Sie sind auf Seite 1von 8

Preprints of the 8th IFAC Symposium on Advanced Control of Chemical Processes The International Federation of Automatic Control Furama

Riverfront, Singapore, July 10-13, 2012

Data-Driven Design of Model-based Fault Diagnosis Systems


S. X. Ding

Institute for Automatic Control and Complex Systems (AKS), University of Duisburg-Essen, Duisburg, 47057, Germany

Abstract: In this paper, recent development of data-driven design of fault detection and isolation (FDI) systems is presented. The major attention and focus are on the design schemes for observer-based FDI systems. Keywords: Fault detection and diagnosis; data-driven methods; observer-based methods; subspace identication methods; parity space methods 1. INTRODUCTION Process monitoring and fault diagnosis are currently receiving considerably increasing attention in the application and research domains. In the monographs by Gertler [1998], Mangoubi [1998], Chen and Patton [1999], Patton et al. [2000], Russell et al. [2000], Gustafsson [2000], Chiang et al. [2001], Blanke et al. [2003], Simani et al. [2003], Isermann [2006], Ding [2008], both the data-driven and model-based schemes and methods are presented and well described. The recent surveys by Venkatasubramanian et al. [2003b,a], Qin [2003], Zhang and Ding [2008], Mangoubi et al. [2009], Qin [2009], Hwang et al. [2010] provide the reader with the excellent review of the current development of the advanced and sophisticated fault diagnosis schemes and their applications. As reported in (Venkatasubramanian et al. [2003a], Qin [2003, 2009]), data-driven techniques are widely applied in the process industry for fault detection and isolation (FDI). Among numerous data-driven FDI schemes, the multivariate analysis technique with PCA (principal component analysis) and PLS (partial least squares) as its representative methods is, thanks to its simplicity and eciency in processing huge amount of process data, recognized as a powerful tool for addressing statistical process monitoring and FDI problems. Although dynamic PCA/PLS (Russell et al. [2000], Li and Qin [2001]), recursive implementation of PCA/PLS (Li et al. [2000], Qin [1998]), fast moving window PCA (Wang et al. [2005]) have been developed in recent years, the multivariate analysis technique seems only ecient in dealing with dynamic processes in the steady state and at higher levels in largescale systems. In comparison, the model-based FDI technique, thanks to the application of advanced system and control theory, provides a more ecient and powerful tool to investigate FDI issues in highly dynamic systems and control loops which are generally located at the process level. The possible high FDI performance is often achieved at the cost of a highly complex process modelling and, based on it, a sophisticated FDI system design. Recently, Qin and Li [2001], Ding et al. [2009b], Dong and Verhaegen [2009], Kulcsar et al. [2009], Dong [2009] have
IFAC, 2012. All rights reserved. 840

proposed SIM (subspace identication methods) based FDI design schemes, in which the FDI systems are directly designed utilizing the collected process data without explicitly identifying a system model. Some of these results have been successfully applied in the industrial processes and also extended for instance to the linear parameter varying systems. The major advantage of these methods is that they can provide high FDI performance similar to the model-based FDI methods but without a sophisticated system design. The objective of this paper is to present recent development of the data-driven design of FDI systems that are applicable for dynamic processes. Our major attention and focus are on the so-called observer-based FDI systems, which provide not only high FDI performance but also high real-time ability. The paper is organized as follows. A review of the parameterization of the observer-based FDI systems is rst presented in Section 2, based on which basic ideas and requirements of data-driven design of FDI systems for dynamic processes are formulated. Section 3 is dedicated to the data-driven schemes for open-loop congured FDI systems, while Section 4 addresses data-driven design of observer-based (and thus closed-loop structured) FDI systems. In Sections 5 and 6, adaptive, fault isolation and identications issues are briey addressed. Notation: The notation adopted throughout this paper is fairly standard. We use Q(i : j, p : q) denoting the sub-matrix consisting of the i-th to the j-th rows and the p-th to the q-th columns of Q. kkF is used for the Frobenius norm of a matrix. E() represents mean value. x N (0, ) means that x is normal distributed with zero mean and covariance (matrix) . X 2 (l) stands for X 2 distribution with l degrees of freedom and, associated to it, prob( > (l)) = for the probability of > (l) equal to (signicance level). 2. BASIC IDEAS AND PROBLEM FORMULATION In this section, we rst review observer-based fault detection schemes briey. It will motivate the problem formulation for the development of data-driven design of modeland observer-based FDI systems.

