an algebraic denoising scheme

5
 An algebraic denoising scheme J. Cort´ es-Romer o, C. Garc´ ıa-Rodr´ ıguez, A. Luvian o-Ju ´ arez, R. Portillo-V´ elez and H. Si ra-Ram´ ırez  AbstractIn thi s paper , the noi se lt eri ng pr obl em (“de- noising” problem) is approached via a suitable modication of the traditional Luenberger observer approach, also known as the “high gain observer” approach (HGO). The HGO observer is, fundamental ly , a Luenb erg er obse rve r with larg e stabl e eigenvalues of the output estimation error dynamics. HGO and some of its modications have been particularly useful in the linear based control of perturbed linear and nonlinear systems. HGO are, however, inappropriate to deal with noisy injection signa ls and noisy plants. To overc ome this fact, the structur e of the Luenbe rger state estimator re con struct ing the time derivatives of the given signal is enhanced against the effects of noise by means of an algebraic ltering scheme. The algebraic lt er ing consi sts in a sui table mod ic ati on of the re ce ntl y introduced algebraic parameter identication methodology. The proposed approach is capable of attenuating the noise effects signicantly. The proposed strategy is illustrated via numerical simulations. I. INTRODUCTION High gain observers constitute an effective signal deriva- tiv e estimation technique in a varie ty of applic ation elds such as robotics [1], motion control [2], position/force con- trol, motor load estimation, induction motor control [3], [4], disturbance and fault estimation [5] among many others. This class of observers offers certain degree of design freedom which can be utilized to achieve properties such as robustness aga ins t external per tur bat ions and other unc ert aintie s. In addi ti on, when the observer is desi gne d, the numbe r of  observed time derivatives can be chosen in connection with the observer dimension. T wo prob lems, arisin g in connection wit h these HGO obs ervers, are : the ini tial “pe aki ng” phe nomenon and the undesired noise amplicat ion. In rel ati on to the pea king phenomenon, there several approaches have been proposed to solve it. See, for instance, [6], where the high-gain differ- entiators are used in combination with a traditional ltering scheme for estimating the higher order time derivatives of a signal. On the othe r hand, high-ga in observ ers have been im- plemented to estimate output time derivatives and unknown plant inputs, such as faults and plant perturbations. In appli- cations where environmental disturbances and measurement J. Cort´ es-Romero is with Unive rsida d Nacional de Colombia. Faculta d de Ingenier´ ıa, Departamento de Ingenier´ ıa El´ ectrica y Electr´ onica. Carrera 30 No. 45-03 Bogot´ a Colombia. [email protected] J. Cort´ es-Romero, C. Garc´ ıa-Rodr´ ıguez, A. Luviano-Ju´ arez, R. Portillo- elez and H. Sir a-Ram´ ıre z are wit h Centro de Inve sti gac on y de Est udi os Ava nza dos del Ins tit uto Pol it´ ecni co Naci onal . Depar tame nto de Ingeni er ´ ıa El´ ect ric a, Sec ci´ on de Me catr´ oni ca. Apa rta do pos tal 14740, ex ic o DF C. P . 07360.  [email protected], [email protected], [email protected], rportillo@cinvestav.mx and hsira@cinvestav.mx noi ses play an import ant r ˆ ole , the tuning of the obser ver gai n bec ome s an impo rta nt and dif cult iss ue. The unde - sired amplicat ion tends to gro w up in the cal cul ati on of each consecuti ve higher order time deri vati ve. A poss ible solution to this problem is using the well known Extended Ka lman Fi lt er , under the assumpti on tha t the nois e is a Gaussian process with known statistics (see [7], [8]). In [9] an alt ernati ve is prop ose d for reduci ng the noi se ef fec ts. A class of “proportional integral observers” are considered which consist of the use of an integral gain in the observer inj ect ion str uct ure to att enuate the ef fec ts of the nois e in the reconstruction error dynamics. This ltering scheme may ach ie ve an ef fec tive reduction of the noise eff ects. Thus, the problem relies on applying adequate ltering techniques to the signal and its time der iv ati ve s. If cla ss ica l lt eri ng schemes are selected, then, undesired signal time delays often arise. On the other hand, numerical algorithms are slow and they are also known to be computationally complex. Algebr aic tec hniq ues ha ve bee n prop ose d to sol ve the proble m of par ame ter es timati on, for linear sys tems wit h constant parameters, and, also, for linear systems with time va ryi ng par ame ter s (se e [10], [11]). The re also exi sts an algebraic approach to carry out the state estimation task in a class of algebraically observable nonlinear systems [12]. The se approac hes are bas ed on the , so cal led, alg ebra ic derivative approach. In this technique, the problem of state esti mation is close ly related to the problem of numeric al dif- ferentiation which is an active area of research in numerous sc ience and engine eri ng are as suc h as control the ory and signal processing. Some numerical differentiation algorithms as linear differentiation, backward nite differences and the Sav itzky -Golay different iation have been analy zed in [13]. T ayl or expans ion approac hes ha ve bee n implement ed to obt ain numeri cal dif fere nti ati on from a noi sy sig nal , see [14], [15]. However, this method becomes ill-conditioned for higher order truncation of the Taylor series expansion. On the other hand, even if the time derivative estimations, obtained from a high -ga in obs erver , are ver y noi sy; the ir avera ge sig nal s are “good” in the sense that the y tend to t the actual underlying signal derivative. This fact can be exploi ted by mea ns of algebr aic manipulat ions similar to those applied in the linear parameter identication problem. The aim of our approach, is to provide a ltering scheme for a noisy measured or estimated derivative signal. In particular, we tak e the est ima te exp res sions from a HGO obs erver [3] and apply an algebraic ltering scheme. This approach is ins pir ed on the algebr aic sta te ide nti cation tec hniq ue prese nted by Fliess and Sira-Ram´ ırez for contin uous time systems (see [10]).

