Journal of Control Science and Engineering
Robustness Issues in Fault Diagnosis and Fault Tolerant Control
Guest Editors: Jakob Stoustrup and Kemin Zhou Robustness Issues in Fault Diagnosis and Fault Tolerant Control Journal of Control Science and Engineering Robustness Issues in Fault Diagnosis and Fault Tolerant Control
Guest Editors: Jakob Stoustrup and Kemin Zhou Copyright © 2008 Hindawi Publishing Corporation. All rights reserved.
This is a special issue published in volume 2008 of “Journal of Control Science and Engineering.” All articles are open access articles distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. Editor-in-Chief Jie Chen, University of California, USA
Associate Editors
Evgeny I. Veremey, Russia Shinji Hara, Japan Joe Qin, USA V. Balakrishnan, USA Jie Huang, China R. Salvador Sanchez´ Pena,˜ Spain Amit Bhaya, Brazil Pablo Iglesias, USA Sigurd Skogestad, Norway Benoit Boulet, Canada Vladimir Kharitonov, Mexico Jakob Stoustrup, Denmark Tongwen Chen, Canada Derong Liu, USA Jing Sun, USA Guanrong Chen, Hong Kong Tomas McKelvey, Sweden Zoltan Szabo, Hungary Ben M. Chen, Singapore Kevin Moore, USA Onur Toker, Turkey EdwinK.P.Chong,USA Silviu Niculescu, France XiaohuaXia,SouthAfrica Nicola Elia, USA Brett Martin Ninness, Australia George G. Yin, USA Andrea Garulli, Italy Yoshito Ohta, Japan Li Yu, China Juan Gomez, Argentina Hitay Ozbay, Turkey Lei Guo, China Ron J. Patton, UK Contents
Robustness Issues in Fault Diagnosis and Fault Tolerant Control, Jakob Stoustrup and Kemin Zhou Volume 2008, Article ID 251973, 2 pages
On Fault Detection in Linear Discrete-Time, Periodic, and Sampled-Data Systems, P. Zhang and S. X. Ding Volume 2008, Article ID 849546, 18 pages
Optimal Robust Fault Detection for Linear Discrete Time Systems, Nike Liu and Kemin Zhou Volume 2008, Article ID 829459, 16 pages
Computation of a Reference Model for Robust Fault Detection and Isolation Residual Generation, Emmanuel Mazars, Imad M. Jaimoukha, and Zhenhai Li Volume 2008, Article ID 790893, 12 pages
Fault Detection, Isolation, and Accommodation for LTI Systems Based on GIMC Structure, D. U. Campos-Delgado, E. Palacios, and D. R. Espinoza-Trejo Volume 2008, Article ID 853275, 15 pages
A Method for Designing Fault Diagnosis Filters for LPV Polytopic Systems,SylvainGrenaille, David Henry, and Ali Zolghadri Volume 2008, Article ID 231697, 11 pages
Design and Assessment of a Multiple Sensor Fault Tolerant Robust Control System, S. S. Yang and J. Chen Volume 2008, Article ID 152030, 10 pages
Design and Analysis of Robust Fault Diagnosis Schemes for a Simulated Aircraft Model, M. Benini, M. Bonfe,` P. Castaldi, W. Geri, and S. Simani Volume 2008, Article ID 274313, 18 pages
Stability of Discrete Systems Controlled in the Presence of Intermittent Sensor Faults, Rui Vilela Dion´ısio and Joao˜ M. Lemos Volume 2008, Article ID 523749, 11 pages
Actuator Fault Diagnosis with Robustness to Sensor Distortion, Qinghua Zhang Volume 2008, Article ID 723292, 7 pages
Stability Guaranteed Active Fault-Tolerant Control of Networked Control Systems, Shanbin Li, Dominique Sauter, Christophe Aubrun, and Joseph Yame´ Volume 2008, Article ID 189064, 9 pages
Reliability Monitoring of Fault Tolerant Control Systems with Demonstration on an Aircraft Model, Hongbin Li, Qing Zhao, and Zhenyu Yang Volume 2008, Article ID 265189, 10 pages
Fault-Tolerant Control of a Distributed Database System, N. Eva Wu, Matthew C. Ruschmann, and Mark H. Linderman Volume 2008, Article ID 310652, 13 pages Hindawi Publishing Corporation Journal of Control Science and Engineering Volume 2008, Article ID 251973, 2 pages doi:10.1155/2008/251973
Editorial Robustness Issues in Fault Diagnosis and Fault Tolerant Control
Jakob Stoustrup1 and Kemin Zhou2
1 Section for Automation and Control, Department of Electronic Systems, Aalborg University, Fredrik Bajers Vej 7C, 9220 Aalborg, Denmark 2 Department of Electrical and Computer Engineering, Louisiana State University, Baton Rouge, LA 70803, USA
Correspondence should be addressed to Jakob Stoustrup, [email protected]
Received 24 January 2008; Accepted 24 January 2008
Copyright © 2008 J. Stoustrup and K. Zhou. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Fault diagnosis and fault tolerant control have become criti- The robust fault detection and isolation problem is stud- cally important in modern complex systems such as aircrafts ied in the paper by E. Mazars et al., where an H criterion is and petrochemical plants. Since no system in the real world used, giving rise to a quadratic matrix inequality problem. A can work perfectly at all time under all conditions, it is crucial jet engine example is provided. to be able to detect and identify the possible faults in the sys- D. Campos-Delgado et al. suggest an active fault-tolerant tem as early as possible so that measures can be taken to pre- control, and a design strategy is provided, which takes model vent significant performance degradation or damages to the uncertainty into account. The methods are illustrated for a system. Fault diagnosis is not relevant only for safety critical DC motor example. systems, but also for a significant number of systems, where Systems that can be described by linear parameter vary- availability is a major issue. ing models are considered by S. Grenaille et al. where robust- In the past twenty some years, fault diagnosis of dynamic ness constraints are included in the design of fault detection systems has received much attention and significant progress and isolation filters. An illustration of the methods is given has been made in searching for model-based diagnosis tech- in terms of an application to a nuclear power plant. niques. Many techniques have been developed for fault detec- A method for design of a diagnosis and a fault tolerant tion and fault tolerant control. However, the issue of robust- control system using an integrated approach is presented in ness of fault detection and fault tolerant control has not been the paper by S. Yang and J. Chen. The design is illustrated for sufficiently addressed. Since disturbances, noise, and model a double inverted pendulum system. uncertainties are unavoidable for any practical system, it is essential in the design of any fault diagnosis/fault tolerant M. Benini et al. present both a linear and a nonlinear fault control system to take these effects into consideration, so that detection and isolation scheme. This paper focuses on robust fault diagnosis/tolerant control can be done reliably and ro- fault diagnosis for an aircraft model, and has extensive simu- bustly. The objective of this special issue is to report some lations. most recent developments and contributions in this direc- The fault tolerant scheme proposed by R. Dionisio and tion. J. Lemos is capable of stabilizing systems with intermittent The special issue is initiated by a paper by P. Zhang and sensor faults. The approach is based on reconstructing the S. Ding, which gives a review of standard fault detection for- feedback signal, using a switching strategy where a model is mulations, focusing on robustness issues for model-based di- used in the intermittent periods. agnosis systems. A design method for actuator fault diagnosis is proposed N. Liu and K. Zhou then study a number of robust fault in the paper by Q. Zhang, where the focus is to obtain robust- detection problems, such as H/H, H2/H,andH/H prob- ness with respect to nonlinear sensor distortion. A numerical lems, and it is shown that these problems share the same op- exampleisgiven. timal filters. The optimal filters are designed by solving an The problem of designing fault tolerant control systems algebraic Riccati equation. for networked systems with actuator faults is treated in the 2 Journal of Control Science and Engineering paper by Li et al. The proposed design method is demon- strated by a numerical example. The notion of a reliability index is introduced by H. Li et al., for monitoring fault tolerant control systems. The re- liability is evaluated based on semi-Markov models, and the approach is applied to an aircraft model. The final paper of the special issue by N. Wu et al. ad- dresses fault tolerant control of a distributed database system. The fault tolerant design relies on data redundancy in the partitioned system architecture. Robustness is represented by the introduction of additional states modeling delays and de- cision errors. The design is based on solving Markovian de- cision problems. Jakob Stoustrup Kemin Zhou Hindawi Publishing Corporation Journal of Control Science and Engineering Volume 2008, Article ID 849546, 18 pages doi:10.1155/2008/849546
Review Article On Fault Detection in Linear Discrete-Time, Periodic, and Sampled-Data Systems
P.Zhang and S. X. Ding
Institute for Automatic Control and Complex Systems, Faculty of Engineering, University of Duisburg-Essen, Bismarckstrasse 81 BB, 47057 Duisburg, Germany
Correspondence should be addressed to S. X. Ding, [email protected]
Received 30 April 2007; Accepted 4 October 2007
Recommended by Kemin Zhou
This paper gives a review of some standard fault-detection (FD) problem formulations in discrete linear time-invariant systems and the available solutions. Based on it, recent development of FD in periodic systems and sampled-data systems is reviewed and presented. The focus in this paper is on the robustness and sensitivity issues in designing model-based FD systems.
