chapter 1shodhganga.inflibnet.ac.in/bitstream/10603/4386/11/11_chapter 1.pdf · condition...
TRANSCRIPT
1
CHAPTER 1
INTRODUCTION
1.1 BACKGROUND
Manufacture of products undergoes different processes before they are assembled to
serve their intended function. Over the years, many new materials and processes have
been introduced. Latest machines have better process capability and are competent of
achieving the right fit between mating parts. Today’s’ industries are equipped with the
state of the art technology machines, with processes catering to the specific needs of the
product requirements, like form, fit, function. Some products demand stricter tolerances
and better surface quality. Designers demand specific surface finishes that produce the
right texture enhancing the product’s aesthetic appeal or satisfying the manufacturers’
assembly requirements. So one can say that, in today’s scenario, where manufacturing
processes have matured and quality of products need to be built into them, every aspect
of the process requires to be studied, monitored and be perfected. With the availability
of computer technology and its allied growth in the software industry, newer computing
techniques and algorithms push the technology to its limits and application engineers
are curious to study the impact of them in various situations that may interest them. As
manufacturing brings to life the various abstract designs, there exist a huge potential to
create newer and newer products by various processes. This moots the study of the
implication of new algorithms and techniques on these processes with a goal to
manufacture better products in a shorter time, keeping the cost aspects low and
complying with the quality requirements. This also opens up another related domain
called ‘condition monitoring’. Condition monitoring studies are carried out on
processes, machines, tools and the like.
Among the entire group of manufacturing processes, metal cutting play a significant
role. Metal cutting is defined as the removal of materials in the form of chips from a
workpiece in order to obtain a finished product with desired attributes of size, shape,
and surface roughness [1]. The cutting of metal requires a harder tool than the
workpiece and may be classified into two categories based on the number of points that
2
are in contact at the time of cutting of metals, viz. single point cutting tool and multi-
point cutting tools. Single point cutting tools are used in processes like, turning, boring,
shaping and planing, while multi-point cutting tools are used in processes like milling
and grinding. Though several investigations have been carried out to study the effect of
cutting parameters of a cutting tool of different materials and geometries, tool condition
monitoring is still a very active area of research. This is due to the developments made
in sensor technologies that assist monitoring of parameters and the availability of newer
and more efficient algorithms to analyse them.
1.2 CONDITION MONITORING OF SINGLE POINT CUTTING TOOL
Condition monitoring studies, in general, are used to reduce the down time of a
production machine, or to avoid the production of waste/rejected components. They are
termed as preventive maintenance, when a check is performed on regular time intervals
to avoid break down or loss of productivity. They are called trend monitoring if they
reveal an index value with time that can give a useful lead time in warning of an
incipient machine failure. They are called fault diagnosis when the focus is to find the
root cause of the problem that leads to the said failure. The root cause study is important
to eliminate or to avoid such failures in the future. Manufacturing processes depend on
a large number of parameters that have to be monitored and controlled to produce the
required results.
Condition monitoring of a single point cutting tool has been practised mainly to study
the wear pattern of the cutting tools, to assess the remaining life of the tool, and to study
the effect of geometry on the quality of the surface and the like. A tool condition
monitoring system is an information flow and processing system in which there is an
amalgamation of
1. Information source selection and acquisition (sensors and data collection)
2. Information processing and refinement (Signal processing and data collection)
3. Decision making based on the refined information (condition identification).
Many researchers have contributed to such studies by using pattern recognition
approaches like artificial neural network (ANN) and fuzzy logic. The general
requirements for these studies are to understand the machining parameters and their
effects on the quality of the surface. With the availability of computers and the
3
developments made in sensor technology, an indirect method to the estimation of tool
wear using different kinds of inputs like force, vibration or acoustic signals has
emerged. The techniques used for the prediction of surface roughness have also
improved considerably and contributed to a large extent.
1.3 SURFACE ROUGHNESS MEASUREMENT
Surface roughness is a quantification of the texture of a surface and is measured by the
vertical deviations of a real surface from its ideal form. Characterization of surface
topography is important in applications involving friction, lubrication, and wear [2]. If
these deviations are large, the surface is rough; if they are small the surface is smooth.
Roughness is the high frequency, short wavelength component of a measured surface.