8th IFAC Symposium on Advanced Control of Chemical Processes Furama Riverfront, Singapore, July 10-13, 2012

2.1 Reviewing observer-based fault detection schemes Consider a process modelled by x(k + 1) = Ax(k) + Bu(k) + w(k) y(k) = Cx(k) + Du(k) + v(k) (1) (2)

2.2 Essential issues of data-driven design of FDI systems It follows from (5) that the core of the residual generation problem is to nd/identify the LCF of a process model. This motivates us to address the data-driven design of FDI systems in a framework, which can be schematically formulated as solving an (algebraic) equation described by Y U, =0 (10) U

where u Rl , y Rm and x Rn represent the process input, output and state variable vectors, respectively. w Rn and v Rm denote noise sequences that are normally distributed and statistically independent of u and x(0). We assume that the above process model is observable. Let Gu (z) = C(zI A)1 B + D. It is well-known (Zhou et al. [1996]) that the pair M (z), N (z) with N (z) = D + C (zI A + LC) M (z) = I C(zI A + LC)1 L
1

(3)

builds a left coprime factorization (LCF) of Gu (z), i.e. Gu (z) = M 1 (z)N (z), where L is a matrix of appropriate dimensions and can be interpreted as the observer gain matrix and thus is selected to ensure the system stability. A fundamental property of the LCF is that in the faultand noise-free case u(z) (z) M (z) u, N =0 (5) y(z) For this reason, the LCF, also called kernel representation of system (1)-(2), is used for the parameterization of residual generators, which is described by u(z) (6) r(z) = R(z) N (z) M (z) y(z) = R(z) Nw (z)w(z) + M (z)v(z) (7) where R(z) (6= 0) is a stable parameterization matrix (Ding and Frank [1990]), which is called post-lter. It is worth mentioning that (6) and (7) are exactly the parameterization forms of all LTI (linear time-invariant) residual generators and their dynamics . Moreover, in order to avoid loss of information about faults in the process, the condition rank(R(z)) = m is to be satised. Since u(z) = y(z) y (z) N (z) M (z) y(z) Nw (z) = C(zI A + LC)1

(B LD)

(4)

with Y, U denoting the (recorded) process output and input data sets and the data-driven realization of the LCF and so that the residual generator to be designed. This is an alternative way of designing a residual generator directly. A further important issue is the realization of the residual generator in a closed-loop conguration and in the (recursive) form of an FDF aiming at ensuring the required robustness on the one hand and real-time ability on the other hand. The last issue addressed in this paper is the development of adaptive schemes, which are used to enhance the robustness of the FDI systems. 3. METHODS OF DATA-DRIVEN DESIGN OF OPEN-LOOP CONFIGURED FD SYSTEMS In this section, we present three data-driven methods for the design of FD systems which are congurated in an open-loop structure and realized in form of an FIR (nite impulse response) lter. The rst two deal with the identication of I/O models, while the third one is based on a direct identication of the LCF in the form of (10). 3.1 I/O data models For the purpose of data-driven design of FD systems, an I/O model that describes the relation between the input and output data sets plays an essential role. Let (k) R be a data vector. We introduce the following notations (k) . . s (k) = (11) R(s+1) . (k + s) k = [ (k) (k + N 1) ] RN k,s = [ s (k) s (k + N 1) ] k . = . R(s+1)N . k+s (12) (13) (14)

with y delivered by a full order observer as an estimate of y (Ding [2008]), we can apply an observer, also called FDF (fault detection lter), of the form x(k + 1) = A(k) + Bu(k) + L (y(k) y (k)) x y (k) = C x(k) + Du(k) (8) (9)

for the on-line implementation of the residual generator. It is remarkable that the FDF realization (8)-(9) requires (considerably) less online computation than a direct implementation of (6). Moreover, thanks to its closed-loop conguration with the feedback of the residual signal y(k) y (k) in (8), the FDF realization is of higher robust ness against disturbances than a feedforward congured residual generator like, for instance, a parity space one. In the literature, the observer-based residual generators are often referred to as closed-loop structured, while the parity space one as open-loop.
841

where s, N are some (large) integer (Qin [2006], Huang and Kadali [2008]). In our study, (k) can be y(k), u(k), x(k), and represents m or l or n given in (1)-(2). The rst I/O data model described by Yk,s = s Xk + Hu,s Uk,s + Hw,s Wk,s + Vk,s C CA s = . R(s+1)mn . . CAs (15)

8th IFAC Symposium on Advanced Control of Chemical Processes Furama Riverfront, Singapore, July 10-13, 2012