Copyright © 2008 P. Zhang and S. X. Ding. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
1. INTRODUCTION thus the central problem in model-based FDI. This is the first difference of FDI problems from control and standard filter- With the increasing requirements of modern complex con- ing problems, where the focus is put on disturbance atten- trol systems on safety and reliability, model-based fault de- uation. Bearing this in mind, full-decoupling problem and tection and isolation (FDI) technology has attracted remark- optimal design of FDI systems have been studied [3–6]and able attention during the last three decades [1–6]. In major different types of indices have been introduced to describe industrial sectors, it has become an important supporting the sensitivity to the faults. Secondly, for the purpose of FDI, technology and is replacing the traditional hardware redun- a fault indicating signal, called residual, needs not only to be dancy technique in part or totally. As a standard functional generated, but also to be evaluated and, based on it, a deci- module, FDI systems are increasingly integrated in mod- sion for the existence, location, and size of the faults needs ern technical systems and provide valuable information for to be made. Therefore, an FDI procedure includes residual condition-based predictive maintenance, higher-level fault generation and residual evaluation. An integrated design of tolerant control, and plant-wide production optimization. these two parts is needed to guarantee the optimal FDI per- Though closely related to the development of control formance [7]. and filtering theory, there are several distinct features of the In this paper, we will first give a review of some stan- model-based FDI problems that justify the efforts made in dard fault detection (FD) problem formulations in discrete- this field. To evaluate the performance of an FDI system in time systems and the available solutions. There are two types practice, miss alarm rate, false alarm rate, and detection de- of discrete-time model-based FD systems: the parity space lay are the most important criteria that decide the accep- and the observer-based ones. The former is, in its original tance of the methods. It is widely accepted that these func- form, specially dedicated to the discrete-time systems [8], tional requirements can be reformulated as a multi-objective while the latter is analogous to the continuous-time sys- problem. Enhancing the robustness of the FDI system to un- tems and its development shares the same essentials with the known disturbances and modeling errors is an essential ob- continuous-time systems. Perhaps for this reason, besides the jective. However, alone the robustness does not guarantee a early research activity on the parity space approaches, only good FDI performance. The sensitivity of the FDI system to fewstudieshavebeenspecificallydevotedtotheFDprob- faults should be simultaneously improved. To find the best lems in discrete-time systems. Recently, the intensive research compromise between the robustness and the sensitivity is on networked control systems (NCS) and embedded systems 2 Journal of Control Science and Engineering considerably stimulates the study on periodic, sampled-data (i) parity space approaches, which are specific for discrete systems [9]. The integration of data communication net- LTI systems and will be dealt to some details; works into control systems introduces natural periodic be- (ii) the parametrization of observer-based FD systems and havior in the system dynamics and the sampling effect is un- postfilter design schemes; derstood not only in view of the behavior of A/D and D/A (iii) fault detection filter schemes, which are mostly studied converters but also in the context of data transmission among and closely related to robust control theory. the subsystems. It can be observed that the recent studies on Thanks to the well-known relationships between the tech- FD in periodic and sampled-data systems are mainly based nical features of the discrete- and continuous-time systems, on the discrete-time model-based FD methods. It is this fact many well-established FD schemes for continuous LTI sys- that motivates us to give an overview of some standard FD tems can be directly applied to the latter two FD schemes. For methods for discrete-time systems and, based on it, to re- this reason, we will restrict ourselves to some representative view and present some recent results on FD in periodic and methods and give a brief view of the analog application of sampled-data systems. Bearing in mind that fault isolation the methods for continuous-time systems to the discrete FD problems can be principally reformulated as a robust fault systems. Another focus in this section is on the comparison detection problem [4, 5], our focus in this paper is on the and interpretation of the FD methods. robustness issues in designing model-based FD systems. The paper is organized as follows. In Section 2,wereview FD methods for discrete-time systems and address some im- 2.1. System models and problem formulation portant relations between different methods. Section 3 is de- Suppose that the discrete LTI systems are described by voted to FD in discrete-time periodic systems. In Section 4, FD in sampled-data systems is addressed. ⎧ ⎨x(k +1)= Ax(k)+Bu(k)+E d(k)+E f (k), Throughout this paper, standard notations of robust con- d f ΣLTI : ⎩ trol theory, for instance those used in [10], are adopted. We y(k) = Cx(k)+Du(k)+Fdd(k)+F f f (k), will use σ(X), σ(X) to denote the minimum and maximum (1) singular values of matrix X,respectivelyandσi(X)tode- n p note any singular value of X that satisfies σ(X) ≤ σi(X) ≤ where x ∈ R is the state vector, u ∈ R the vector of con- m nd σ(X). x E denotes the Euclidean norm of vector x, ξ 2 trol inputs, y ∈ R the vector of process outputs, d ∈ R n the l2-norm of discrete-time signal ξ or the L2-norm of the vector of unknown disturbances, and f ∈ R f the vec- continuous-time signal ξ, ξ 2,[a,b) the 2 norm of ξ over the tor of faults to be detected, A, B, Ed, E f , C, D, Fd,andF f are interval [a, b), and G∞ the H∞-norm of transfer function known constant matrices of appropriate dimensions. In the matrix G. The superscript T denotes the transpose of matri- frequency domain, system ΣLTI can be equivalently described ces and the superscript ∗ denotes the adjoint of operators. by RH∞ stands for the subspace that consists of all proper and real rational stable transfer function matrices. In this paper, y(z) = Gu(z)u(z)+Gd(z)d(z)+G f (z) f (z), (2) we call a state-space model (A, B, C, D) regular, if it is de- tectable and has no invariant zeros on the unit circle and no where Gu(z), Gd(z), and G f (z) denote, respectively, the unobservable modes at the origin. transfer function matrices from u, d,and f to y. Although the design of a model-based FD systems mainly consists of three tasks: (a) residual generation, (b) residual 2. FD OF DISCRETE LTI SYSTEMS evaluation, (c) threshold determination, major research at- tention has been focused on the residual generation with the Linear time-invariant (LTI) systems are the simplest class of following issues. systems. Although the handling of FD problems in discrete LTI systems can often be done along the well-established (i) Full decoupling problem, which deals with the design of framework of FD schemes for continuos LTI systems, study a residual generator, so that the residual signal r satis- on FD in discrete LTI systems is of primary importance from fies the following three aspects: ∀u, d,limr(k) = 0iff = 0, k→∞ (i) it gives insight and often motivates extensions to more (3) r(k) =0iff (k) =0, i = 1, ..., n . complex systems like periodic and sampled-data sys- i f tems addressed in the subsequent sections; If a full decoupling is realized, then the residual evalu- (ii) there are some methods that have been developed spe- ation reduces to detect the nonzeroness of the residual cially for discrete LTI systems; signal. (iii) due to its practical form for the direct online imple- (ii) Optimal FD problem, which is to design the residual mentation, the discrete-time system form is often fa- generator so that the residual signal r is as small as pos- vored in the applications. sible if f = 0anddeviatesfrom0asmuchaspossible if f (k) =0, i = 1, ..., n . In this section, basic ideas and solution procedures of ad- i f vanced FD methods for discrete LTI systems, divided into Considering that in the fault-free case the residual signal r three groups, will be reviewed: would, due to the existence of d,differ from zero, evaluation P. Zhang and S. X. Ding 3 of the size of r is necessary in order to distinguish the influ- coupling is not achievable or not desired, the FD problem is ence of the faults from that of the disturbances. In this paper, often formulated as to solve the optimization problem the norm-based evaluation of the residual signal, denoted by T T = sup = = r (k)r(k)/f fk,s J r and, based on it, threshold determination satisfying dk,s 0, fk,s 0 k,s max JPS(v ) = max = s = T T vs,vsHo,s 0 vs,vsHo,s 0 sup f =0,d =0 r (k)r(k)/dk,sdk,s = k,s k,s Jth sup r (4) T T f =0,d vsH f ,sH , vs = max f s = T T vs,vsHo,s 0 vsHd,sHd,svs will be briefly reviewed. (10)
