Adaptive Filtering and Change Detection Fredrik Gustafsson Copyright © 2000 John Wiley & Sons, Ltd ISBNs: 0-471-49287-6 (Hardback); 0-470-84161-3 (Electronic) Part I: Introduction Adaptive Filtering and Change Detection Fredrik Gustafsson Copyright © 2000 John Wiley & Sons, Ltd ISBNs: 0-471-49287-6 (Hardback); 0-470-84161-3 (Electronic) Extended summary 1.1 About the book 1.1.1 Outlook 1.1.2 Aim 1.1.3 Background knowledge 1.1.4 Outline and reading advice 1.2 Adaptive linear filtering 1.2.1 Signal estimation 1.2.2 Parameterestimation using adaptive filtering 1.2.3 State estimationusingKalman filtering 1.3 Change detection 1.3.1 Filtersasresidualgenerators 1.3.2 Stopping rules 1.3.3 One-model approach 1.3.4 Two-model approach 1.3.5 Multi-model approach 1.4 Evaluation and formal design 1.4.1 General considerations 1.4.2 Performance measures 3 11 13 17 17 18 19 22 23 26 26 28 1.1 Aboutthebook 1.1.1 Outlook The areas of adaptive filtering and change (fault) detection are quite active fields both in research and applications Some central keywords of the book are listed in Table 1.1, and the figures illustrated in Figure 1.1, give an idea of the relative activity in the different areas For comparison the two related and well established areas of adaptive control and system identification are included in the table Such a search gives a quick idea of the size of the areas but there are of course many shortcomings and the comparison may be unfair at several instances Still it is interesting to see that the theory has reached many successful applications whichis directly reflected in the Extended summary Table 1.1 Keywords andnumber of hits (March 2000) in different databases For ScienceDirect the maximum numberof hits is limited to 2000 On some of the rows, the logical ‘or’ is used for related keywords like ‘adaptive signal processing or adaptive estimation or adaptive filter’ Keyword Adaptive filter/estimation/SP Kalman filter Adaptive equalizer (Eq) Target tracking Fault diagnosis (FDI) Adaptive control (AC) (System) Identification (SI) Total number of items I EL IEL 4661 1921 479 890 2413 4563 8894 588683 ScienceDirect Kalman AC ScienceDirect 952 1642 74 124 417 2000 2000 856692 IBM patent 871 317 291 402 74 666 317 2582588 IBM patents Kalman AC Figure l Relative frequency of keywords in different databases number of patents Browsing the titles also indicates that many journal and conference publications concern applications Figure 1.2 reveals the, perhaps well known, fact that thecommunication industry is more keen to hold patents (here: equalization) Algorithms aimed at real-time implementation are also, of course, more often subject to patents, compared to, for instance, system identification, which is a part of the design process Table 1.2 lists a few books in these areas It is not meant to be comprehensive, only to show a few important monographs in the respective areas 1.1.2 Aim The aim of the book is to provide theory, algorithms and applicationsof adaptive filters with or without support from change detection algorithms Applications in these areas can be divided into the the following categories: l About the book Patents per publication SI Adaptive AC Figure 1.2 Relative ratio of number of patents found in the IBM database compared to publications in IEL for different keywords Table 1.2 Related books Keyword Adaptive filters Kalman filter Adaptive equalizer Target tracking Fault diagnosis Adaptive control System identification 0 Books Haykin (1996), Mulgrew and Cowan (1988), Widrow and Stearns (1985), Cowan and Grant (1985) Kailath et al (1998), Minkler and Minkler (1990), Anderson and Moore (1979), Brown and Hwang (1997), Chui and Chen (1987)7 Proakis (1995), Haykin (1994), Gardner (1993), Mulgrew and Cowan (1988) Bar-Shalom and Fortmann (1988), Bar-Shalom and Li (1993), Blackman (1986) Basseville and Nikiforov (1993), Gertler (1998), Chen and Patton (1999), Mangoubi (1998) Wstrom and Wittenmark (1989), Goodwin and Sin (1984) Ljung (1999), Soderstrom and Stoica (1989), Johansson (1993) Surveillance and parametertracking Classical surveillanceproblems consist in filteringnoisy measurements of physical variables asflows, temperatures, pressures etc, which will be called signal estimation Modelbased approaches, where (time-varying) parameters in a model of a nonstationary signal need to be estimated, is a problem of parameter trackis blind ing Adaptive control belongs tothisarea.Anotherexample equalization in digital communication State estimation The Kalman filter provides the best linear state estimate, and change detection support can be used to speed up the re- Extended summary sponse after disturbances and abrupt state changes Feedback control using state feedback, such as Linear Quadratic Gaussian LQG control, belongs to this area Navigation and target tracking are two particular application examples Fault detection Faults can occur in almost all systems Change detection here has the role of locating the fault occurrence intime and togive a quick alarm After the alarm, isolation is often needed to locate the faulty component The combined task of detection and isolation is commonly referred to as diagnosis Fault detection can be recast to one of parameter or state estimation Faults in actuators and sensors are most easily detected in a state