORMED PARAMETRIC SYNTHESIS OF SOUND EFFECTS

M.KAMAL BASHA.,

Principal, Arasu Engineering College, Kumbakonam.

DR.R.RANI HEMA MALINI.,

Head of the Department /E&I, St.Peter’s University, Chennai.

K.ANUJA BERKATH BANU.,

Asst.Professor/E&I, St.Peter’s University, Chennai.

ABSTRACT:

This paper presents algorithms and systems for automatic analysis and parametric synthesis of walking and other (gesture-rate) periodically modulated noisy sounds. A recording of walking is analyzed, extracting the gait (tempo and left/right asymmetries), heel-toe events, etc. Linear prediction is used to extract the basic resonances. Wavelet decomposition is performed, and a high frequency-subband is used to calculate statistics for a particle re synthesis model. Control envelopes are extracted from the original sound. A real-time synthesis program allows flexible re synthesis of walking sounds, controlled by a score extracted from a sound file, a graphical user interface, or


parameters from game/animation/VR data. Results for the analysis algorithm are presented for synthesisized data, and for hand-crafted real experimental sounds of gravel.

key words:

Graphical User Interface, Low Pass Filter, walking Sounds, Physically Inspired Stochastic Event Modeling.

INTRODUCTION

A staple of production for movies, stage performance, television, and radio dramas is the addition (often last minute) of artificial and natural sound effects. For, movies, “Foley” artists put shoes on their hands and “walk” in boxes of gravel, leaves, cornstarch (for the sound of snow), etc. in real time, effectively “acting” the sounds as they watch the scenes. Radio and stage sound effects performer/engineers use tape cartridgesor CDs of pre-recorded sounds, and physical noisemakers to add sounds in real time. For offline production and real-time dramas, these might indeed be the best means to add sounds. However, for virtual reality, training simulations, and games, the lack of parametric control, along with the need for large libraries of pre-recorded sounds and complex software for selecting the right sounds, forces inherent compromises in responsiveness and sonic quality.

One clear benefit of parametrized analysis/synthesis of sound effects is the possibility of significantly reducing the memory storage required for a large library of digital walking sounds. Greater benefits, however, include the ability to economically synthesize sounds in real time, coupled to parameters generated by sensors (virtual reality and training simulations), or data from games, animations, etc [1]. The synthesis method presented here is based on prior work by the author, but the primary novelty presented in this paper is the presentation of an algorithm for estimating the sonic granularity parameter (previously it had to be adjusted manually by ear), and the integration of the entire system into an interactive Graphical User Interface workbench. Related work [2][3] has investigated various background and noise like sounds such as walking, but the work presented

here is unique in the fundamental algorithm Linear

prediction is performed on each footstep sound to

determine the overall resonances for a resynthesis

filter, and to yield “whitened” (spectrally flat) step

sounds. Wavelet extraction of a high frequency band

is performed, and the (rectified) peaks are used to

estimate a “particle density.” New files with arbitrary

lengths, gaits, materials etc. can be generated

automatically, using the extracted average and

deviation envelopes. A real-time walking synthesis

program can be controlled by parameters from a

score file, a Graphical User Interface (GUI), data

from foot/ground collision detections in a game or

simulation, sensors in a virtual reality system.

2 ENVELOPE ANALYSIS

and parametric architecture, testing the

multi-step analysis algorithm on real and synthesized

data, and the integration of the analysis/ re synthesis

algorithms into a complete interactive system (for

which all source code is made publicly available for

free).

1 SYSTEM ARCHITECTURE

A sound file is analyzed to determine the individual footfalls, and the average gait (tempo plus left/right asymmetries). A control envelope score file is written. The sound file is marked and segmented into individual footstep events. The individual envelopes are averaged to yield an average control envelope and a standard envelope deviation.


Figure 1: Walking sound system architecture.The first stage of signal processing involves

extracting and analysing the overall envelope of a

walking sound. Figure 2 shows a simple non-linear

low-pass filter for performing envelope extraction.