0 .. .. CB . . R(s+1)m(s+1)l Hu,s = . .. .. . . . 0 . CAs1 B CB D follows directly from (1)-(2), where Hw,s Wk,s + Vk,s represents the inuence of the noise vectors on the process output with Hw,s having the same structure like Hu,s and Wk,s , Vk,s as dened in (11)-(13). D In the SIM framework, the so-called innovation form, instead of (1)-(2), is often applied to build an I/O model (Qin [2006], Huang and Kadali [2008]). The core of the innovation form is a Kalman lter, which is written as x(k + 1) = A(k) + Bu(k) + K (y(k) y (k)) x (16) y (k) = C x(k) + Du(k) (17) with the innovation y(k) y (k) := e(k) being a white noise sequence and K the Kalman lter gain matrix. Based on (16)-(17), the I/O relation of the plant can be alternatively written into x(k + 1) = A(k) + Bu(k) + Ke(k) x (18) x(k + 1) = AK x(k) + BK u(k) + Ky(k) (19) AK = A KC, BK = B DK y(k) = C x(k) + Du(k) + e(k) (20) The following two I/O data models follow respectively from (18), (20) and (19), (20): Yk,s = s Xk + Hu,s Uk,s + He,s Ek,s (21) I 0 .. .. CK . . R(s+1)m(s+1)m He,s = . .. .. . . . 0 . CAs1 K CK I K K (22) I Hy,s Yk,s = K Xk + Hu,s Uk,s + Ek,s s 0 0 .. .. CK . . K R(s+1)m(s+1)m Hy,s = . .. .. . . . 0 . CAs1 K CK 0 K D 0 C .. .. CAK . . K K . , Hu,s = CBK s = . . .. .. . . . . 0 . s CAK CAs1 BK CBK D K with Ek,s R(s+1)mN having the same structure as dened in (11)-(13). 3.2 Data-driven design of FIR fault detection systems Due to the unmeasurable Xk or Xk , the I/O models (15), (21) and (22) cannot be directly identied and applied for the residual generation purpose. To eliminate the inuence of Xk or Xk and further to generate residual signals, the following schemes are available. Scheme I It follows from (19) that X u(k i) i1 AK [ BK K ] x(k) = AK x(k ) + y(k i)
i=1

As a result, (21) can be re-written into Yk,s s,1 Zk,1 + Hu,s Uk,s + He,s Ek,s (24) Zk,1 = [ z1 (k ) z1 (k + N ) ] u(k ) y(k ) . . z1 (k ) = . u(k 1) y(k 1) 1 s,1 = s AK [ BK K ] [ BK K ] Now, identifying s,1 , Hu,s using process data Yk,s , Uk,s , Zk,1 (Huang and Kadali [2008]) allows the construction of a residual generator and further the computation of the covariance matrix of the residual vector as well as the threshold setting, as summarized in the following algorithm. Algorithm 3.1: FIR FD system design S1: Collect process data and build Zk,1 , Uk,s , Yk,s S2: Solve the least squares problem for s,1 , Hu,s mins,1 ,Hu,s kYk,s s,1 Zk,1 Hu,s Uk,s kF S3: Compute r =
1 N He,s Ek,s

Since the eigenvalues of AK are all located within the unit circle, for a large integer X u(k i) i1 (23) AK 0 = x(k) AK [ BK K ] y(k i)
i=1

(He,s Ek,s )T

S4: Set the threshold Jth = 2 ((s + 1)m)) In the above algorithm, He,s Ek,s = Yk,s s,1 Zk,1 Hu,s Uk,s with s,1 , Hu,s being identied in S2, which are then also applied to for the (on-line) residual generation rs (k +) = ys (k +)s,1 z1 (k)Hu,s us (k+) (25) Scheme II This scheme has been proposed by Dong [2009]. It is based on the I/O model (22), which is further re-written into, using (23),
K K Yk,s = K Zk,1 + Hy,s Yk,s + Hu,s Uk,s + Ek,s s,1 (26) 1 K K s,1 = s AK [ BK K ] [ BK K ]

K K Having constructed K , Hy,s , Hu,s using the Markov s,1 parameters CA1 B CA1 K CB CK D := , = s + K K which is identied applying the model X Uki i1 CAK [ BK K ] + DUk + Ek Yk Yki i=1

Uki = [ u(k i) u(k i + N ) ] Rl(N+1) i = 0, , s, Ek = [ e(k) e(k + N ) ] residual vector can then be generated by K rs (k + ) = I Hy,s ys (k + ) K zk,k+1 s,1

Yki = [ y(k i) y(k i + N ) ] Rm(N +1)

(27) + )

K Hu,s us (k

842

8th IFAC Symposium on Advanced Control of Chemical Processes Furama Riverfront, Singapore, July 10-13, 2012

Algorithm 3.2: FD system design (Dong [2009]) S1: Collect process data and build Zk,1 , Uk, , Yk S2: Solve the least squares problem for Zk,1 min Yk Uk
F K K S3: Form Hy,s , K , Hu,s using s,1 1 T S4: Compute r = N Ek,s Ek,s K K Ek,s = I Hy,s Yk,s K Zk,1 + Hu,s Uk,s s,1

Let = , R((s+1)mn)(s+1)m . It s s,y s,u s,y has been proven (Ding et al. [2009b]) that s = 0, = Hu,s (34) s,y s,u s,y