Upload: gelfanduss

Post on 06-Oct-2015

13 views

Category:

Documents


0 download

DESCRIPTION

Artículo

TRANSCRIPT

  • An algebraic denoising scheme

    J. Cortes-Romero, C. Garca-Rodrguez, A. Luviano-Juarez, R. Portillo-Velez and H. Sira-Ramrez

    Abstract In this paper, the noise filtering problem (de-noising problem) is approached via a suitable modification ofthe traditional Luenberger observer approach, also known asthe high gain observer approach (HGO). The HGO observeris, fundamentally, a Luenberger observer with large stableeigenvalues of the output estimation error dynamics. HGO andsome of its modifications have been particularly useful in thelinear based control of perturbed linear and nonlinear systems.HGO are, however, inappropriate to deal with noisy injectionsignals and noisy plants. To overcome this fact, the structureof the Luenberger state estimator reconstructing the timederivatives of the given signal is enhanced against the effects ofnoise by means of an algebraic filtering scheme. The algebraicfiltering consists in a suitable modification of the recentlyintroduced algebraic parameter identification methodology. Theproposed approach is capable of attenuating the noise effectssignificantly. The proposed strategy is illustrated via numericalsimulations.

    I. INTRODUCTION

    High gain observers constitute an effective signal deriva-tive estimation technique in a variety of application fieldssuch as robotics [1], motion control [2], position/force con-trol, motor load estimation, induction motor control [3], [4],disturbance and fault estimation [5] among many others. Thisclass of observers offers certain degree of design freedomwhich can be utilized to achieve properties such as robustnessagainst external perturbations and other uncertainties. Inaddition, when the observer is designed, the number ofobserved time derivatives can be chosen in connection withthe observer dimension.

    Two problems, arising in connection with these HGOobservers, are: the initial peaking phenomenon and theundesired noise amplification. In relation to the peakingphenomenon, there several approaches have been proposedto solve it. See, for instance, [6], where the high-gain differ-entiators are used in combination with a traditional filteringscheme for estimating the higher order time derivatives of asignal.