space context, while system dynamic changes often require parametric models These problems are usually treated separately in literature in the areas of signal processing,mathematical statistics, automaticcontrol, communicationsystems and quality control However, the tools for solving these problems have much in common, and the same type of algorithms can be used (C.R Johnson, 1995) The close links between these areas are clearly under-estimated in literature The main difference of the problem areas above lies in the evaluation criteria In surveillance the parameter estimate should be as close as possible to the true value, while in fault detection it is essential to get an alarm from the change detector as soon as possible after the fault, and at the same time generating few false alarms In fault detection, isolation of the fault is also a main task The combination of fault detection and isolation is often abbreviated to FDI, and the combined task can be referred to as diagnosis More terminology used in this area is found in Appendix B The design usually consists of the following steps: Modeling the signal or system Implementing an algorithm Tuning the algorithm with respect to certain evaluation criteria, either using real or simulated data The main focus is on algorithms and their properties, implementation, tuning and evaluation Modeling is covered only briefly, but the numerous examples should give an idea of the possibilities of model-based signal processing 1.l Background knowledge The derivations and analysis can be divided intothe following areas, and some prior knowledge, or at least orientation, of these is required: l About the book Statistical theory: maximum likelihood, conditional distributions etc Calculus: integrations, differentiations, equation Matrix algebra: projections, subspaces, matrix factorizations etc Signal modeling: transfer functions and state space models solving etc Classical filter theory: the use of a low-pass filter for signal conditioning, poles and zeros etc Transforms and frequency domain interpretations occur, but are relatively rare To use the methods, it is essential to understand themodel and the statistical approach These are explained in each chapter in a section called ‘Basics’ These sections should provide enough information for understanding and tuning the algorithms A deeper understanding requires the reader to go through the calculus and matrix algebra in the derivations The practitioner who is mainly interested in what kind of problems can be addressed is advised to start with the examples and applications sections 1.l Outline and reading advice There are certain shortcuts to approaching the book, and advice on how to read the book is appropriate Chapter is a summary and overview of the book, while Chapter overviews possible applications and reviews the basic mathematical signal models These first two chapters should serve as an overview of the field, suitable for those who want to know what can be done rather than how it is done Chapters 3, and - the first chapter in each part are the core chapters of the book, where standard approaches to adaptive filtering are detailed These can be used independently of the rest of the material The other chapters start with a section called ‘Basics’, which can also be considered as essential knowledge Part V is a somewhat abstract presentation of filter theory in general, without using explicit signal models It is advisable to check the content at an early stage, but the reader should in no way spend too much time trying to digest all of the details Instead, browse through and return to the details later However, the ideas should be familiar before starting with the other parts The material can be used as follows: ~ 0 Chapters and are suitable for people from within industry who want an orientation in what adaptive filtering is, and what change detection can add to performance An important goal is to understand what kind of practical problems can be solved Chapters 5, and 13 are suitablefor an undergraduatecourse in adaptive filtering Extended summary Table 1.3 Organization of the book chapters Estimation of Approach Adaptive filtering and whiteness based change detection Maximum likelihood based change detection Multiple-model based change detection Algebraic (parity space) change detection Parameter Signal State Chapter Chapter Chapter Chapter Chapter Chapter Chapter Chapter Chapter 10 Chapter 11 Chapters 1, 2, 3, 5, 8, 12, 13 andthe ‘Basics’ sectionsin theother chapters can be included in a graduate course on adaptive filtering with orientation of change detection, while a more thorough course for students specializing in the area would include the whole book This matrix organization is illustrated in Table 1.3 Part I1 on signal estimation has many interesting signalprocessing applications, but it also serves as a primer on the change detection chapters in Parts I11 and IV The approach in Chapter 11 is algebraic rather than statistical, and can be studied separately Appendix A overviews the signal models used in the book, and presents the main notation, while Appendix B summarizes notation used in the literature on fault detection The only way in which the book should not be approached is probably a reading from cover to cover The theory in the last part is