Thus, = is the so-called parity subspace (Ding s,y s [2008]). Moreover, He,s Ek,s = R33 Q3 (35) s,y s,y Algorithm 3.3: LCF based FD system design S1: Collect process data and build Zksp ,sp1 , Uk,s , Yk,s S2: Do QR-decomposition (28)
T S3: Do SVD (32) and set = U2 s

S5: Set the threshold Jth = 2 ((s + 1)m))

It is worth to remark that in the above two algorithms the least squares estimations using process data sets build the major computations. As pointed out in (Huang and Kadali [2008]), they can be implemented in a numerically robust way with a QR-decomposition of the process data as follows " # " #" # Zksp ,sp 1 R11 0 0 Q1 Q2 Uk,s = R21 R22 0 (28) R31 R32 R33 Q3 Yk,s where sp is some large integer. For instance, [ s,1 Hu,s ] in Algorithm 3.1 can be determined by R11 0 [ s,1 Hu,s ] = [ R31 R32 ] R21 R22

S4: Compute r =

1 T N s,y U2 R33

S5: Set the threshold Jth = 2 ((s + 1)m n))

T T s,y U2 R33

Moreover, the relation (Qin [2006]) (29) R33 Q3 = He,s Ek,s = Hw,s Wk,s + Vk,s can be used for the determination of the covariance matrix needed for the threshold setting. Scheme III Inspired by the work in Qin and Li [2001], Wang and Qin [2002], Ding et al. [2009b], the idea behind this scheme is the identication of the data-driven form of the LCF as described in (10). To this end, re-write (15) or (21) into Uk,s Uk,s (30) = s + Hw,s Wk,s + Vk,s Yk,s Xk Uk,s I 0 = s + He,s Ek,s , s = (31) Hu,s s Xk s = 0, R((s+1)mn)(s+1)(m+l) = s s Uk,s = (Hw,s Wk,s + Vk,s ) = He,s Ek,s s s s Yk,s s

On-line realization and computations The on-line realization of the above introduced FD systems can be described in the following general form u (k ) r(k) = (36) y (k ) where = s + for the rst two schemes and = s for the last one. Using the generated residual vector r(k), the T 2 statistic J can be built and a fault detection follows the detection logic: J > Jth =faulty, otherwise fault-free. It is clear that residual generator (36) is an FIR lter and open-loop congurated. With the measurement data in the time interval [k , k] as its input, (36) delivers a residual vector that may contain redundant information and is less robust against disturbances. 4. DATA-DRIVEN DESIGN OF OBSERVER-BASED FD SYSTEMS Although less data are needed than the rst two FD systems, the on-line computation of the LCF based FD system is still involved. In this section, we present schemes for the construction/realization of observer-based FD systems with the aid of the identied LCF . We begin s with the generation of a scalar residual signal and then introduce the design of a Kalman lter based FD system. 4.1 A parity vector based construction of an observer-based residual generator Given process model (1)-(2) and a parity vector s = [ s,0 s,1 s,s ] R(s+1)m with s,i Rm , i = 0, 1, , s, then 0 0 0 s,0 1 0 0 s,1 Az = . . . . . . Rss , Lz = . . . . . . . . . 0 1 0 s,s1 C s,1 s,2 s,s1 s,s 0 CA s,2 s,s T = . . . . . . . . . . . . s,s 0 0 CAs1

Since s R(s+1)(m+l)(n+(s+1)l) and for s n, (s + 1)(m + l) > n + (s + 1)l, there exists so that s

In other words, is a data-driven realization of the LCF and thus builds a residual generator. The identication of and associated with it He,s Ek,s can be done s s using QR-decomposition (28). By an SVD (singular value decomposition) T R21 R22 1 0 V1 = [ U1 U2 ] (32) 0 2 R31 R32 V2T R21 R22 and noting that s and have the same (left) R31 R32 null matrix, we have
T 2 0, = U2 R((s+1)mn)(s+1)(l+m) s

(37)

(38)

(33)
843

8th IFAC Symposium on Advanced Control of Chemical Processes Furama Riverfront, Singapore, July 10-13, 2012

solve Luenberger equations T A Az T = Lz C (39) (40) cz T = gC, cz = [ 0 0 1 ] , g = s,s A direct application of this result is the construction of an observer-based residual generator for the given parity vector s as follows z(k + 1) = Az z(k) + Bz u(k) + Lz y(k) Rs (41) r(k) = gy(k) cz z(k) dz u(k) R (42) Bz = T B Lz D, dz = gD (43) Note that D s,0 s,2 s,s1 s,s CB 0 CAB s,1 s,s Bz . = . . . . . dz . . . . . . s,s 0 0 s1 CA B s Hu,s (:, 1 : l) s Hu,s (:, l + 1 : 2l) = (44) . . . s Hu,s (:, sl + 1 : (s + 1)l)