    On the other hand, high-gain observers have been im-plemented to estimate output time derivatives and unknownplant inputs, such as faults and plant perturbations. In appli-cations where environmental disturbances and measurement

    J. Cortes-Romero is with Universidad Nacional de Colombia. Facultadde Ingeniera, Departamento de Ingeniera Electrica y Electronica. Carrera30 No. 45-03 Bogota Colombia. [email protected]

    J. Cortes-Romero, C. Garca-Rodrguez, A. Luviano-Juarez, R. Portillo-Velez and H. Sira-Ramrez are with Centro de Investigacion y deEstudios Avanzados del Instituto Politecnico Nacional. Departamentode Ingeniera Electrica, Seccion de Mecatronica. Apartado postal14740, Mexico DF C.P. 07360. [email protected],[email protected], [email protected],[email protected] and [email protected]

    noises play an important role, the tuning of the observergain becomes an important and difficult issue. The unde-sired amplification tends to grow up in the calculation ofeach consecutive higher order time derivative. A possiblesolution to this problem is using the well known ExtendedKalman Filter, under the assumption that the noise is aGaussian process with known statistics (see [7], [8]). In [9]an alternative is proposed for reducing the noise effects.A class of proportional integral observers are consideredwhich consist of the use of an integral gain in the observerinjection structure to attenuate the effects of the noise inthe reconstruction error dynamics. This filtering scheme mayachieve an effective reduction of the noise effects. Thus,the problem relies on applying adequate filtering techniquesto the signal and its time derivatives. If classical filteringschemes are selected, then, undesired signal time delays oftenarise. On the other hand, numerical algorithms are slow andthey are also known to be computationally complex.

    Algebraic techniques have been proposed to solve theproblem of parameter estimation, for linear systems withconstant parameters, and, also, for linear systems with timevarying parameters (see [10], [11]). There also exists analgebraic approach to carry out the state estimation task ina class of algebraically observable nonlinear systems [12].These approaches are based on the, so called, algebraicderivative approach. In this technique, the problem of stateestimation is closely related to the problem of numerical dif-ferentiation which is an active area of research in numerousscience and engineering areas such as control theory andsignal processing. Some numerical differentiation algorithmsas linear differentiation, backward finite differences and theSavitzky-Golay differentiation have been analyzed in [13].Taylor expansion approaches have been implemented toobtain numerical differentiation from a noisy signal, see[14], [15]. However, this method becomes ill-conditioned forhigher order truncation of the Taylor series expansion.

    On the other hand, even if the time derivative estimations,obtained from a high-gain observer, are very noisy; theiraverage signals are good in the sense that they tend tofit the actual underlying signal derivative. This fact can beexploited by means of algebraic manipulations similar tothose applied in the linear parameter identification problem.The aim of our approach, is to provide a filtering scheme fora noisy measured or estimated derivative signal. In particular,we take the estimate expressions from a HGO observer[3] and apply an algebraic filtering scheme. This approachis inspired on the algebraic state identification techniquepresented by Fliess and Sira-Ramrez for continuous timesystems (see [10]).

  • The remainder of the article is organized as follows:In Section II, the problem formulation is presented; then,an example of the filtering methodology is considered inSection III. To illustrate the methodology, in Section IV somenumerical results are given for two different levels of additivenoise and a comparison with other filtering alternatives isdepicted. Finally, the conclusions and suggestions for furtherwork are comprised in Section V.

    II. PROBLEM FORMULATIONConsider a smooth signal x(t), affected by an additive,

    zero mean high frequency noise, (t)

    y(t) = x(t) + (t) (1)Carrying out any signal processing, or implementing a

    control law through this noisy measurement is undesirablebecause of the noise effects downstream. There exist lowpass filtering alternatives, but they need a certain knowledgeabout the noise characteristics while producing a significantdelay of the filtered signal.

    Problem Formulation. Given a noisy measurement, y(t),of the smooth signal x(t) as in (1), with, (t), being anadditive, zero mean, high frequency measurement noise;devise an online filtering process for the measured signal,y(t), such that it recovers the signal x(t), with no delay.

    A. An observer-based filtering algorithm: An algebraic ap-proach

    The proposed algorithm is carried out in two steps.The first step consists in devising an asymptotic estimationscheme for the time derivatives of, y(t), via a traditionalhigh gain observer as if y were uncorrupted by noise.Since the additive signal, (t), is a zero mean noise, weknow that the observer based estimated derivatives are alsocrucially affected by the noise, but, their average valuesare asymptotically coincident with the actual derivatives ofx(t). The second step consists of considering the algebraicexpression relating the j th time derivative of y(t) and itsestimated value from the observer equations. This expressionreceives an algebraic treatment to extract the required filteredsignal. One generally proceeds by multiplying the appropri-ate expression by the j-th positive power of the time variable,(t)j , then, the resulting product is integrated by parts jtimes thus obtaining an expression which is free of the initialconditions and of any time derivatives of y(t). From theobtained expression, the signal y(t) can be expressed as thequotient of two time functions. This filtering can be carriedout quite fast and with no involved delays, as it is commonin many of the traditional low pass filtering schemes.