important to grasp at an early stage, and so are the basics in change detection Some of the parts on change detection will appear rather repetitive, since the basic ideas are quite similar for signal, parameter and state estimation More specifically, Part I1 can be seen as a special case (or an illustrative first order example) of Part 111 1.2 Adaptive linearfiltering Three conceptually different (although algorithmically similar) cases exist: Signal estimation Parameter estimation in an unknown model State estimation in a known model The following sections will explain the basic ideas of theseproblems, and introduce one central example to each of them that will be used throughout the chapter linear 1.2 AdaDtive 1.2.1 Signalestimation The basic signal estimation problem is to estimate the signal part Ot in the noisy measurement yt in the model An example of an adaptive algorithm is Here At will be referred to as the forgetting fuctor It is a design parameter that affects the tracking speed of the algorithm Aswill become clear from the examples to follow, it is a trade-off between trackingspeed and noise attenuation The archetypical example is to use At = X, when this also has the interpretation of the pole in a first order low-pass filter More generally, any(low-pass)filtercanbe used If it is known that the signal level has undergone an abrupt change, as might be indicated by a change detection algorithm, then there is a possibility to momentarily forget all old information by setting At = once This is an example of decision feedback in an adaptive filter, which will play an important role in change detection An illustrative surveillance problem is given below Example 7.7 Fuelconsumption The following application illustrates the use of change detection for improving signal quality The data consist of measurements of instantaneous fuel consumption available from the electronic injection system ina Volvo 850 GLT used as a test car The raw data are pulse lengths of a binary signal, called t,, which is the control signal from the electronic injection system to the cylinders When t , = 1, fuel is injected with roughly constant flow, so the length of the t , pulses is a measure of fuel consumption The measured signal contains a lot of measurement noise and needs some kind of filtering before being displayed to the driver on the dashboard Intuitively, the actual fuel consumption cannot change arbitrarily fast, and the measured signal must be smoothed by a filter There are two requirements on the filter: 0 Good attenuation of noise is necessary to be able to tune theaccelerator during cruising Goodtracking ability Tests show that fuel consumption very often changes abruptly, especially in city traffic Extended summary 10 25 - Measurement I Y Time - Slow filter [S] Figure 1.3 Measurements of fuel consumption and two candidate filters Data collected by the author in a collaboration with Volvo Theserequirements arecontradictory for standard linear filters The thin lines in Figure 1.3 show measurements of fuel consumption for a test in city traffic The solid lines show the result of (1.2) for two particular values of the forgetting factor X The fast filter follows the abrupt changes well, but attenuates the noise unsatisfactorily, and it is the other way around for the slow filter The best compromise is probably somewhere in between these filters The fundamental trade-off between speed and accuracy is inherent in all linear filters Change detectors provide a tool to design non-linear filters with better performance for the type of abruptly changing signal in Figure 1.3 Figure 1.4 shows the raw data, together witha filter implemented by Volvo (not exactly the samefilter, but the principal functionality is the same) Volvo uses a quite fast low-pass filter to get good tracking ability and then quantizes the result to a multiple of 0.3 to attenuatesome of the noise To avoid a rapidly changing value in the monitor, they update themonitored estimate only once a second However, the quantization introduces a problem when trying to minimize fuel consumption manually, and the response time to changes of one second makes the feedback information to the driver less useful linear 1.2 AdaDtive 11 25 20 - - Measurement n - Volvo's filter 40 20 60 Time 80 101 [S] Figure 1.4 Measured fuel consumption and a filtered signal similar to Volvo's implemented filter 1.2.2 Parameter estimation using adaptive filtering A quite general parametric model of a linear system is Here G(q;0) and H ( q ;0) are two filters expressed in the delay operator q defined by qut = ut+l The parameters I9 in the model are assumed to be time-varying, and are to be estimated recursively Here and in the sequel ut denotes measured inputs to the system, if available, and et is an unknown input supposed to be modeled as a stochastic noise disturbance The generic form of an adaptive algorithm is et,, Et = et + = Yt - KtEt, !?to The outputfrom the estimatedmodel is compared to thesystem, which defines the residual E t The adaptive filter acts as a system inverse, as depicted in Figure 1.5 One common filterthat does this operation for linear in parameter models is the recursive least squares (RLS) filter Other well known filters are Least Mean Square (LMS) (unnormalized or normalized), the Kalman filter and