2.2. Parity space approach whose solution can be obtained by solving a generalized eigenvalue-eigenvector problem [11]. The parity space approach is based on the so-called parity relation. Let s be an integer denoting the length of a moving Solution to optimization problem (10) time window. The output of system (1) over the moving win- dow [k − s, k] can be expressed by the initial state x(k − s), Let Nbasis denote the basis of the left null space of Ho,s.As- the stacked control input vector uk,s, the stacked disturbance sume that λmax and ps,max are the maximal generalized eigen- vector dk,s, and the stacked fault vector fk,s as value and the corresponding eigenvector to the generalized eigenvalue-eigenvector problem yk,s = Ho,sx(k − s)+Hu,suk,s + Hd,sdk,s + H f ,s fk,s,(5) T T − T T = ps,max NbasisH f ,sH f ,sNbasis λmax NbasisHd,sHd,sNbasis 0, where (11) ⎡ ⎤ ξ(k − s) then optimization problem (10)issolvedby ⎢ ⎥ ⎢ξ(k − s +1)⎥ = ⎢ ⎥ = (12) ξk,s ⎢ . ⎥ with ξ standing for y, u, d, f , vs ps,max Nbasis. ⎣ . ⎦ ξ(k) It is pointed out in [12] that the solutions of a full decou- ⎡ ⎤ pling or (10) are achieved at the cost of (considerably) re- ⎡ ⎤ DO··· O C ⎢ ⎥ duced fault detectability. This can be immediately seen with ⎢ ⎥ ⎢ . . ⎥ ⎢ CA⎥ ⎢ .. . ⎥ a look at the dynamics of the residual signal = ⎢ ⎥ = ⎢ CB D . ⎥ Ho,s ⎢ . ⎥ , Hu,s ⎢ ⎥ , ⎣ . ⎦ ⎣ . .. .. ⎦ . . . . O r(k) = vs yk,s − Hu,suk,s = vs H f ,s fk,s + Hd,sdk,s (13) s CA CAs−1B ··· CB D which shows that the influence of the fault expressed (6) by vsH f ,s is structurally reduced to a minimum, that is, rank (v H ) = 1. Reference [12] proposed the use of a par- H and H are constructed similarly as H and can be s f ,s d,s f ,s u,s ity matrix V , achieved by replacing , ,respectively,by , and , . s B D Ed Fd E f F f To satisfy the requirement on the residual signal, a residual r(k) = Vs yk,s − Hu,suk,s = Vs H f ,s fk,s + Hd,sdk,s (14) generator can be constructed as instead of a parity vector aiming at enhancing the influence r(k) = vs yk,s − Hu,suk,s ,(7)of the faults on the residual signal. To this end, the following optimization problems are formulated as where a design parameter vs called parity vector is introduced to modulate the residual dynamics and improve the sensitiv- max JPS,∞/∞ Vs Vs,VsHo,s=0 ity of the residual to the faults and the robustness to the dis- sup T ( ) ( ) T = dk,s=0, fk,s =0 r k r k /fk,s fk,s turbances and the initial state. Usually, vsHo,s 0isrequired = max = T T (15) to eliminate the influence of the initial state and the past in- Vs,VsHo,s 0 sup = = r (k)r(k)/d dk,s fk,s 0,dk ,s 0 k,s put signals (before the time instant k − s). 2 σ VsH f ,s If the existence condition = max , = 2 Vs,VsHo,s 0 σ VsHd,s rank Ho,s Hd,s H f ,s > rank Ho,s Hd,s (8) max JPS,−/∞ Vs V ,V H =0 s s o,s T T inf d =0, f =0 r (k)r(k)/f fk,s is satisfied, then a full decoupling from both the initial state = max k,s k,s k,s = T T (16) and the disturbances can be achieved by solving Vs,VsHo,s 0 sup = = r (k)r(k)/d dk,s fk,s 0,dk ,s 0 k,s 2 σ VsH f ,s vs Ho,s Hd,s = 0, vsH f ,s =0, (9) = max , = 2 Vs,VsHo,s 0 σ V H s d,s 2 for , that is, lies in the intersection between the left null σ VsH f ,s vs vs max JPS, ∞ V i = i/ s = max . (17) Vs,VsHo,s 0 = 2 space of [Ho,s Hd,s] and the image space of H f ,s.Ifafullde- Vs,VsHo,s 0 σ VsHd,s 4 Journal of Control Science and Engineering
The difference in optimization problems (15)-(16)con- In practice, false alarm rate and miss detection rate are sists in that the former considers the maximal influence of two important technical features for the performance eval- the faults on the residual amplitude, while the latter considers uation of a fault detection system. Below, we will introduce the minimal influence. Optimization problem (17)isagen- these two concepts in the context of the norm-based resid- eralization of (15)-(16) and takes into account the fault sen- ual evaluation (22) and briefly compare the above-presented sitivity in different directions. The achievable optimal perfor- parity space solutions. mance index of optimization problem (15) is the same with Jth setting under a given false alarm rate.Consider(23) that of (10). At the end of this subsection, we will show that and denote the upper bound of dk,s E by δd. In the con- the solution of (17) would lead to maximizing the fault de- text of norm-based evaluation, the objective of Jth setting is tectability in the context of a tradeoff between false alarm rate to ensure that any disturbance whose size is not larger than and fault detectability. the tolerant limit should not cause an alarm. To express the strongest disturbance that is allowed without causing a false Solution to optimization problems (15), (16),and(17) alarm in relation to δd, we define false alarm rate (FAR) as α α A solution to optimization problem (17) that also solves FAR = 1 − ,0≤ α ≤ δd =⇒ 0 ≤ 1 − ≤ 1, (25) (15)-(16) simultaneously is given in [12] as follows, which δd δd is derived based on the observation that for any matrices X1 that is, those disturbances whose size is not larger than α = and X2 of compatible dimensions, (1 − FAR)δd should not cause an alarm. Suppose that the allowable FAR is now given. It is straightforward that the σ X X ≤ σ X σ X , 1 2 1 2 threshold should be set as σ X X ≤ σ X σ X , (18) 1 2 1 2 Jth = ασ VsHd,s = (1 − FAR)δdσ VsHd,s . (26) σi X1X2 ≤ σ X1 σi X2 . Note that in the norm-based residual evaluation, Jth is often Assume that there is the following singular value decompo- set as δ σ(V H ) which leads to a zero FAR but may result sition (SVD): d s d,s in a very conservative Jth setting. T NbasisHd,s = U SOV , (19) To express the miss detection rate (MDR), we introduce the set of detectable faults. Note that a fault can be detected where U and V are unitary matrices, S = diag{σ1, ..., σγ}, if and only if then optimization problems (15), (16), and (17)aresolved ⇐⇒ by r E >Jth Vs Hd,sdk,s + H f ,s fk,s E >Jth. (27)
−1 T Vs = PsS U Nbasis, (20) Hence, the set of detectable faults (SDF) ΩDE(Vs, Jth, d)isde- fined as follows: given Vs and Jth where P is any unitary matrix of compatible dimensions. s = | Note that the solutions to the above optimization prob- ΩDE Vs, Jth, d fk,s (27) is satisfied . (28) lems are not unique. For instance, an alternative optimal so- Given Jth, a parity space matrix V delivers a residual sig- lution for problem (16)is s,opt nal with the lowest MDR if
Vs = PsPNbasis, (21) ∀Vs, ΩDE Vs, Jth, d ⊆ ΩDE Vs,opt, Jth, d , (29) where P is the left inverse of NbasisH f ,s. On the other side, only solution (20)solves(15), (16), and (17) simultaneously. that is, ΩDE(Vs,opt, Jth, d) includes the largest number of de- For this reason, (20) is called unified parity space solution. tectable faults, which is equivalent with the lowest MDR. To detect the faults successfully, the generated residual The subsequent comparison study is done in the con- signal should be further evaluated. For a residual signal gen- text of maximizing SDF (i.e., minimizing MDR) under a given erated by means of the parity space approach, the Euclidean FAR. norm defined by Note that (27) can be, according to (19), rewritten into Vs Hd,sdk,s + H f ,s fk,s J = r(k) = rT (k)r(k) (22) E E > (1 − FAR)δ σ V H d s d,s is a reasonable evaluation function. It follows from (14)and T −1 T ⇐⇒ Q IOV dk,s + QS U NbasisH f ,s fk,s (4) that the corresponding threshold is determined by E (30) − T = = > (1 FAR)δdσ Q IOV Jth sup r(k) E σ VsHd,s max dk,s E. (23) f =0,d −1 T by setting Vs = QS U Nbasis,forsomeQ. Based on the decision logic It turns out that ∀Q =0, (30) holds only if ( ) ≤ th =⇒ fault-free, otherwise faulty, (24) r k E J T −1 T IOV dk,s + S U NbasisH f ,s fk,s > (1 − FAR)δd. E a decision for the occurrence of a fault can be finally made. (31) P. Zhang and S. X. Ding 5
It means that a parity matrix that ensures (31) would pro- (iv) The algebraic form of the parity-space-based FD sys- vide a maximal SDF. Note that the unified parity space solu- tem allows a statistic test and norm-based resid- tion (20)deliversexactly(31). Thus, theunifiedparityspace ual evaluation and threshold determination [19]. It solution maximizes SDF (i.e., minimizes MDR) under a given may well bridge the statistical methods [1] and the FAR. observer-based methods. For comparison, denote the SVD of NbasisH f ,s by (v) In the framework of parity-space-based FD system de- sign, system dynamic features like transmission zeros, = T NbasisH f ,s U f S f O V f . (32) zeros in the right half plane (RHP), and so forth are not taken into account. This may cause trouble at the Then the vector-valued solution (12) to optimization prob- online implementation. Also for this reason, we are lem (10) and the matrix-valued solution (21) to optimization of the opinion that the strategy of parity space de- problem (16) can be, respectively, rewritten into sign, observer-based implementation would be helpful
−1 T to solve this problem. vs = psS U Nbasis,
−1 T 2 T −1 − = ps S U U f S f U f US λmax I 0, (33) 2.3. Parametrization of FD systems and
= −1 T −1 T postfilter design Vs PsS f U f US S U Nbasis.