    1) Obtaining an average derivative: Consider a Taylorseries expansion of the noise-free signal x(t), which isassumed to be infinitely differentiable in a neighborhood oft0

    x(t) = x(t0) +(t t0)

    1!x(t0) +

    (t t0)2

    2!x(t0) + (2)

    For an (n1)th order approximation of x(t), it is possibleto consider the following n th order model

    dnx(t)

    dtn= 0 (3)

    Rewriting (3) in state variable form, we obtainx1 = x2x2 = x3

    .

    .

    .

    .

    .

    .

    xn = 0

    x(t) = x1

    (4)

    A Luenberger observer can be proposed in order to esti-mate the unknown states of (4) as follows

    x1 = x2 + n1x2 = x3 + n2

    .

    .

    .

    .

    .

    .

    xn = 0

    (5)

    where the output estimation error is defined by = y x1.The asymptotic convergence of this estimation error dependson the parameters i, which can be selected according toa pre-specified Hurwitz polynomial governing the resultingoutput estimation error dynamics.

    2) Obtaining an on-line smooth signal: As it is known,under noisy output measurement conditions the output timederivative estimates, obtained from a HGO, are significantlydistorted. Nevertheless, their realizations are close, in the av-erage to their actual values, provided a zero mean assumptionhas been established on the nature of the noise. In this paper,a algebraic filtering process is proposed using the estimatedderivatives supplied by an observer of the high gain type.Abusing the notation1, a j th order algebraic filtering isobtained from

    (j)tjy(j) =

    (j)tj xj+1 for 1 < j < n (6)

    where xj+1 is the jth estimated derivative of y(t) suppliedby (5). Integrating by parts (j)0 tjy(j), it is possible to finda filtered version for y(t).

    In order to illustrate we consider the particular case whenj = 3.

    Integration by parts of the corresponding expression allowsus to check that (3)

    (t)3y(3) = 6

    (3)y 18

    (2)ty + 9

    (1)t2y t3y

    1We denote the quantity t0

    10

    j10

    (j)dj . . . d1

    by the expression: (j)

    (t)

  • and

    y =

    (3)t3y(3) + 6

    (3)y 18

    (2)ty + 9

    (1)t2y

    t3

    given that the third time derivative estimate of y is x4, afiltered version of y is given by

    yFilt3 =

    (3)t3x4 + 6

    (3)y 18

    (2)ty + 9

    (1)t2y

    t3

    Remark 1: The algebraic filtering expressions are in termsof the time variable t and some of finite power termswhich may produce high numerical values as time elapses.To avoid having numerical computing problems, instead ofmultiplying by tj , the expression can be multiplied by anadmissible bounded time function (i.e, a function whichallows to apply the algebraic manipulations to generate avalid filtering expression free of initial conditions) so thatnumerator and denominator signals remain bounded. Forinstance, multiplying by (1 et)j instead of (t)j .

    III. EXAMPLELet y(t) be a signal of the form (1). It is desired to filter the

    noisy output signal by means of the proposed combinationof a Luenberger observer and the algebraic filter approach.Proceeding with the methodology, an approximation to thesignal, x(t), can be modelled as

    d7x

    dt7= 0 (7)

    A. Observer designAn associated state space representation of (7) can be

    readily conformed as

    x1 = x2

    x2 = x3

    x3 = x4

    x4 = x5

    x5 = x6

    x6 = x7

    x7 = 0

    y = x1 +

    A Luenberger observer for the polynomial model of x(t)is then synthesized as follows:

    x1 = x2 + 6(t)

    x2 = x3 + 5(t)

    x3 = x4 + 4(t)

    x4 = x5 + 3(t)

    x5 = x6 + 2(t)

    x6 = x7 + 1(t)

    x7 = 0(t)

    (t) = y x1

    where the corresponding output estimation error dynamics isclearly given by the linear dynamics,

    (7) + 6(6) + 5

    (5) + + 1 + 0 = 0

    whose corresponding characteristic polynomial is

    p(s) = s(7) +6s6 +5s

    5 +4s4 +3s

    3 +2s2 +1s+0

    The gain parameters of the observer are chosen such thatthe characteristic polynomial of the output observation errormatches the known polynomial (s2 + 2ns+ 2n)3(s+ p),with , n, p being positive real values.