least squares over sliding window This book focuses on parametric models that are linear in the parameters (notnecessarily linear in the measurements) The reason for this is that the statistical approaches become optimal in this case How to obtain sub-optimal algorithms for the general linear filter model will be discussed Extended summary 16 X 104 Initial filter design 104 3.5 3.5 3 2.5 2.5 2 > 1.5 0.5 0 X X -0.5 X 104 104 (4 Figure 1.9 Radar measurements (a) and estimates from a Kalman filter (b) for an aircraft in-flight manoeuvre space model /l xt+T = \o Yt = T 0) OIXttl 0 1) T IWt (1.14) + et The state vector used here is X = ( ~ ~ ~ , 1 , 1where ) ~ ~ xi is the position in 2D, z i l the corresponding velocity The state equation is one example of a motion model describing the dynamics of the object to be tracked More examples are given in Chapter The measurements are transformedto this coordinate system and a Kalman filter is applied The resulting estimates are marked with stars We can note that the tracking is poor, and as will be demonstrated, thefilter can be better tuned To summarize what has been said, conventional adaptive filters have the following well-known shortcoming: Fundamental limitation of linear adaptive filters The adaptation gain in a linear adaptive filter is a compromise between noise attenuation and tracking ability 1.3 Change detection 17 1.3 Changedetection Algorithmically, all proposed change detectorscanbeputinto following three categories: 0 one of the Methods using one filter, where a whiteness test is applied to the residuals from a linear filter Methods using two filters, one slow and one fast one, in parallel Methods using multiple filters in parallel, each one matched to certain assumption on the abrupt changes In the following subsections, these will be briefly described Let us note that the computational complexity of the algorithm is proportional to how many filters are used Before reviewing these methods, we first need to define what is meant by a residual in this context, and we also need a tool for deciding whether a result is significant or not - a stopping rule 1.3.1 Filtersasresidualgenerators A good understanding of the Kalman and adaptivefilters requires a thorough reading of Chapters and However, as a shortcut to understanding statistical change detection, we only need to know the following property, also illustrated in Figure 1.10 Residual generation Under certain model assumptions, the Kalman and adaptivefilters take the measured signals and transform them to a sequence of residuals that resemble white noise before the change occurs From a change detection point of view, it does not matter which filter we use and the modeling phase can be seen as a standard task The filters also computes other statistics that are used by some change detectors, but more on this later Filter Figure 1.10 A whitening filter takes the observed input ut and output yt and transforms them to a sequence of residuals E t 18 Extended summary In aperfect world, the residuals would be zero before a changeand non-zero afterwards Since measurement noise and process disturbances are fundamental problems in the statistical approach to change detection, the actual value of the residuals cannot be predicted Instead, we have to rely on their average behavior If there isno change in the system, and the model is correct, then the residuals are so-called white noise, that is a sequence of independent stochastic variables with zero mean and known variance After the change either the mean or variance or both changes, that is, the residuals become ‘large’ in some sense The main problem in statistical change detection is to decide what ‘large’ is Chapter 11 reviews how state space models can be used for filtering (or residual generation as it will be referred to in this context) The idea is to find a set of residuals that is sensitive to the faults, such that a particular fault will excite different combinations of the residuals The main approach taken in that chapter is based on parity spaces The first step is to stack all variables into vectors The linear signal model can then be expressed as where Yt is a vector of outputs, Ut is a vector of inputs, Dt the disturbances and Ft the faults The residual is then defined as a projection With proper design of W , the residual will react to certain faults in specific patters, making fault isolation possible A simple example is when the measurement is two-dimensional, and the state disturbance and the fault are both scalar functions of time Then, under certain conditions, it is possible to linearly transform the measurement to ~t = (dt, f t ) T A projection that keeps only the second component can now be used as the residual to detect faults, and it is said that the disturbance is decoupled It should also be noted that the residual is not the only indicator of a change (that is, it is not a sufficient statistic) inall cases So even though residual based change detection as outlined below is applicable in many cases, there might be improved algorithms The simplified presentation in this chapter hides the fact that the multi-model approaches below actually use other statistics, but the residual still plays a very important role 1.3.2 Stopping rules Many change detection algorithms, among these algorithms