−1 T Observer-based FD system design for continuous LTI sys- Since, generally, ps, PsS f U f US are not unitary matrices, we tems has been widely studied in the literature [3–6]. In this have finally and the next subsections, the analog form of those known results will be briefly reviewed. Attention will be paid to the Ω V , Jth, d ⊂ Ω V , Jth, d , V = p N , DE s DE s,opt s s,max basis comparison study when it is special for discrete LTI systems. −1 T Vs,opt = PsS U Nbasis, Let (Mu(z), Nu(z)) be a left coprime factorization pair of −1 ( ), that is, ( ) = ( ) ( ), ( ), ( ) ∈ RH∞ Ω V , Jth, d ⊂ Ω V , Jth, d , V = P PN , Gu z Gu z Mu z Nu z Mu z Nu z DE s DE s,opt s s basis [10]. In [20], a parametrization of all LTI residual generators −1 T Vs,opt = PsS U Nbasis. for system (1)describedby (34) r(z) = R(z) M (z)y(z) − N (z)u(z) (35) With the following remarks we would like to conclude this u u subsection. is presented, where R(z) ∈ RH∞ is the so-called post- (i) Parity-space-based FD system design is characterized filter that is arbitrarily selectable. Suppose that Gu(z) = by the simple mathematical handling. It only deals (A, B, C, D) is a detectable state space realization of Gu(z). with matrix- and vector-valued operations. This fact Then Mu(z), Nu(z) can be computed as follows [10]: attracts attention from the industry for the application of parity-space-based methods. −1 Mu(z) = I − C(zI − A + LC) L, (ii) There is a one-to-one relationship between the parity- (36) −1 space approach and the observer-based approach that Nu(z) = D + C(zI − A + LC) (B − LD). allows the design of an observer-based residual gen- eratorbasedonagivenparityvector[13, 14]. Based It is now a well-known result that on this result, a strategy called parity-space design, − = − observer-based implementation has been developed, (i) Mu(z)y(z) Nu(z)u(z) y(z) y(z), where y(z) is the which makes use of the computational advantage of output estimation delivered by a full-order observer; RH parity-space approaches for the FD system design (se- (ii) given L1 and L2, there exists an ∞-invertible post- = − −1 − lection of a parity vector or matrix) and then realizes filter Q(z) I + C(zI A + L1C) (L2 L1) so that the solution in the observer form to ensure a numer- − ically stable and less consuming online computation. Mu,L1 (z)y(z) Nu,L1 (z)u(z) This strategy has been successfully used in the sensor- (37) = Q(z) M (z)y(z) − N (z)u(z) , fault detection in vehicles [15]. u,L2 u,L2 (iii) In the parity-space approaches, a high order s will improve the optimal performance index JPS but, on where (Mu,L1 (z), Nu,L1 (z)) and (Mu,L2 (z), Nu,L2 (z)) are the other side, increase the online computational ef- the left coprime factorization pair of Gu(z), which are fort [16]. By introducing a low-order IIR (infinite im- computed according to (36)withL = L1 and L = L2, pulse response) filter, the performance of the parity- respectively. relation-based residual generator can be much im- (iii) all LTI residual generators can be expressed by a se- proved without significant increase of the order of ries connection of a full-order observer and a postfilter, the parity relation [17]. Similar effect can be achieved and are therefore called observer-based residual gener- by the closed-loop-observer-based implementation, as ators. Moreover, the selection of the postfilter can be pointed out by [18]. done independent of the observer design. 6 Journal of Control Science and Engineering
Due to the latter fact, we concentrate in this subsection on becomesmorepopular,whereγ, β are some constants. The the selection of R(z). The dynamics of a residual generator FD system design is often formulated as maximizing β under (35)isgovernedby agivenγ. The third index type is often met in the robust control theory and formulated as = r(z) R(z)Mu(z) Gd(z)d(z)+G f (z) f (z) . (38) J f −d = α f R(z)Mu(z)G f (z) − αd R(z)Mu(z)Gd(z) , The full decoupling problem is to find the postfilter R(z)so (46) that where α f , αd > 0 are some given constants. The FD sys- R(z)M (z)G (z) = 0, R(z)M (z)G (z) =0. (39) tem design is then achieved by maximizing J f −d. In [22], it u d u f has been demonstrated that the above three types indices are
If the l2-norm of the residual signal is used as evaluation equivalent in a certain sense. With this fact in mind, in this function, then the optimal FD problem is formulated as op- paper we only consider optimizations under ratio-type in- timization problems dices (40)–(43). R(z)M (z)G (z) Solution to optimization problems – = u f ∞ (40) (42) sup JFRE,∞/∞(R) sup , R(z)∈RH∞ R(z)∈RH∞ R(z)Mu(z)Gd(z) ∞ Let Mu(z)Gd(z) = Gdo(z)Gdi(z) be a co-inner-outer factor- (40) RH ization of Mu(z)Gd(z)[10], where Gdo(z) is the ∞-left- R(z)Mu(z)G f (z) − invertible co-outer, Gdi(z) is the co-inner containing all the sup JFRE,−/∞(R) = sup , right half complex plane zeros of Mu(z)Gd(z) and satisfying R(z)∈RH∞ R(z)∈RH∞ R(z)Mu(z)Gd(z) ∞ ∗ Gdi(z)G (z) = I. Based on the relations (41) di G (z)G (z) ≤ G (z) G (z) , sup JFRE,i/∞(R) 1 2 ∞ 1 ∞ 2 ∞ R(z)∈RH∞ ≤ G1(z)G2(z) − G1(z) ∞ G2(z) − σ R e jω M e jω G e jω (42) = i u f σ G e jω G e jω ≤ G (z) σ G e jω , ∀i, sup i 1 2 1 ∞ i 2 R(z)∈RH∞ R(z)Mu(z)Gd(z) ∞ (47) sup JFRE,2/2(R) 1× ithasbeenprovenin[23, 24], similar to the results for con- R(z)∈RH∞ m tinuous LTI systems given in [22] and recently in [25], that 2π = jω jω jω ∗ jω optimization problems (40)–(42) are solved simultaneously sup R e Mu e G f e G f e 1× R(z)∈RH∞ m 0 by 2π = −1 × ∗ jω ∗ jω jω R(z) Gdo (z). (48) Mu e R e dω R e 0 For this reason, (48) is called unified solution. × M e jω G e jω G∗ e jω M∗ e jω u d d u Another solution to optimization problem (41), if G f (z) is RL∞-left-invertible, is × R∗ e jω dω , = −1 R(z) G fo(z), (49) (43) where G fo(z) is the co-outer of Mu(z)G f (z). jω jω jω The main purpose of the co-inner-outer factorization is where σi(R(e )Mu(e )G f (e )) represents the fault sensi- tivity at different levels at the frequency ω, to separate the nonminimum phase zeros so that the rest part RH of Mu(z)Gd(z)orMu(z)G f (z)is ∞-left invertible. Note jω jω jω R(z)Mu(z)G f (z) − = inf σ R e Mu e G f e that the co-inner-outer factorization is not unique. There- ω fore, the optimal postfilter R(z) is also not unique [26]. jω jω jω = inf σi R e Mu e G f e ω,i (44) Solution to optimization problem (43) though not a norm, it is interpreted as the worst-case fault The optimal solution to optimization problem (43)isafre- sensitivity. Optimization problems (40), (41), and (43)are quency selector as follows [27]: often called the H∞/H∞, H−/H∞,andH2/H2 optimization, = = R(z) fω0 (z)p(z), JFRE,2/2,opt sup λmax (ω), (50) respectively. ω The ratio-type performance index given in (40)and(43) where fω0 (z) is an ideal frequency-selective filter with the se- is the first one that was introduced for the FD purpose [11, lective frequency at ω0, which satisfies 21]. Currently, the index of the form jω jω = = fω0 e q e 0, ω ω0 R(z)Mu(z)Gd(z) <γ, R(z)Mu(z)G f (z) >β 2π jω jω ∗ jω ∗ jω (51) fω0 e q e q e fω0 e dω 0 (45) jω0 ∗ jω0 T = q e q e , ∀q z ∈ RH2, P. Zhang and S. X. Ding 7
jω λmax (ω), p(e ) are, respectively, the maximal generalized where the observer gain matrix L and the weighting matrix eigenvalue and corresponding eigenvector of the following W are design parameters. Due to its state space expression generalized eigenvalue-eigenvector problem: and close relation to the observer design, FDF study receives most research attention. The dynamics of residual generator jω jω jω ∗ jω ∗ jω p e λmax ω Mu e Gd e Gd e Mu e (57)isgovernedby (52) jω jω ∗ jω ∗ jω − Mu e G f e G e M e = 0 f u r(z) = Grd(z)d(z)+Grf(z) f (z),
−1 Grd(z) = W Fd + C(zI − A + LC) Ed − LFd , (58) and ω0 is the frequency at which λmax (ω) achieves its maxi- −1 mum, that is, Grf(z) = W F f + C(zI − A + LC) E f − LF f .