    B. Filter designIn this subsection, two different algebraic filters are pre-

    sented. Each algebraic filter arises from a different timederivative estimate expression found from the observer:Namely, a first order time derivative estimate based filterand a fourth order time derivative estimate based filter willbe examined.

    First order time derivative based filter. From the esti-mation of the first order time derivative, x2, we have

    yFilt1 =

    tx2 +

    y

    t(8)

    Fourth order time derivative based filter. From theestimation of the fourth order time derivative, x5, we have

    yFilt4 =n4(t)

    d4(t)(9)

    with n4(t) = (4)

    t5x524 (4)

    y+96 (3)

    ty72 (2)

    t2y+16

    t3y, and d4(t) = t4 respectively.

    IV. NUMERICAL RESULTSNumerical simulations are performed in order to show

    the performance of the proposed approach. A comparisonis also carried out with the performances of two classicfiltering methods: one of them consisting of an eight orderButterworth low pass filter and a second one consisting of aneight order Chebyshev low pass filter, both filtering schemesare provided with a cut-off frequency of 700 [s1]. Thesignal to be taken has the following form

    y(t) =10

    1 + .5 cos(5t2)+ (t) (10)

    with (t) being is a zero mean noise.The observer gain parameters were selected as a function

    of the desired characteristic polynomial, with the desiredparameters: n = 200, p = 200 and = 1. Two differentnoise levels were taken for the simulations, one consistingin a 20 [dB] amplitude and another one of 10 [dB]. Figure1 depicts the analyzed signal. Figures 2 and 3 show thecomparison between the different order algebraic filters forthe 20 [dB] additive noise, and the 10 [dB] additive noiserespectively. The fourth order filtering scheme achieves lessharmonic components in relation to the first order filter. Thecomparison regarding the classic filtering approaches is given

  • in Figures 4 and 5 for the 20 [dB] noise and finally, thesame comparison for the 10 [dB] noise is shown in Figures6 and 7. To highlight the filtering effects the images showa zoom of the filtered response. Notice that classic filteringresponses exhibit a significant delay effect. The delay effectsare absent in the proposed combination of observer plusalgebraic filtering approach.

    0 0.5 1 1.5 2 2.5 30

    5

    10

    15

    20

    25

    Time [s]

    y(t)+(t) y(t)

    Fig. 1. Noisy chirp signal.

    0 0.2 0.4 0.6 0.8 1

    5

    10

    15

    20

    25

    Time [s]

    y(t)+(t) y(t)Filt1

    y(t)Filt4

    y(t)

    Fig. 2. Different order Algebraic Filtering response: SNR=20 dB.

    V. CONCLUDING REMARKS

    In this paper, a novel algebraic filtering technique for noisysignals has been proposed. The filtering scheme is based ona high-gain observer and algebraic manipulations on someof its expression to obtain estimates of the required signals.The algebraic filtering approach has several advantages overclassical filtering methods. First, there is no delay in the

    0 0.2 0.4 0.6 0.8 10

    5

    10

    15

    20

    25

    Time [s]

    y(t)+(t) y(t)Filt1

    y(t)Filt4

    y(t)

    Fig. 3. Different order Algebraic Filtering response: SNR=10 dB.

    0.5 0.6 0.7 0.8 0.9 1 1.15

    10

    15

    20

    Time [s]

    y(t) y(t)Butter y(t)Cheby y(t)Filt4

    Fig. 4. Comparison with classical low pass filters: SNR=20 dB.

    0.5 0.6 0.7 0.8 0.9 1 1.15

    10

    15

    20

    Time [s]

    y(t) y(t)Butter y(t)Cheby y(t)Filt4

    Fig. 5. Comparison with classical low pass filters: SNR=10 dB.