The two signal dependent pole positions are similar

to rise/fall times on audio compressors and many

historical analog envelope followers [4], allowing the

output to rise faster when the signal is rising, and fall

slower when the signal is falling. This helps to ensure

that peaks are tracked accurately, while still

eliminating spurious high-frequency components.

Due to the low-pass filtering effects of the envelope

follower, and also the generally slow nature of

walking gestures and steps themselves, the extracted

envelope can be stored at a much lower sample rate

than the original sound.

Figure 2: Digital Envelope follower.

An autocorrelation of the envelope signal is performed,to get a rough periodicity estimate. Significant peaks (local maxima over a threshold) are marked in the envelope, then a set of “best peaks” is selected, to meet various criteria of periodicity and expected walking tempos. At this point the event-

marked envelope can be inspected, and peaks can be edited by hand if desired, though the algorithm is quite reliable over a fairly large class of walking sounds. Figure 3 shows some envelopes, automatically marked by the system.


Figure 3: Some extracted/marked envelopes

The envelope can be written out as a constant update

rate “gestural” score file (30 Hz. updates of MIDI

after touch, for example). To make the representation

parametric, the envelope is also segmented into

individual step events, bounded by a set of lowest

significant values coming just before marked peaks.

Again the same on straints on periodicity used to

mark the original peaks are applied to mark the cut

points.

The original footfall sound segments are stored as

individual sound files to allow identity re synthesis

later. Next, the individual envelopes are compared to

yield an average and standard deviation envelope, as

shown in Figure 4. Significant sub-events can

sometimes be found within the envelopes, such as the

Heel/Toe events marked in Figure 3. Events like this

have been shown to be important in the perception of

identity, age, sex, etc. purely from walking sounds

[5]. The parametric

prototype envelopes are stored as multiple (6-10)

break-

point linear envelopes based on ordered triplets of (time, amplitude, deviation),

Figure 4: Envelope prototyping.

3 PARTICLE MODELING The heart of the parameterization and resynthesis of

walking sounds described here is based on PhISEM


(Physically Inspired Stochastic Event Modeling) [6].

This type of modeling is quite natural since much of

the sound produced by walking on various textures

involves particle interactions, excited by friction/

pressure from the feet. Stochastic parameterization

means that we don’t have to model all of the particles

explicitly, but rather only the probability that

particles will make noise. For many particle systems,

this is modeled well by a simple Poisson distribution,

where sound probability is constant at each time step,

giving rise to an exponential probability waiting time

between events as shown in Figure 5.

Figure 5: Poisson sound event probability.

As a first sound analysis step in PhISEM, the

resonant structure is removed using Linear Prediction

[7], keeping the “whitened” residual as the “raw”

particle sound. To determine the proper filter order,

the LPC prediction order can be incrementally

increased until the residual power does not decrease

significantly. Figure 6 shows the waveforms and

spectra of a particle system before and after 2nd order

LPC processing. LPC is performed on a per-footstep

basis, writing the filter coefficients into the score file

along with the excitation envelope. For a given

walking material, and for generic resynthesis later, an

average LPC filter is also computed and stored.

3.1 Resonance Modeling

As a first sound analysis step in PhISEM, the

resonant structure is removed using Linear Prediction

[7], keeping the “whitened” residual as the “raw”

particle sound. To determine the proper filter order,

the LPC prediction order can be incrementally

increased until the residual power does not decrease

significantly. Figure 6 shows the waveforms and

spectra of a particle system before and after 2nd order

LPC processing. LPC is performed on a per-footstep

basis, writing the filter coefficients into the score file

along with the excitation envelope. For a given

walking material, and for generic resynthesis later, an

average LPC filter is also computed and stored.

3.2 Particle Modeling

Each particle noise event can be modeled as an

impulse, but to better “fill in” the sonic events , a

short exponentially decaying (ác) noise burst is used.

To model the total sound, we do not explicitly

overlap and add these events, because the sum of

exponentially decaying indepenent noises (with the

same exponential time constant) is an exponentially

decaying noise (with the same time constant). On

each collision, we simply add energy to the

exponential sound state, corresponding to the current

system energy (net kinetic energy of all particles in

the system). We keep and calculate the exponentially

decaying (ás) state variable representing the system

energy, modified (added to) by the control envelope.