λmax (ω0) = sup λmax (ω). (53) In the FDF design, the full-decoupling problem is to design ω L and W such that In practice, usually a narrow bandpass filter is implemented G (z) = 0, G (z) =0 (59) as frequency selector. From the viewpoint of FAR and MDR, rd rf the frequency selector may cause loss of fault sensitivity and whiletheoptimalFDproblemsareformulatedsoasto restrict the application of the H2/H2 optimal residual gener- choose matrices L, W that solve the optimization problem ator. [5, 7, 31] Recently, [27] reported a very interesting result on the re- lationship between the parity-space vector and the solution Grf(z)∞ sup J ∞ ∞(L, W) = sup , (60) to optimization problem (43). It has been shown that OBS, / G (z) L,W L,W rd ∞ T T Grf(z) − vsH f ,sH vs f ,s sup JOBS,−/∞(L, W) = sup , (61) lim min = λmax ωopt (54) s→∞ v T T L,W L,W Grd(z) ∞ s vsHd,sHd,svs jω σi Grf e sup JOBS, ∞(L, W) = sup . (62) and lim s→∞vs correspond to a bandpass filter. This result not i/ L,W L,W Grd(z) ∞ only reveals the physical interpretation of the standard op- timal selection of parity vectors but also provides us with Solution to optimization problems (60)–(62) an efficient tool to approximate the optimal solution to op- timization problem (43).Moreover,basedonit,advanced Because the FDF (57)isaspecialcaseof(35), the optimal parity-space approaches using wavelet transform have been solutions to optimization problems (60)–(62)canbederived proposed [28, 29]. based on the state space realization of the optimal postfilter As to the residual evaluation and threshold determina- R(z)givenby(48), as shown in [24]. A unified optimal so- tion, the l2-norm is the mostly used evaluation function, lution to optimization problems (60)–(62) is associated to a which leads to discrete-time algebraic Riccati system (DTARS) [24, 32]. As- sume that (A, Ed, C, Fd) is regular, then J =r2, = = (55) =− T = Jth sup r 2 R(z)Mu(z)Gd(z) ∞ max d 2. L Ld , W Wd (63) f =0,d solve optimization problems (60)–(62) simultaneously, In case of applying the unified solution (48), we have where Wd is the left inverse of a full column rank matrix Hd T T T satisfying HdH = CXdC + FdF and (Xd, Ld) is the stabi- = = d d Jth sup r 2 max d 2. (56) lizing solution to the DTARS = f 0,d AX AT − X + E ET AX CT + E FT I Analog with the results in [30] and the discussion in the last d d d d d d d = 0. (64) CX AT + F ET CX CT + F FT L subsection, it can be proven that the unified solution (48) d d d d d d d minimizes MDR under a given FAR, where MDR and FAR An alternative solution to optimization problem (61), if are defined in the context of the norm-based residual evalu- (A, E , C, F )isregular,isgivenby ation. f f =− T = L L f , W W f , (65) 2.4. fault detection filter design where W f is the left inverse of a full column rank matrix H f fault detection filter (FDF) is a special kind of observer-based T = T T satisfying H f H f CXf C + F f F f and (X f , L f ) is the stabi- FD systems (35) with a constant postfilter and constructed as lizing solution to the DTARS
= − x(k +1) Ax(k)+Bu(k)+L y(k) y(k) , AX AT − X + E ET AX CT + E FT f f f f f f f I = = − T T T T 0. r(k) W y(k) y(k) , (57) CXf A + F f E f CXf C + F f F f L f y(k) = Cx(k)+Du(k), (66) 8 Journal of Control Science and Engineering
Recently, application of LMI-technique (linear matrix in- The FD problem of periodic systems has been considered in equality) to solve (60)and(61) for continuous LTI systems [24, 32, 43–45]. Basically, there are two ways to handle the has been reported [33–36]. The core of those approaches FD problem of LTP systems, as shown below. consists in formulating (60)or(61) as a multiobjective op- 3.1. FD schemes based on lifted LTI reformulation timization problem and solving them based on an iterative computation of two LMIs. It can be proven that these so- It is well known that there is a strong correspondence be- lutions can, in the ideal case, converge to the optimal solu- tween discrete LTP systems and discrete LTI systems [42]. tion (63). The extension of these results to the discrete FDF Therefore, FD system design for the LTP system (72)canbe is straightforward and will not be discussed in this paper. carried out as follows: At the end of this subsection, we would like to introduce (i) lift the LTP system (72) into a discrete LTI system, a very interesting result achieved by the comparison study (ii) design residual generator(s) based on the lifted LTI re- between the well-known Kalman-filter-based residual gener- formulation, ation [37] and the unified solution (63). Given system (iii) using either parallel residual generators or select the parameters of the residual generator to satisfy the x(k +1)= Ax(k)+Bu(k)+Edζ(k), (67) causality condition. y(k) = Cx(k)+Du(k)+Fdη(k), Let ΨA(k)(k1, k0)(k1 ≥ k0) denote the state transition where ζ(k), η(k) are independent zero-mean Gaussian white matrix of LTP system (72) noise processes with cov(ζ(k)ζT (k)) = I,cov(η(k)ηT (k)) = if = , I, then a Kalman filter with = I k1 k0 ΨA(k) k1, k0 A k1 − 1 A k1 − 2 ···A k0 if k1 >k0. x k+1 | k =Ax k | k−1 +Bu(k) +L y(k)− y k | k−1 , (73) (68)
y k | k − 1 = Cx k | k − 1 + Du(k), (69) Periodic system (72) can be lifted into a discrete LTI system described by [42] −1 L = APCT CPCT + R , R = F FT (> 0), (70) d d = − xτ (k +1)θ + τ Aτ xτ (kθ + τ)+Bτ uτ (kθ + τ) T − − T T 1 T T = APA P APC CPC +R CPA +EdEd 0, P>0, + Ed,τ dτ (kθ + τ)+E f ,τ fτ (kθ + τ), (71) (74) y (kθ + τ) = C x (kθ + τ)+D u (kθ + τ) delivers an innovation as residual signal, where x(k | k − 1) is τ τ τ τ τ a prediction of x(k) based on the data up to k−1. Comparing + Fd,τ dτ (kθ + τ)+F f ,τ fτ (kθ + τ), the above Kalman filter algorithm with the unified solution where τ is an integer between 0 and θ − 1 denoting the initial (63) makes it clear that both solutions are quite similar. In- time, the state vector of the lifted system is x (kθ+τ) = x(kθ+ deed, (67) can be brought into the general form of (1)by τ τ), η with η standing for u, y, d, f is the augmented signal = T T T τ letting d [ ζ η ] and, as a result, (71) can be regarded as a defined by special case of (64). Remember that the Kalman filter deliv- ⎡ ⎤ η(kθ + τ) ers, in the context of