Surface measurement techniques are grouped into contact and noncontact methods. The
measurement is more often done as an off-line process than on-line, but it is essential to
study the online surface roughness prediction without interrupting the ongoing process
so that it can be used for flexible manufacturing systems and mass production. Some of
the ways by which the measurement problems are approached are [3]:
1. By correlating the cutting vibrations with the surface roughness.
2. By the use of image processing techniques.
3. By using a non-contact inductance pickup.
4. By using ultrasonic sensing approach
5. Direct stylus measurement.
The most popular method of surface measurement is by using an amplified stylus
profilometer. It is a contact type of measurement device and is easy to use and relatively
inexpensive [4]. This instrument uses a tracer or pickup incorporating a diamond stylus
and a transducer. Running the stylus tip across the workpiece surface generates
electrical signals corresponding to surface roughness. The electrical signals are
amplified, converted from analog to digital, processed according to an algorithm, and
displayed. The measurement has a fairly good resolution and a large range that satisfies
the measurements of most manufactured surfaces. However, this stylus profilometer has
its own limitations because it takes a huge amount of time to scan large areas and has a
limited range of use on non-flat surfaces, and more importantly, its usage is restricted to
4
off-line processes [5]. Off-line and in-process measurements are compared in the next
section.
1.4 IN-PROCESS VERSUS OFF-LINE MEASUREMENT
Monitoring can be performed in-process or off-line using direct or indirect methods [6].
Critical need for in-process tool and process monitoring has developed since computer
numerically controlled (CNC) machines and automated machining centers have become
more widespread [7]. Monitoring individual machining processes in real time is critical
to integrating those processes into the overall machining system. The in-process sensing
method means that the measuring is carried out while the workpiece is being machined
(or during normal disengagement) without interrupting the process. Off-line methods
can be performed on the machine or away from the machine. In either case, off-line
methods require either scheduled idle time or interruption of the process for
measurement. In-process and off-line methods are effective in gathering important
information about surface characteristics, but in-process methods are largely preferred.
In-process monitoring provides real-time information concerning the machining
process. This real-time feedback enables the machinist to adjust the appropriate
machining parameters to produce the desired surface roughness, reduce tool wear,
reduce the probability of tool breakage or reduce the production components with poor
surface quality. However, monitoring or measurement conducted in-process or off-line
would not be possible without the use of sensory devices.
1.5 PROBLEM DEFINITION:
The problem consists of two phases of study; a tool status classification problem
using supervised learning method and a surface roughness prediction problem using
vibrations signals, flank wear and cutting parameters.
The tool status classification study focuses on bringing out the best “feature-
classifier” combination of the vibration signals acquired during turning operation
under different tool conditions that give the maximum classification accuracy.
The second phase consists of prediction of surface roughness using cutting
parameters, flank wear and the statistical features obtained from vibration signals
acquired during the turning process. The study sets out to determine the
5
predictability of surface roughness using three regression techniques viz. multiple
linear regression (MLR), support vector regression (SVR) and radial basis network
(RBF). The influence of statistical parameterlearnings extracted from the vibration
signals in predicting the surface roughness is determined. The effect of coating on
surface roughness is compared by using carbide tipped tool and coated carbide
tipped tool.
1.6 SALIENT FEATURES OF WORK DONE:
As mentioned in the above section there are two main objectives that this research has
set out to investigate:
1. The best possible ‘feature-classifier’ combination that has the highest classification
accuracy in determining the status of the tool wear in a single point carbide cutting
tool.
2. The statistical features, histogram features and wavelet features that were extracted
out of the vibration signals acquired during turning operation were studied for their
classification accuracy with different classifiers like ID3 algorithm, Naive Bayes,
Bayes Net, Support Vector Machine, Artificial Neural Network (ANN) and fuzzy
classifier.
3. Four pre-determined classes viz., Good (Good tool), TblentH (TB1- Tool that is
blunt high), TblentL (TB2 - Tool that is blunt low) and Ttiploose (TTL - Tool tip
that has got loosened) were considered.
4. Feature reduction was carried on extracted features to reduce the computational
effort.
5. SVM algorithm was used to find the effect of kernel selection (linear, sigmoidal,
radial basis function and polynomial) among nu-SVM and c-SVM. The degree of
Polynomial kernel was varied and the appropriate degree that produces the highest
classification accuracy with the lowest computational time was obtained.
6. Wavelet transforms was used to study the signals in the time-frequency domain to
consider the spectral variations. Two types of features were defined using Discrete
Wavelet Transform; the DWT-energy feature and the DWT-entropy feature and
were investigated using, Daubechies, Symlets, Coiflets, Biorthogonal, Reverse
Biorthogonal and Meyer (dmey) wavelets. The classification accuracies were
6
compared for all the different wavelets and tabulated to find the best among them.