Figure 6: Removing resonance by Linear Prediction.

So the collisions are modeled by calculating a

random number, and if that number is < the particle

constant N, the switch in Figure 7 closes for one

sample,

exciting the exponential noise source with

energy equal to the current system kinetic energy. At

this point the synthesis algorithm is complete (Fig 7).

Both exponential functions are modeled using simple

one-pole recursive filters (poles at αc and αs).

Figure 7: PhISEM synthesis.

4.1 Estimating N

N is estimated by inspecting a high-frequency band (5.5 – 11kHz) of the whitened footstep sounds. A Daub echies 4 wavelet filter bank [8] is used to split the signal into sub bands, and these sub bands are rectified. As can be seen by inspecting the top three rectified sub band outputs in Figure 8, the 5.5-11kHz band seems to capture the rapid collision events best. The method can be used to analyze sound files of any sample rate, but 22.05kHz or 44.1kHz files are best

because of the presence of the 5-11k band (best band for any sample rate above 22.05k.). For sample rates below 22k (of interest because of a growing number of lower bandwidth sound effects files available on the Web), the highest octave sub band is used for forming the best estimate of N. Next a threshold value is computed from the maximum peak and average values of the rectified sub band signal, and the number of peaks per second over the threshold are counted. This estimate of N, the density of collisions, tends to miss collisions as the probability N increases.

To correct for this, the actual N estimate is

calculated to be:

Nest = 2avgPeaks/blocks/10

Where “blocks” means that the average number of

peaks are only counted in blocks where significant

peaks occur. Equation 1 was arrived at by trial and

error, in a set of experiments involving extracting

known N values from simulations, as described in the

next subsection.

4.2 Verifying the System ID Techniques

Using the basic synthesis model shown in Figure

7,1050 soundfiles were synthesized using a simple

raised cosine excitation envelope. System parameters

were:

αc = constant at 0.95 (60 dB decay of 130 samples)

N = 2,4,8,16,64,256,1024

αs = 0.95,0.99,0.995,0.999,0.9995,0.9999

r = 0.7,0.8,0.9,0.95,0.99

f = 1000,2000,3000,5000,8000

where r and f are the pole radius and center frequency

of a single 2nd order resonant filter. The steps of


envelope extraction, 2nd order Linear Prediction,

rectified high-frequency 5.5-11kHz subband

extraction, and N estimation were performed on all

1050 synthesized files.As Figure 9 shows, the

average frequencies and resonances yielded by

Linear prediction are more accurate for higher

resonances, and for middle frequencies. Most of the

frequency errors were experienced on files with very

low resonance (pole = 0.7 or 0.8). In those cases,

frequency mis adjustment is not so important

anyway.

Table 1: Particle density estimates for synthesized data.

Table 1 and Figure 10 show the results of calculating

estimates of N (using Equation 1) for all 1050 files.

Figure 9: Resonance estimates for synthesized data.

Given that the synthesis is stochastic, the estimated averages for N are quite good. However, the relatively high standard deviations suggest that we should collect and average as many sound files as possible for any given walking condition. This means that the more individual footsteps we have on a recording, the better the overall estimate of particle density N for the material will be.

Figure 10: Particle density estimates, synthesized data.

4.3 “Real World” Analysis

To test the analysis techniques on non-synthesized data, two types of gravel were analysed. Large gravel rocks, averaging 1.5 cm diameter and 14 gm. weight, and small gravel rocks, averaging 0.3 cm diameter and 1 gm. weight were analysed. Ten shaking sounds of different gravel samples were analysed for each gravel type. For the large rocks, the center frequency and resonance was estimated by LPC to be 6460 +/- 701 Hz, and r = 0.932. N was estimated to be 23.22, with a standard deviation of 14.53. For the small rocks, the center frequency and resonance were estimated to be 12670 +/- 3264 Hz. and 0.843, and N was estimated to be 1068 with a standard deviation of 755.