statistic tests, a minimum MDR under a ⎢ ⎥ given FAR. In comparison, in the context of norm-based def- ⎢ η(kθ + τ +1) ⎥ = ⎢ ⎥ inition of MDR and FAR, the unified solution (63)provides ητ (kθ + τ) ⎢ . ⎥ , ⎣ . ⎦ us with the same result. η(kθ + τ + θ − 1) 3. FD OF DISCRETE-TIME LINEAR PERIODIC SYSTEMS A = Ψ (τ + θ, τ), τ A(k) ! ··· In this section, we review some recent results on FD in dis- Bτ = X Y Z , crete linear time periodic (LTP) systems. LTP is a special kind ⎡ ⎤ C(τ) of linear time-varying systems described by ⎢ ⎥ ⎢ C(τ +1)Ψ (τ +1,τ) ⎥ ⎧ ⎢ A(k) ⎥ ⎪ Cτ = ⎢ . ⎥ , ⎪x(k +1) ⎣ . ⎦ ⎨⎪ . = A(k)x(k)+B(k)u(k)+E (k)d(k)+E (k) f (k), d f C(τ + θ − 1)ΨA(k)(τ + θ − 1, τ) ΣLTP : ⎪ ⎪y(k) ⎡ ⎤ ⎪ D O ··· O ⎩ = ⎢ τ,1,1 ⎥ C(k)x(k)+D(k)u(k)+Fd(k)d(k)+F f (k) f (k), ⎢ ⎥ ⎢ .. . ⎥ ⎢Dτ,2,1 Dτ,2,2 . . ⎥ (72) D = ⎢ ⎥ , τ ⎢ . . . ⎥ ⎣ . .. .. ⎦ where A(k), B(k), C(k), D(k), Ed(k), E f (k), Fd(k), and F f (k) . O are known bounded and real periodic matrices of period θ, D , ,1 ··· D , , −1 D , , ⎧ τ θ τ θ θ τ θ θ that is, ∀k, A(k + θ) = A(k), and so forth. There is not ⎪D(τ + i − 1) for i= j, only continuous interest and development of periodic con- ⎨ = − − Dτ,i,j ⎪C(τ + i 1)ΨA(k)(τ + i 1, τ + j) trol and filtering theory [38–42], but also increasing appli- ⎩⎪ cations of periodic control in practice like helicopter vibra- B(τ + j − 1) for i>j, tion control, satellite attitude control as well as wind turbine. (75) P. Zhang and S. X. Ding 9
where Bτ,1 = ΨA(k)(τ + θ, τ +1)B(τ), Bτ,2 = ΨA(k)(τ + θ, τ + r(kθ + τ + j) = − 2)B(τ +1),Bτ,θ B(τ + θ 1) ,andEd,τ , Fd,τ and E f ,τ , F f ,τ = [ Wτ,j+1,1 ··· Wτ,j+1,j+1 ] ⎛⎡ ⎤ ⎡ ⎤⎞ are defined in a way similar to Cτ , Dτ . y(kθ + τ) y(kθ + τ) ⎜⎢ ⎥ ⎢ ⎥⎟ ⎜⎢ . ⎥ ⎢ . ⎥⎟ × ⎝⎣ . ⎦ − ⎣ . ⎦⎠ , 3.1.1. Observer-based FD system design and y(kθ + τ + j) y(kθ + τ + j) implementation y(kθ + τ + j) Assume that (A(k), C(k)) is detectable. Then (A , C )isde- = C(τ + j)Ψ ( )(τ + j, τ)x (kθ + τ) τ τ A k τ ⎡ ⎤ tectable and an observer-based LTI residual generator can be u(kθ + τ) designed based on lifted reformulation (74)as ! ⎢ ⎥ ··· ⎢ . ⎥ + Dτ,j+1,1 Dτ,j+1,j+1 ⎣ . ⎦ ,
u(kθ + τ + j) x τ (k +1)θ + τ = Aτ x τ (kθ + τ)+Bτ u τ (kθ + τ) = − j 0, 1, ..., θ 1. − + Lτ yτ (kθ + τ) yτ (kθ + τ) , (78) (76) r (kθ + τ) = W y (kθ + τ) − y (kθ + τ) , In this case, the state estimation is still updated at every τ τ τ τ θ time instants, but at each time instant kθ + τ + j, j = 0, 1, ..., θ − 1, a residual signal r(kθ + τ + j)is y (k) = C x (kθ + τ)+D u (kθ + τ), τ τ τ τ τ calculated from control input and measured outputs available up to the time instant kθ + τ + j. where Lτ and Wτ are constant matrices and can be de- signed with FD approaches for the discrete LTI systems intro- 3.1.2. Parity-relation-based FD system design and duced in the last section to realize full decoupling or optimal implementation FD. Observer (76) reconstructs the outputs over one period − y(kθ + τ), ..., y(kθ + τ + θ 1) based on the estimation of Similarly, a parity-relation-based LTI residual generator can state vector x(kθ + τ). Both the state vector of observer (76) be built as follows: and the residual signal are updated every θ time instants. In fault detection, the detection delay should be as small rτ (kθ + τ) = Vτ,s y τ,k,s − Hu,su τ,k,s , as possible. Therefore, it is advantageous if a residual signal ⎡ ⎤ y τ (k − s)θ + τ can be obtained at each time instant. To this aim, ⎢ ⎥ ⎢ − ⎥ ⎢yτ (k s +1)θ + τ ⎥ (i) a bank of LTI residual generators (76)canbeused,each ⎢ ⎥ y = ⎢ ⎥ , of which is designed for τ = 0, 1, ..., θ −1, respectively τ,k,s ⎢ . ⎥ ⎣ . ⎦ [43]. This scheme is characterized by a simple design but needs much online computational efforts. yτ (kθ + τ)
(ii) If the weighting matrix Wτ is designed to satisfy the ⎡ ⎤ u τ (k − s)θ + τ causality constraint, that is, Wτ is a lower block trian- ⎢ ⎥ ⎢ − ⎥ ⎢uτ (k s +1)θ + τ ⎥ gular matrix in the form of ⎢ ⎥ (79) u = ⎢ ⎥ , ⎡ ⎤ τ,k,s ⎢ . ⎥ W O ··· O ⎣ . ⎦ ⎢ τ,1,1 ⎥ ⎢ ⎥ ⎢ . . ⎥ uτ (kθ + τ) ⎢W W .. . ⎥ ⎡ ⎤ = ⎢ τ,2,1 τ,2,2 ⎥ Wτ ⎢ ⎥ , (77) D O ··· O ⎢ . ⎥ ⎢ τ ⎥ ⎢ . .. .. ⎥ ⎢ ⎥ ⎣ . . . O ⎦ ⎢ . . ⎥ ⎢ .. . ⎥ ··· ⎢ Cτ Bτ Dτ . ⎥ Wτ,θ,1 Wτ,θ,θ−1 Wτ,θ,θ H = ⎢ ⎥ , u,s ⎢ ⎥ ⎢ . .. .. ⎥ ⎣ . . . O ⎦ then, for a given integer τ, the residual generator (76) s−1 ··· can be implemented as Cτ Aτ Bτ Cτ Bτ Dτ where is a constant parity matrix, Vτ,s x τ (k +1)θ + τ ⎡ ⎤ ⎢ Cτ ⎥ = ⎢ ⎥ Aτ xτ (kθ + τ)+Bτ uτ (kθ + τ) ⎢Cτ Aτ ⎥ V ⎢ ⎥ = 0. (80) ⎛⎡ ⎤ ⎡ ⎤⎞ τ,s ⎢ . ⎥ y(kθ + τ) y(kθ + τ) ⎣ . ⎦ ⎜⎢ ⎥ ⎢ ⎥⎟ s ⎜⎢ ⎥ ⎢ ⎥⎟ Cτ Aτ ⎜⎢ . ⎥ − ⎢ . ⎥⎟ + Lτ ⎜⎢ . ⎥ ⎢ . ⎥⎟ , ⎝⎣ ⎦ ⎣ ⎦⎠ To get a residual signal at each time instant, we can use a y(kθ + τ + θ − 1) y(kθ + τ + θ − 1) bank of parity-relation-based residual generators, each one 10 Journal of Control Science and Engineering
for τ = 0, 1, ..., θ − 1, respectively. Alternatively, we can also where Vk is a θ-periodic parity matrix (or vector) that sat- impose a structural constant on parity matrix Vτ,s as isfies VkHo,s,k = 0, equation (85) represents the residual dy- ··· namics. Vτ,s = Vτ,s,0 Vτ,s,1 Vτ,s,s−1 Vτ,s,s , ⎡ ⎤ If the rank condition (Vτ,s,s) O ··· O ⎢ 1,1 ⎥ ⎢ . . ⎥ ⎢ .. . ⎥ rank Ho,s,k Hd,s,k H f ,s,k > rank Ho,s,k Hd,s,k (86) = ⎢(Vτ,s,s)2,1 (Vτ,s,s)2,2 . ⎥ Vτ,s,s ⎢ ⎥ , ⎣ . .. .. ⎦ . . . O holds for any k, then a full decoupling can be achieved by ··· (Vτ,s,s)θ,1 (Vτ,s,s)θ,θ−1 (Vτ,s,s)θ,θ solving (81) = = that is, the last block in Vτ,s is a lower triangular matrix, then Vk Ho,s,k Hd,s,k 0, VkH f ,s,k 0 (87) for a fixed τ, residual generator (79) can be implemented in such a way that only control inputs and measured outputs for Vk over one period at k, k +1,..., k + θ − 1. The residual available up to the time instant kθ + τ + j are needed for the evaluation consists in detecting the deviation of residual r(k) calculation of r(kθ + τ + j), j = 0, 1, ..., θ − 1. from 0. Especially, if 3.2. FD schemes based on periodic model rank Ho,s,0···Ho,s,θ−1 Hd,s,0···Hd,s,θ−1 H f ,s,0···H f ,s,θ−1 In this subsection, we will show that the parity-space ap- > rank Ho,s,0 ··· Ho,s,θ−1 Hd,s,0 ··· Hd,s,θ−1 , proach and the observer-based FD approach can be directly (88) extended to periodic systems, which do not need the tempo- rary step of lifted LTI reformulation and lead to a simplified then the full decoupling can be achieved by a constant par- design and implementation. ity matrix (or vector) Vk. However, condition (88) is rather restrictive in practice. 