Recall that delivered by Algorithm 3.3 is composed s of , with = . Applying Algorithm OR to s,y s,u s,y s yields Q1 , n . Now, multiply Q1 to and denote s,y s,u 1 1 the last row of Q1 by n,u . It holds, following (44), s,u 1 n,u (1 : l) n,u (l + 1 : 2l) Bz = n,u = n Hu,s = . . dz . n,u (nl + 1 : (n + 1)l) (47) In summary, we have Algorithm 4.2: Observer-based FD system design S1: Run Algorithm 3.3 with output , s,y s,u S2: Run Algorithm 4.1 S3: Compute Q1 for n,u s,u 1 S4: Form Bz ,dz , Lz according to (47), (38) S5: Construct residual generator (41)-(42) Although Algorithm 4.2 only delivers a scalar residual signal, it is fundamental for the development of multiple residual generation. It is well-known that an mdimensional residual vector is necessary for a reliable fault detection and isolation in the framework of observer-based FDI systems (Ding [2008]). For this purpose, m vectors should be selected from . Due to the space limitation, s this topic cannot be addressed in this paper. We refer the reader to (Ding et al. [2011, 2012]) for some recent results on this topic. 4.2 Data-driven design of Kalman lter based FD systems It is well known that a Kalman lter (18)-(20) delivers an innovation e(k) = y(k) y (k) with (48) E e(i)eT (j) = e ij

Equations (37), (40) and (44) allow a direct construction of an observer-based residual generator given by (41)-(42) using a row of and without any additional computation s and design eorts.

Remember that s is generally selected suciently large (typically much larger than n) during the identication phase. From the real-time computational viewpoint, it is of practical advantage to reduce the order of the residual generator as much as possible. For this purpose, Ding et al. [2009b] have proposed the following algorithm. Algorithm 4.1: Order reduction S1: Do a QR decomposition of V = [ Q1 R1 Q2 ] s 0 0 0 1 0 0 1 0 . with V = . . . Rm(s+1)m(s+1) . . . . . 0 1 0 0 1 0 0 0 R1 being a upper triangular, Q1 a with orthonormal columns and Q2 a n matrix S2: Compute Q1 = R1 Q1 Q2 V 1 s 1 1

As proven in (Ding et al. [2009b], Theorem 1), the result of the above algorithm is (45) Q1 = [ Q3 ] s 1 where = m(s + 1) n, Q3 is a n matrix and is of the form 1,1 1,2 1,1 1, 2,2 2,1 0 2,1 . . . R (46) . . . = . . . 0 0 1,1 1,2 ,1 0 0 0 It is evident that the last row of Q1 is of the form s 1 [ n 0 0 ] R

and builds, together with test statistic J = eT (k)1 e(k) e and the threshold Jth = 2 (m) an FD system with the optimal trade-o between the false alarm rate and fault detectability. This sub-section addresses (i) identication of the Kalman lter gain matrix K and the covariance matrix e based on the I/O model (21) and Algorithm 3.3 (ii) construction of a Kalman lter-based FD system.

Now, solve

Given , Zk,s = R33 Q3 = He,s Ek,s delivered s s s,y s,y by Algorithm 3.3. It turns out e 0 1 1 T T Ek,s Yk . = Zk,s Yk = . N N s,y . 0 e CKe 1 T (49) s,y He,s Ek,s Yk . s,y . N . CAs1 Ke (50)

and thus n R(n+1)m is a parity vector whose order is n.


844

s = 0 s,y for s and further compute

8th IFAC Symposium on Advanced Control of Chemical Processes Furama Riverfront, Singapore, July 10-13, 2012

for e , K. Without loss of generality, suppose that using a n-dimensional state vector z(k) is built which s,y satises (Ding et al. [2011]) z(k) = T x(k), Cz T = C, rank(T ) = n It follows from Luenberger equation (39) that

1 e 0 I T = s Zk,s Yk (51) Ke 0 s (1 : sm, :) N s 0 0 I I =I s s 0 s (1 : sm, :) 0 s (1 : sm, :)

T A = Az T + Lz C T AT 1 = Az + Lz Cz We now apply the identied s and parity vector n to form T according to (38) n,1 n,2 n,n1 n,n 0 n,2 n,n (52) T = . . . s . . . . . . n,n 0 0 Note that although s identied by solving (50) is not unique (up to a regular transformation), but the product T K is unique with K identied based on (51). As a result, we are able to construct a Kalman lter based residual generator as follows z(k + 1) = Az z(k) + Bz u(k) + T Ke(k) (53) r(k) = e(k) = y(k) Cz z(k) Dz u(k) (54) Az = Az + Lz Cz , Bz = Bz + Lz Dz It is straightforward to prove that the dynamics of the residual generator is governed by x(k + 1) = AK x(k) + T w(k) T Kv(k) (55) r(k) = Cx(k) + v(k) (56) 1 x(k) = T x(k) z(k), AK = T (A KC) T Considering that e(k) N (0, e ) , the following statistic and threshold are applied for the fault detection purpose J = rT (k)1 r(k), Jth = 2 (m) e Algorithm 4.3. Kalman lter based FD S1: Run Algorithm 3.3 S2: Compute
1 T N s,y R33 Q3 Yk