in the classes of one-model and two-model approaches below, can be recast into the problem 1.3 Chanae detection 19 of deciding on the following two hypotheses: H0 : E(st) = 0, H1 :E(st) > A stopping rule is essentially achieved by low-pass filtering st and comparing this value to a threshold Below, two such low-pass filters are given: The Cumulative SUM (CUSUM) test of Page (1954): gt = max(gt-1 + st - v,0), alarm if gt > h The drift parameter U influences the low-pass effect, and the threshold h (and also v ) influences the performance of the detector The Geometric Moving Average (GMA) test in Roberts (1959) gt = Xgt-1 + (1- X)st, alarm if gt > h Here, the forgetting factor X is used to tune thelow-pass effect, and the threshold h is used to tune the performance of the detector Using no forgetting at all (X = 0), corresponds to thresholding directly, which is one option 1.3.3 One-model approach Statistical whiteness tests can be used to test if the residuals are white noise as they should be if there is no change Figure 1.11 shows the basic structure, where the filter residuals are transformed to a distance measure, that measures the deviation from the no-change hypothesis The stopping rule decides whether the deviation is significant or not The most natural distance measures are listed below: Change in the mean The residual itself is used in the stopping rule and St = Et Change invariance The squaredresidual subtracted by a known residual variance X is used and st = E; - X Data Yt,Ut c Filter Et c Distance meas St Stopping rule Alarm c k,ta Figure 1.11 Change detection based on a whiteness test from filter residuals Extended summary 20 0 Change in correlation The correlation between the residual and past outputs and/or inputs are used and st = Etyt-k or st = ~ t u t - k for some k Change in sign correlation For instance, one can use the fact that white residuals should change sign every second sample in theaverage and use st = sign(etet-l) A variant of this sign test is given in Section 4.5.3 Example 7.4 Fuelconsumption To improve on the filter in Example 1.1, the CUSUM test is applied to the residuals of a slow filter, like that in Figure 1.3 For the design parameters h = and v = 0.5, the response in Figure 1.12 is obtained The vertical lines illustrate the alarmtimes of the CUSUM algorithm The lower plot shows how the test statistic exceeds the threshold level h at each alarm The adaptive filter in this example computesthe mean of the signal fromthe latest alarm to the current time With a bad tuning of the CUSUM algorithm, we get either the total mean of the signal if there are no alarms at all, or we get the signal back as the estimate if the CUSUM test gives an alarm at each time instant These are the two extreme points in the design Note that nothing worse can happen, so the stability of the filter is not an issue here To avoid the first situation where the estimate will converge to theoverall signal mean, a better 30 E g 20 z I 10 U 0 100 200 300 400 500 ~~ Time [samples] Figure 1.12 Response of an adaptive filter restarted each time a CUSUM test ( h = , v = 0.5), fed with the filter residuals, gives an alarm The lower plot shows the test statistic of the CUSUM test 1.3 Chanae detection 21 design is to use a slow adaptive filter of the type illustrated in Figure 1.3 To avoid the second degenerate case, an alarm can trigger a fast filter instead of a complete filter restart That is, the algorithm alternates between slow and fast forgetting instead of complete or no forgetting In contrast to the example above, the next example shows a case where the user gets important information from the change detector itself Example 1.5 frictionestimation Figure 1.13 shows how a whiteness test, used for restarting the filter, can improve the filtering result in Example 1.2 quite considerably Here the CUSUM test statistics from the residuals is used, and the test statistics gt are shown in the lower plot in Figure 1.13 Note that the test statistics start to grow at time 200, but that the conservative threshold level of is not reached until time 210 0.04- - Estimated parameters True parameters I I I v-100 200 150 250 ' 4I 41 I "0 50 100 150 200 300 250 300 Time [samples] Figure 1.l Estimated friction parameters (upper plot) and test statistics from the CUSUM test (lower plot) Note the improved tracking ability compared to Figure 1.7 In the following example, change detection is a tool for improved tracking, and the changes themselves not contain much information for the user Example 1.6 Target tracking This example is a continuation of Example 1.3, and the application of the CUSUM test is analogous to Example 1.5 Figure 1.14(a) shows the estimated ... like ? ?adaptive signal processing or adaptive estimation or adaptive filter’ Keyword Adaptive filter/estimation/SP Kalman filter Adaptive equalizer (Eq) Target tracking Fault diagnosis (FDI) Adaptive. .. been said, conventional adaptive filters have the following well-known shortcoming: Fundamental limitation of linear adaptive filters The adaptation gain in a linear adaptive filter is a compromise... publication SI Adaptive AC Figure 1.2 Relative ratio of number of patents found in the IBM database compared to publications in IEL for different keywords Table 1.2 Related books Keyword Adaptive filters