5 CONTROLLING THE RESYNTHESIS

Once analysis of a walking sound is complete, the original soundfile can be exactly reassembled by


concatenating the stored wave segments. Depending on the amount and nature of the background noise in the original recording, often a highly realistic semiparametric resynthesis of arbitrary length and tempo can be made by randomly editing together the individual original footstep waves. For truly flexible parametric resynthesis, the extracted envelope parameters can be used to generate new step

envelopes, using the prototype average multibreakpoint envelope, perturbing it by the standard deviation parameters. These new envelopes drive thePhISEM model (Figure 7) set to the analysed resonance and particle density, yielding infinite possible new syntheses. Any analyzed parameter can be interactively edited and changed as desired to modify the resynthesis. Figure 12 shows “Bill’s GaitLab,” a TCL/Tk Graphical User Interface for controlling synthesis/re synthesis in real time. Much of the power of parametric synthesis comes from driving the parameters from controllers, or from other algorithmic processes. Examples might include automatically synthesizing footsteps for various materials and characters in a video game, based on the same physical processes that are calculated to animate the walking of those characters. A specific example involves using a pressure sensitive floor surface in an immersive virtual environment to automatically provide walking sounds,programmatically tied to the environment being simulated and displayed. Figure 13 shows the Princeton “PhOLIEMat” (Physically Oriented Library of Interactive Effects). The base mat senses the location and pressure of each of a variety of moveable tiles. Each tile feeds envelope parameters in real time to the appropriate walking sound, responsive to walking pressure.

6 CONCLUSIONS, FUTURE WORK AND A NOTE ON PERCEPTUAL REALISM

Work remains to be done in refining the envelope,

resonance, and particle parameter extraction

subsystems. The system fails on some sound files,

but still provides a good starting point for hand-

crafting the synthesis parameters. Clearly the largest

amount of

work remains to be done in evaluating the perceptual

quality of the synthesized sounds.

The purpose of this paper has been to describe a

system for analysis and synthesis of noisy,

enveloped, periodic sounds such as those created by

walking. Also this paper endeavoured to analyze (via

signal processing and statistical measures) the

techniques used for estimating the parameters.

Validation of the “realness” of the synthesized

interactive environmental sounds lies in the realm of

a huge psychoacoustic research agenda on the

perception of “realistic sound” itself. In fact,

there is not yet a proposed or accepted experimental

paradigm for asking questions like “is this sound

real?” or “is this sound more real than that one?”

Further, perception of sound effects, like much of

perception, is inherently a multi-modal phenomenon.

For example, presenting the sound of breaking glass

while displaying video of a dropped glass bowl

breaking could improve the perceived quality of

sound, or the sound could change the perception of

the image. Of course, Foley artists know this

instinctively and exploit the interaction between

sound and image for emotion effects. Interaction

with the sound synthesis adds even more perceptual

modes, most importantly the gestures and the

responsive “feel” of the controllers. Thus a true

understanding of the “sonic realness” of synthesized


sounds needs a large amount of perceptual research

to be designed and carried out.

There have been efforts in the past [9][10] and

various new efforts are underway [11][12] [13], to

evaluate the salience, dispensability vs.

indispensability, and other aspects of perceptual and

synthesis control features of environmental and

interactive sounds. The author looks forward to

performing more perceptual experiments in this area,

and to integrating any new findings from sound

perception research into future signal-processing and

systems work.

REFERENCES

[1] P. R. Cook, “Toward Physically-Informed

Parametric Synthesis of Sound Effects,” Invited

Keynote Address, IEEE Workshop on Applications

of Signal Processing to Audio and Acoustics,

October, 1999.

[2] Casey, M., Auditory Group Theory: with

Applications to Statistical Basis Methods for

Structured Audio, Ph.D. Thesis, MIT Media Lab,

February 1998.

[3] N. Miner, 1998, Creating Wavelet-based Models

for Real-time Synthesis of Perceptually Convincing

Environmental Sounds, Ph.D. Diss., University of

New Mexico.

[4] C. Anderton, Electronic Projects for Musicians,

Guitar Player Books, Saratoga, CA, 1978.


engineering today - · pdf filearshia azam1*, ch. d.v. paradesi rao2, j. amarnath3 finite...

Documents