also an SVD. Motivated by this observation, Naik et al. [2010] proposed two schemes for recursively updating the identication of parity vectors. One of them is based on the rst-order perturbation (FOP) theory, which is a rank-one update of the eigenpairs of the data covariance matrix, and another one on the data-projection method (DPM) which serves as a simple approach for recursively updating the singular values and associated eigenvectors of interests. These two scheme can be used either to construct an (open-loop) adaptive residual generator of the form (36) or for the implementation of an adaptive observer-based residual generator. In this section, we present an approach proposed in (Ding et al. [2009a]) to the design and implementation of an adaptive residual generator. Dierent from the above-mentioned methods, the basic idea of this approach is the application of the well-established adaptive control methods (Astrm and Wittenmark [1995]) to the observer-based residual generator. 5.1 Problem formulation For the simplicity of our discussion, we only consider the residual generator (41)-(42) designed using Algorithm 4.2. It is evident that parameter changes in the original system matrices are now represented by the changes in Bz , Lz , g, dz . In the sequel, we assume that there exists no structural change, i.e. change e.g. in the observability, and the changes in parameters are slow and can be considered nearly constant in a (large) time interval. For our purpose, we now extend the residual generator (41)-(42) to z(k + 1) = Az z(k) + Bz u(k) + Lz y(k) + L0 r(k) (57) r(k) = gy(k) cz z(k) dz u(k) (58) where L0 provides additional degree of design freedom and should ensure the stability of Az = Az L0 cz . Let Bz L = u R(s+1)(m+l) , u = col , y = col g y dz Q (u(k), y(k)) = U(k) L0 uT (k) Y(k) L0 y T (k) U(k) = [ u1 (k) Iss ul (k) Iss ] Y(k) = [ y1 (k) Iss ym (k) Iss ] (57)-(58) can be re-written into z(k + 1) = Az z(k) + Q (u(k), y(k)) (59) T dz r(k) = uT (k) y T (k) cz z(k) (60) gT The task consists in designing a residual generator which is adaptive to and delivers a residual signal r(k) satisfying limk r(k) = 0 and, if possible, with an exponential converging speed independent of a constant change in . 5.2 The adaptive residual generator algorithm The adaptive residual generator scheme given in (Ding et al. [2009a]) is inspired by the adaptive observer schemes proposed by Zhang [2002]. It consists of three sub-systems: Residual generator z (k + 1) = Az z (k) + Q (u(k), y(k)) (k) +V (k + 1) + 1) (k (k) T d (k) T T z cz z (k) r(k) = u (k) y (k) g T (k)

S3: Solve (50) and (51) S4: Generate an m-dimensional residual vector S5: Compute (52) 5. THE ADAPTIVE SCHEME Among the existing adaptive schemes, the recursive technique has been well developed and widely applied to the standard data-driven methods like PCA, PLS. Since the early work by Helland et al. [1992], Qin [1998], numerous recursive schemes have been reported. The recent research focus is mainly on the recursive computation of SVD which is used both in the PCA and PLS. Li et al. [2000] proposed to apply the rank-one modication technique for a recursive updating of the covariance matrix and its SVD used in the PCA. It is remarkable that the key step in Algorithm 3.3 and all those associated algorithms is
845

(61) (62)

8th IFAC Symposium on Advanced Control of Chemical Processes Furama Riverfront, Singapore, July 10-13, 2012

Auxiliary lter (63) V (k + 1) = Az V (k) + Q(u(k), y(k)) (k) = cz V (k) 0 uT (k) 0 y T (k) (64) T dz 0 uT (k) 0 y T (k) = uT (k) y T (k) gT Parameter estimator + 1) = (k)T (k)r(k) + (k) (65) (k , 0, 0 < < 2 (66) (k) = + (k)T (k) Algorithm 5.1: Adaptive residual generator S0: Set the initial values k = 0, z (0), (0), V (0) = 0 (0) = 0, r(0) = g (0)y(0) cz z (0) dz (0)u(0) S1: Compute V (k + 1) according to (63) S2: Compute + 1) according to (65) (k S3: Compute z (k + 1) according to (61) S4: Increase k by one, receive y(k), u(k) S5: Compute r(k), (k) according to (62) and (64) 5.3 Stability and exponential convergence The major advantage of applying the above adaptive technique is that the convergency of the parameter estimation and the whole system stability are guaranteed. In order to demonstrate it, let (k) = z (k) V (k) (k) with z (k) = z(k) z (k), (k) = (k)