This completed the first phase of the study.
7. The prediction of surface roughness using cutting parameters, taking into account
the flank wear (FW) of the tool and the statistical features extracted from the
vibration signals formed the second phase of the study.
a. Prediction of surface roughness for a particular speed, feed and depth of cut
combination, and by considering flank wear of the tool along with the statistical
features extracted from the vibration signal was carried out using MLR, SVR
and the RBF techniques. The root mean square error values (RMSE) obtained
for the three techniques were compared to find the most suitable regression
technique for this kind of study.
b. Study using entire data set and considering all machining parameters (speed,
feed, depth of cut (DOC)), FW, and statistical features of the vibration signal
using MLR and by including
only the mean of the vibrations (Case 1)
the first four moments of the statistical features and speed, feed, DOC,
FW (Case 2)
all the statistical features and speed, feed, DOC, FW (Case 3)
features selected by PCA and speed, feed, DOC, FW (Case 4)
The purpose of this part of the study was to find the effect of including various
statistical parameters in the regression equation. PCA was used to find the effect
of feature reduction on the coefficients of determination and RMSE values
obtained.
c. Surface roughness was predicted using the entire data set considering all
parameters (speed, feed, DOC, flank wear of the tool) and the statistical features
extracted from the vibration signal to evaluate the best regression technique
among, MLR, SVR and RBF techniques.
d. Comparison between coated and non-coated RMSE results of the entire set to
find the effect of coating using MLR.
7
1.7 STRUCTURE OF THE THESIS:
The structure of the thesis has been organised as follows:
An overview of the thesis has been presented in Chapter -1, comprising of an
introduction to condition monitoring of cutting tool, the definition of the problem under
study, salient features of the work done and the structure of the thesis.
The background of the work done by other researchers in the field of condition
monitoring of cutting tool and prediction of surface roughness during turning is
presented in Chapter - 2. It also discusses the scope of the present work under the said
backdrop.
Chapter - 3 discusses the methodology involved in machine learning approach towards
tool wear classification involving feature extraction and feature selection. An overview
of the various classification algorithms that are used to classify the data is also
presented here.
Chapter - 4 describes the details of experimental set up, the CNC machine, FFT (fast
Fourier transform) analyser, the vibration signal acquisition method and the design of
experiments using different cutting parameters. Two sets of experiments were
conducted; 1) study on the classification of tool wear status, intended to bring out the
feature-classifier combination, and 2) prediction of surface roughness using MLR, SVR
and RBF techniques to find the best among them that has the highest predictability. The
first set consisted of four sets of readings corresponding to the four tool wear states. The
second set of experiment was a full factorial study for two types of tool tips, viz. carbide
and coated carbide tipped tools. For each tool 81 sets of vibration signals were taken
along with their corresponding surface roughness (Ra) readings.
Chapter - 5 presents the results and discussion of the classification results of tool wear
states through various classification methods using statistical and histogram features.
The classifiers used were J48, ANN, SVM, Fuzzy, Bayes Net, Naive Bayes etc. It also
discusses the effect of kernel function selection among c-SVM and nu-SVM.
8
Chapter - 6 contains a detailed study using wavelet features. Two types of DWT feature
definition were considered, the DWT-energy and the DWT-entropy feature and a
comparison among them were carried out. The wavelets considered were, Coiflet
wavelets, Daubechies wavelets, bi-orthogonal wavelets, reverse bi-orthogonal wavelets,
symlet wavelets, and Meyer wavelets. Amongst these, all except Meyer wavelet have
child wavelets, thus a total of 52 wavelets were considered using all combinations into
account. The best among these that produces the highest classification accuracy for the
study of tool condition monitoring was chosen.
Chapter-7 explains the studies related to the prediction of surface roughness using
multiple linear regression and its related terminologies. The mathematical
representation of Support Vector Regression and Radial basis network algorithms are
also discussed.
Chapter -8 presents the results and discussion regarding the prediction of surface
roughness using cutting parameters, flank wear of the tool and the statistical features of
the vibration signals for both non-coated and coated carbide tipped tool using multiple
regression, SVR and RBF algorithm and the comparison among the three algorithms.
Chapter-9 gives the conclusion of the study and summary of findings. Suggestions for
enhancement of the present study and future scope are discussed. This is followed by a
reference section and Appendix
1.8 CONCLUDING REMARKS
This chapter introduces the topic of condition monitoring and pattern recognition
methods. The problems that this research is set out to investigate is mentioned. The
salient features of the work done along with the structure of thesis were briefed.