3.2.1. Periodic parity-space approach In case that a full decoupling is not achievable, optimiza- tion problems similar to (15)–(17) are formulated as The extension of the parity-space approach to periodic sys- tems is straightforward, because the parity-space approach 2 can handle each time instant independently [45]. The input- σ VkH f ,s,k max JLTP,PS,∞ ∞ V = max , = / k = 2 output relation of periodic system (72) during the moving Vk,VkHo,s,k 0 Vk ,VkHo,s,k 0 σ VkHd,s,k window [ − , ] can be expressed by (89) k s k 2 σ VkH f ,s,k y , = H , , x(k − s)+H , , u , + H , , d , + H , , f , . (82) k s o s k u s k k s d s k k s f s k k s max JLTP,PS,− ∞ V = max , = / k = 2 Vk,VkHo,s,k 0 Vk ,VkHo,s,k 0 σ VkHd,s,k While the vectors yk,s, uk,s, dk,s,and fk,s in (82)arebuiltin (90) exactly the same way as in the LTI case according to (6), the 2 σ V H , , matrices H , , , H , , , H , , ,andH , , in (82)arenotcon- i k f s k o s k u s k d s k f s k max JLTP,PS, ∞ V = max , = i/ k = 2 stant matrices but the following periodic matrices: Vk,VkHo,s,k 0 Vk ,VkHo,s,k 0 σ VkHd,s,k ⎡ ⎤ (91) C(k − s) ⎢ ⎥ ⎢ − − ⎥ ⎢ C(k s +1)A(k s) ⎥ = ⎢ ⎥ which are solved over one period to get the optimal peri- Ho,s,k ⎢ . ⎥ , ⎣ . ⎦ odic parity matrix Vk. Because the parity-space approach handles each time instant independently and there is no sta- C(k)A(k − 1) ···A(k − s +1)A(k − s) ⎡ ⎤ bility problem, the solutions of problems (87)–(91)ateach D(k − s) O ··· O time instant are independent of each other and can be ob- ⎢ ⎥ ⎢ ⎥ tained following the procedures introduced in Section 2.2. ⎢ .. . ⎥ ⎢C(k − s +1)B(k − s) D(k − s +1) . . ⎥ The threshold Jth can be calculated by (23)or(26)accord- ⎢ ⎥ = ⎢ ⎥ Hu,s,k ⎢ . .. .. ⎥ , ing to the requirement on FAR, while the residual evaluation ⎢ . . . O ⎥ ⎢ ⎥ function is selected as the amplitude (22) of the residual sig- ⎣ C(k)A(k − 1) ··· ··· D(k)⎦ nal. A(k − s +1)B(k − s) If the order s of the parity relation (82) is an integer mul- (83) tiple of the period θ, then the periodic parity-space approach is equivalent with a bank of residual generators (79). In com- Hd,s,k,andH f ,s,k are similar as Hu,s,k with B(j), D(j)replaced, parison, the periodic parity space approach provides more = − respectively, by Ed(j), Fd(j), and E f (j), F f (j), j k s, ..., k. flexibility. The order of the parity relation s needs not to be A periodic residual generator can be built as related to the period .Moreover, may take different val- θ s ues at different time instants. In this case, the threshold for r(k) = Vk yk,s − Hu,s,kuk,s , (84) the residual evaluation may need to be chosen differently at (a) = Vk Hd,s,kdk,s + H f ,s,k fk,s , (85) different time. P. Zhang and S. X. Ding 11
3.2.2. Periodic observer-based approach where W f (k) is the left inverse of a full column rank ma- T = T trix H f (k) satisfying H f (k)H f (k) C(k)X f (k)C (k)+ Assume that (A(k), C(k)) is detectable. A periodic observer- T based residual generator can be constructed as F f (k)F f (k), and (X f (k), L f (k)) is the stabilizing solution to the DPRS x(k +1)= A(k)x(k)+B(k)u(k)+L(k) y(k) − y(k) , = − Ω f ,11 Ω f ,12 I r(k) W(k) y(k) y(k) , × = T 0, (99) y(k) = C(k)x(k)+D(k)u(k), Ω f ,12 Ω f ,22 L f (k) (92) = T − T where L(k)andW(k)areθ-periodic observer gain matrix where Ω f ,11 A(k)X f (k)A (k) X f (k +1)+E f (k)E f (k), = T T = and weighting matrix, respectively. The residual dynamics is Ω f ,12 A(k)X f (k)C (k)+E f (k)F f (k), and Ω f ,22 T T governed by C(k)X f (k)C (k)+F f (k)F f (k). It is interesting to note the following connections between different approaches. e(k +1)= A(k)−L(k)C(k) e(k)+ E (k)−L(k)F (k) d(k) d d (i) Periodic observer-based residual generator (92)canbe + E f (k) − L(k)F f (k) f (k), rewritten into the form of lifted reformulation-based = r(k) W(k) C(k)e(k)+Fd(k)d(k)+F f (k) f (k) , LTI observer (76)with (93) ··· where e(k) = x(k) − x(k). Lτ = Lτ,1 Lτ,2 Lτ,θ , To enhance the robustness of the FD system to the un- = − ( + , + ) ( + − 1), known disturbances without loss of the sensitivity to the Lτ,i ΨA(k) L(k)C(k) τ θ τ i L τ i ⎡ ⎤ faults, the optimal design problem is formulated as W O ··· O ⎢ τ,1,1 ⎥ ⎢ ⎥ sup J ∞ ∞ L(k), W(k) ⎢ . . ⎥ LTP,OBS, / ⎢ . . ⎥ L(k),W(k) ⎢Wτ,2,1 Wτ,2,2 . . ⎥ W = ⎢ ⎥ , (94) τ ⎢ . ⎥ sup d=0, f ∈l −{0} r 2/ f 2 ⎢ . .. .. ⎥ = 2 ⎣ . . . O ⎦ sup , L(k),W(k) sup f =0,d∈l2−{0} r 2/ d 2 W , ,1 ··· W , , −1 W , , τ θ τ θ θ τ θ θ sup JLTP,OBS,− ∞ L(k), W(k) / Wτ,i,j = W(τ + i − 1), if i = j, L(k),W(k) (95) =− − − inf d=0, f ∈l −{0} r2/ f Wτ,i,j W(τ + i 1)C(τ + i 1) = sup 2 2 . sup = ∈ −{ } r /d L(k),W(k) f 0,d l2 0 2 2 × ΨA(k)−L(k)C(k)(τ + i − 1, τ + j)L(τ + j − 1) The solutions of optimization problems (94)-(95)arede- if i>j, i = 1, 2, ..., θ, j = 1, 2, ..., θ. rived by solving an equivalent optimization problem for the (100) cyclically lifted LTI systems first and then recover the periodic matrices L(k)andW(k)[24]. Recalling the discussion in Section 3.1.1, the physical meaning is that the periodic observer-based residual Solution to optimization problems (94)-(95) generator naturally satisfies the causality condition. It is further proven that, if the parameters ( ( ), ( )) Assume that (A(k), Ed(k), C(k), Fd(k)) is regular. Then L k W k of the periodic observer-based residual generator (92) =− T = L(k) Ld (k), W(k) Wd(k) (96) solve optimization problems (94)or(95), then the pa- rameters (L , W )oftheLTIobserver(76)gotby(100) solve optimization problems (94)-(95) simultaneously, τ τ will solve optimization problems in the form of (60)- where Wd(k) is the left inverse of a full column rank ma- T = T (61). trix Hd(k) satisfying Hd(k)Hd (k) C(k)Xd(k)C (k)+ T (ii) Similar as in LTI systems, the periodic parity-space Fd(k)Fd (k), and (Xd(k), Ld(k)) is the stabilizing solution to the difference periodic Riccati system (DPRS) approach and periodic observer-based approach are closely related. Assume that the periodic vector Ωd,11 Ωd,12 I × = 0 , (97) T ··· Ωd,12 Ωd,22 Ld(k) vk = vk,0 vk,1 vk,s (101) = T − T where Ωd,11 A(k)Xd(k)A (k) Xd(k +1)+Ed(k)Ed (k), = T T = satisfies vkHo,s,k = 0. Then a periodic functional Ωd,12 A(k)Xd(k)C (k)+Ed(k)Fd (k), and Ωd,22 T T observer-based residual generator in the form of C(k)Xd(k)C (k)+ Fd(k)Fd (k). An alternative solution to problem (95), if (A(k), E f (k), C(k), F f (k)) is regular, is given by z(k +1)= G(k)z(k)+H(k)u(k)+L(k)y(k), (102) =− T = L(k) L f (k), W(k) W f (k), (98) r(k) = w(k)z(k)+q(k)u(k)+p(k)y(k) 12 Journal of Control Science and Engineering