to identify faults in a dynamic process. With the aid of the parity space identied using Algorithm 3.3, the algebraic fault isolation methods e.g. given in (Ding [2008]) can also be used. It is remarkable that these approaches have been developed on the assumption of an available fault model. Alternatively, Ding et al. [2009b], Wang et al. [2011] have proposed data-driven design schemes for the isolation of sensor and actuator faults in dynamic systems. In practice, data collected during faulty process operations are seldom available or at least often not available in a well synchronized form with a detailed fault description. It makes a data-driven (fault) modelling or data-driven fault isolation and identication very dicult. Data-driven fault isolation and identication is a changing topic both in the application and research domains. Due to the space limitation, it will not be addressed in this paper. 7. CONCLUSIONS In this paper, we have presented a number of data-driven methods for the design of FDI systems. The major focus has been on the data-driven design of observer-based FD systems. Most of the data-driven design algorithms presented in this paper have been realized in the software form and studied on dierent benchmark processes like Tennessee Eastman (TE) process (Russell et al. [2000]) and continuous stirred tank heater (CSTH) (Thornhill et al. [2008]). Some of them have also been tested on the real laboratory systems including Three-Tank-System and CSTH that are available at AKS. In addition, they have been applied to the identication and monitoring of process key variables and to the data-driven design of feedback control systems. Due to the space limitation, the achieved results cannot be included in this paper. It is remarkable that the data-driven design technique for dynamic FDI systems is still in its early stage. For a successful application in practice as an established technique, great eorts should be made to address, for instance, the fault isolation and identication issues as mentioned previously and the nonlinear and non-Gaussian issues. The latter is very important and useful for improving the FDI performance during the process operation. The successful methods based on machine learning technique show promising results (Yu and Qin [2008, 2009]). Acknowledgement: The author would like to thank S. Yin and Y. Wang for their valuable technical support, and the reviewers for their valuable comments. REFERENCES C. F. Alcala and S. J. Qin. Reconstruction-based contribution for process monitoring. Automatica, 45:15931600, 2009. K. J. Astrm and B. Wittenmark. Adaptive Control. Addison-Wesley Publishing Company, 1995. M. Blanke, M. Kinnaert, J. Lunze, and M. Staroswiecki. Diagnosis and Fault-Tolerant Control. Springer, 2003. J. Chen and R. J. Patton. Robust Model-Based Fault Diagnosis for Dynamic Systems. Kluwer Academic Publishers, Boston, 1999. L.H. Chiang, EL: Russell, and R. D. Braatz. Fault Detection and Diagnosis in Industrial Systems. Springer, London, Berlin, 2001.

Then r(k) can be re-written into dT (k) T T z cz z (k) r(k) = u (k) y (k) g T (k) = gy(k) cz z (k) dz (k)u(k) (g g (k)) y(k) dT (k) T T z = cz z (k) u (k) y (k) g T (k) (67) = cz (k) + (k)(k) dz (k) = dz dz (k), g (k) = g g (k) Moreover, + 1) = (k)(k) + I (k)T (k)(k) (k) (k (68)

(k) = (k)T (k)cz (69) (k + 1) = Az (k) (70) Based on (67)-(70) and the results given in (Astrm and Wittenmark [1995], Zhang [2002]), Ding et al. [2009a] have proven that limk r(k) = 0. 6. ON FAULT ISOLATION AND IDENTIFICATION In practice of (statistical) process monitoring, contribution plots are widely applied as a standard PCA-based fault isolation and identication technique (Chiang et al. [2001]). Recently, Qin [2009], Alcala and Qin [2009], Li et al. [2010] have proposed to estimate/identify the faults in a statistical process by means of reconstruction. In (Dong [2009]), a data-driven approach has been developed
846

8th IFAC Symposium on Advanced Control of Chemical Processes Furama Riverfront, Singapore, July 10-13, 2012