  • 2.75 2.8 2.85 2.9 2.955

    10

    15

    20

    Time [s]

    y(t) y(t)Butter y(t)Cheby y(t)Filt4

    Fig. 6. Comparison with classical low pass filters: SNR=20 dB.

    2.75 2.8 2.85 2.9 2.955

    10

    15

    20

    Time [s]

    y(t) y(t)Butter y(t)Cheby y(t)Filt4

    Fig. 7. Comparison with classical low pass filters: SNR=10 dB.

    filtered signal, which is a major advantage for practical appli-cations. Besides, the algebraic filter does not need any specialfeature, as for classical filter design, like cutoff frequencyor the need for gain tuning. Thus, once the filter has beenbuilt, it can work over a wide range of signal frequencieswithout any redesign. The comparison with some classicalfilters demonstrates the above conclusions. Moreover, thepresented approach can be extended for time derivativesfiltering, which is an still an open problem. In addition, theproposed scheme is easy to implement, then, it is ideal forengineering applications like signal processing and automaticcontrol.

    REFERENCES

    [1] K. Lee and H. Khalil, Adaptive output feedback control of robot ma-nipulators using high-gain observer, International Journal of Control,vol. 67, no. 6, pp. 869886, 1997.

    [2] K. Tan and S. Zhao, Precision motion control with a high gaindisturbance compensator for linear motors, ISA Transactions, vol. 43,pp. 399412, 2004.

    [3] J. Cortes-Romero, A. Luviano-Juarez, and H. Sira-Ramirez, RobustGPI controller for trajectory tracking for induction motors, in Mecha-tronics, 2009. ICM 2009. IEEE International Conference on, April2009, pp. 16.

    [4] J. A. Solsona and M. I. Valla, Disturbance and nonlinear Luenbergerobservers for estimating mechanical variables in permanent magnetsynchronous motors under mechanical parameters uncertainties, IEEETransactions on Industrial Electronics, vol. 50, no. 4, pp. 717725,August 2003.

    [5] R. Martnez-Guerra and S. Diop, Diagnosis of nonlinear systems:An algebraic and differential approach, IEE Control Theory andApplications, vol. 151, pp. 130135, 2004.

    [6] Y. Chitour, Time-varying high-gain observers for numerical differen-tiation, IEEE Transactions on Automatic Control, vol. 47, no. 9, pp.15651569, 2002.

    [7] A. Alessandri, Design of sliding-mode observers and filters fornonlinear dynamic systems, in IEEE 39th Conference on Decisionand Control, vol. 3, no. 11, 2000, pp. 25932598.

    [8] F. Lewis, Optimal Estimation With An Introduction To StochasticControl Theory. John Wiley, 1986.

    [9] K. Busawon and P. Kabore, Disturbance attenuation using propor-tional integral observers, International Journal of Control, vol. 74,no. 6, pp. 618627, 2001.

    [10] M. Fliess and H. Sira Ramrez, An algebraic framework for linearidentification, ESAIM: Control, Optimization and Calculus of Varia-tions, vol. 9, pp. 151168, 2003.

    [11] W. Yang Tian; Floquet, T.; Perruquetti, Fast state estimation in lineartime-varying systems: An algebraic approach, in Proceedings of the47th IEEE Conference on Decision and Control, Cancun Mexico,December 2008, pp. 2539 2544.

    [12] M. Fliess, C. Join, and H. Sira-Ramirez, Non-linear estimation iseasy, International Journal of Modelling, Identification and Control,vol. 4, no. 1, pp. 1227, 2008.

    [13] M. Braci and S. Diop, On numerical differentiation algorithms fornonlinear estimation, in IEEE 42th Conference on Decision andControl, 2003, pp. 15651569.

    [14] M. Mboup, C. Join, and M. Fliess, Numerical differentiation withannihilators in noisy environment, Numerical Algorithms, vol. 50,pp. 439467, 2009, dOI 10.1007/s11075-008-9236-1.

    [15] C. J. M. Mboup and M. Fliess, A revised look at numerical dif-ferentiation with an application to nonlinear feedback control, inMediterranean Conference on Control and Automation, 2007, pp. 626.