with z ∈ Rs, r ∈ R, can be readily obtained as [45] d(t) f (t) ⎡ ⎤ 00··· 0 g ⎢ k,1 ⎥ ⎢ ⎥ u(t) y(t) ⎢ . . ⎥ Continuous-time ⎢ . . ⎥ process ⎢10 . . gk,2 ⎥ ⎢ ⎥ = ⎢ ⎥ G(k) ⎢ ...... ⎥ , ⎢ . . . 0 . ⎥ ⎢ ⎥ ⎢ ⎥ ⎣0 ··· 10gk,s−1⎦ υ(k) ψ(k) D/AControl unit A/D 0 ··· 01gk,s ⎡ ⎤ ⎡ ⎤ vk+s,0 gk,1 ⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢vk+s−1,1⎥ ⎢gk,2⎥ (103) Information FDI unit ⎢ ⎥ ⎢ ⎥ L(k) =−⎢ ⎥ − ⎢ ⎥ v , about fault ⎢ . ⎥ ⎢ . ⎥ k,s ⎣ . ⎦ ⎣ . ⎦ Computer
v − g k+1,s 1 k,s Figure 1: Schematic description of the application of an FDI system w(k) = 00··· 0 −1 , in a process control system.
p(k) = vk,s, H(k) = T(k +1)B(k) − L(k)D(k), 4.1. System description q(k) =−p(k)D(k), Assume that, in the SD systems, the process is a continuous LTI process represented by where periodic scalars gk,1, ..., gk,s appearing in ma- = trices G(k), L(k) are free parameters and should be x˙(t) Acx(t)+Bcu(t)+Edcd(t)+E fcf (t), (104) selected to guarantee the stability of G(k). More- y(t) = Cx(t), over, if vk realizes a full decoupling from the un- known disturbances, that is, vk[Ho,s,k Hd,s,k] = 0, then where Ac, Bc, Edc, E fc,andC are known constant matrices of the functional observer-based residual generator (102) appropriate dimensions. In single-rate sampled-data (SSD) also achieves a full decoupling, that is, r(k) = 0, systems, the A/D converter and the D/A converter are, re- ∀u(k), d(k). This provides an approach to design full spectively, described by decoupling observer-based residual generator. = (iii) We would like to point out that, for the LTI system ψ(k) y(kh), (105) (1), a residual generator with periodic gain matrix L(k) u(t) = υ(k), kh t<(k +1)h, (106) and periodic weighting matrix W(k) will not improve the FD performance under performance index (94). where h is the sampling period, ψ ∈ Rm is the sampled pro- cess output signal, υ ∈ Rp is the discrete-time control in- 4. FD OF SAMPLED-DATA SYSTEMS put sequence delivered by the controller program. In mul- tirate sampled-data (MSD) systems, the A/D converters and The study on FD problems of sampled-data (SD) systems has the D/A converters may work with different sampling rates been motivated by the digital implementation of controllers and thus modeled, respectively, by and monitoring systems. Figure 1 sketches a typical applica- l = l = l = tion of an FD system in a process control system. The pro- ψl(k ) yl(k Ty,l), l 1, 2, ..., m; k 0, 1, 2, ..., (107) cess under consideration is a continuous-time process. Both j j j the controller and the FD system are discrete-time systems uj (t) = υ j k , k Tu,j ≤ t< k +1 Tu,j , which are implemented on a computer system or on an em- (108) j = 1, 2, ..., p; k j = 0, 1, 2, ..., bedded microprocessor. The sensor output signals are dis- cretized by the A/D converters and then fed to the controller where Ty,l and Tu,j denote, respectively, the sampling peri- as well as to the FD system. The D/A converters convert the ods of the A/D converter in the lth output channel and the discrete-time control input signals into continuous-time sig- D/A converter in the jth input channel. A more general class nals. Since both continuous-time and discrete-time signals of systems are nonuniformly sampled-data (NSD) systems, exist in the system, the system design should be indeed con- where the sampling instants may be multirate, asynchronous, sidered from the viewpoint of an SD system [46, 47]. The and nonequidistantly distributed, that is, intersample behavior is the main factor that should be con- l l sidered in developing FD algorithms for the SD systems. In ψ k = y t l , l = 1, 2, ..., m; k = 0, 1, 2, ..., l l y,k practice, it happens often that the A/D and D/A converters = j ≤ are working at different sampling rates [48–53]. In this sec- uj (t) υ j k , tu,k j t where ty,kl represent the sampling instants in the lth output nal δ(t)withδ standing for d and f ,anoperatorisdefined channel and tu,k j the time instants at which the jth control as follows: input is updated. It is worth mentioning that a special kind ⎡ ⎤ − of NSD systems, where the sampling instants are nonequidis- ⎢ δ(k s) ⎥ ⎢ ⎥ tant spaced but periodic, has been studied in the literature ⎢δ(k − s +1)⎥ δ = ⎢ ⎥ Ψ δk,s(t) ⎢ . ⎥ rather intensively [54–57]. ⎣ . ⎦ δ(k) ⎡ ⎤ 4.2. FD of SSD systems h − Ac(h τ) − ⎢ 0e Eδc δ (k s)h + τ dτ ⎥ ⎢ h A (h−τ) ⎥ ⎢ e c Eδcδ (k − s +1)h + τ dτ⎥ Conventionally, an FD system can be designed for the SSD = ⎢ 0 ⎥ ⎢ . ⎥ . system by indirect approaches, that is, ⎣ . ⎦ . h Ac(h−τ) (i) analog design and SD implementation,or 0e Eδcδ(kh + τ)dτ (ii) discrete-time design based on the discretization of the (113) process model. The residual dynamics can be expressed with the help of op- Motivated by the development of sampled-data control [46, erators as 47], in the last years the FD problem of the SSD systems have d f been studied from the viewpoint of direct design to take into r(k) = VsH Ψ dk,s(t)+Ψ fk,s(t) , ⎡ ⎤ account the intersample behavior and eliminate the approxi- OO··· O ⎢ ⎥ mation made during the design [26, 58–60]. ⎢ . . ⎥ The dynamics of the SSD system at the sampling instants ⎢ CO.. . ⎥ (114) H = ⎢ ⎥ . can be described by ⎢ . . . ⎥ ⎣ . .. .. O⎦ CAs−1 ··· CO x(k +1)= Ax(k)+Bυ(k)+d(k)+ f (k), (110) ψ(k) = Cx(k), The influence of the continuous-time signal δ(t) over the time interval [(k − s)h,(k +1)h) on the discrete-time signal δ Ψ δk,s(t)ismeasuredby where ∗ rT (k)r(k) σ V HΨδ Ψδ HT V T = sup , h s s 2 δ∈L2,[(k−s)h,(k+1)h) δ 2,[(k−s)h,(k+1)h) = Ach = Act A e , B e Bcdt, ∗ T 0 δ δ T T = r (k)r(k) σ VsHΨ Ψ H Vs inf 2 , h δ∈L − − 2,[(k s)h,(k+1)h) δ 2,[(k−s)h,(k+1)h) = Ac(h τ) (111) d(k) e Edcd(kh + τ)dτ, (115) 0 h − ∗ = Ac(h τ) δ δ f (k) e E fcf (kh + τ)dτ. where (Ψ ) denotes the adjoint of the operator Ψ which 0 uniquely satisfies