S. X. Ding. Model-Based Fault Diagnosis Techniques - Design Schemes, Algorithms, and Tools. SpringerVerlag, 2008. S. X. Ding, S. Yin, P. Zhang, E.L. Ding, and A. Naik. An approach to data-driven adaptive residual generator design and implementation. Proc. of IFAC Symp. SAFEPROCESS, 2009a. S. X. Ding, P. Zhang, A. Naik, E.L. Ding, and B. Huang. Subspace method aided data-driven design of fault detection and isolation systems. Journal of Process Control, 19:14961510, 2009b. S. X. Ding, S. Yin, Y. Wang, Y. Wang, Y. Yang, and B. Ni. Data-driven design of observers and its applications. Preprint of the 18th IFAC World Congress, pages 1144111446, 2011. S. X. Ding, Y. Wang, S. Yin, P. Zhang, Y. Yang, and E.L. Ding. Data-driven design of fault-tolerant control systems. Proc. of the 8th IFAC Sypm. SAFEPROCESS, 2012. X. Ding and P. M. Frank. Fault detection via factorization approach. Syst. and Contr. Letters, 14:431436, 1990. J. Dong. Data Driven Fault Tolerant Control: A Subspace Approach. PhD thesis, Technische Universieit Delft, 2009. J.F. Dong and M. Verhaegen. Subspace based fault detection and identication for lti systems. the 7th IFAC Symposium on Fault Detection, Supervision and Safety of Technical Processes, pages 330335, 2009. J. J. Gertler. Fault Detection and Diagnosis in Engineering Systems. Marcel Dekker, 1998. F. Gustafsson. Adaptive Filtering and Change Detection. John Wiley and Sons, LTD, 2000. D. M. Helland, H. E. Bernstein, O.S. Borgen, and H. Martens. Recursive algorithm for partial least squares regression. Chemometrics Intell. Lab. Syst., 14: 129137, 1992. B. Huang and R. Kadali. Dynamic Modelling, Predictive Control and Performance Monitoring, a Data-Driven Subspace Approach. Springer-Verlag, London, 2008. I. Hwang, S. Kim, Y. Kim, and C.E. Seah. A survey of fault detection, isolation, and reconguration methods. IEEE Trans. Contr. Syst. Tech., 18:636653, 2010. R. Isermann. Fault Diagnosis Systems. Springer-Verlag, 2006. B. Kulcsar, J.F. Dong, and M. Verhaegen. Model-free fault tolerant control approach for linear parameter varying systems. the 7th IFAC Symposium on Fault Detection, Supervision and Safety of Technical Processes, pages 876881, 2009. G. Li, S. J. Qin, Y. Ji, and D. Zhou. Reconstruction based fault diagnosis for continous processes. Control Engineering Practice, 2010. W. Li and S. J. Qin. Consistent dynamic PCA based on errors-in-variables subspace identication. Journal of Process Control, 11:661678, 2001. W. Li, H. H. Yue, S. Valle-Cervantes, and S. J. Qin. Recursive PCA for adaptive process monitoring. Journal of process control, 10:471486, 2000. R. Mangoubi. Robust Estimation and Failure Detection. Springer, 1998. R. Mangoubi, M. Desai, A. Edelmayer, and P. Sammak. Robust detection and estimation in dynamic systems and statistical signal processing: Intersection, parallel

paths and applications. European journal of control, 15: 348369, 2009. A. S. Naik, S. Yin, S. X. Ding, and P. Zhang. Recursive identication algorithms to design fault detection systems. Journal of process control, 20:957965, 2010. R. J. Patton, P. M. Frank, and R. N. Clark (Eds.). Issues of Fault Diagnosis for Dynamic Systems. Springer, 2000. S. J. Qin. An overview of subspace identication. Computers and Chemical Engineering, 30:15021513, 2006. S. J. Qin. Recursive PLS algorithms for adaptive data modeling. Computers chem. Eng., 22:503514, 1998. S. J. Qin. Data-driven fault detection and diagnosis for complex industrial processes. Proc. of IFAC SAFEPROCESS Symp., pages 11151125, 2009. S. J. Qin and W. Li. Detection and identication of faulty sensors in dynamic processes. AIChE Journal, 47:1581 1593, 2001. S.J. Qin. Statistical process monitoring: Basics and beyond. Journal of Chemometrics, 17:480502, 2003. E.L. Russell, L.H. Chiang, and R.D. Braatz. DataDriven Techniques for Fault Detection and Diagnosis in Chemical Processes. Springer-Verlag, London, 2000. S. Simani, S. Fantuzzi, and R. J. Patton. Model-Based Fault Diagnosis in Dynamic Systems Using Identication Techniques. Springer-Verlag, 2003. N. Thornhill, S. Patwardhan, and S. Shah. A continous stirred tank heater simulation model with applications. J. of Process contr., 18, 2008. V. Venkatasubramanian, R. Rengaswamy, S.N. Kavuri, and K. Yin. Review of process fault detection and diagnosis part III: Quantitative model-based methods. Computers and Chemical Engineering, 27:327346, 2003a. V. Venkatasubramanian, R. Rengaswamy, K. Yin, and S.N. Kavuri. Review of process fault detection and diagnosis part i: Quantitative model-based methods. Computers and Chemical Engineering, 27:293311, 2003b. J. Wang and S. J. Qin. A new subspace identication approach based on principle component analysis. Journal of process control, 12:841855, 2002. X. Wang, U. Kruger, and G. W. Irwin. Process monitoring approach using fast moving window PCA. Ind. Eng. Chem. Res., 44:56915702, 2005. Y. Wang, G. Ma, S. X. Ding, and C. Li. Subspace aided data-driven design of robust fault detection and isolation systems. Automatica, 47:24742480, 2011. J. Yu and S.J. Qin. Multimode process monitoring with bayersian inference-based nite gaussian mixture models. AICHe Journal, 54:18111829, 2008. J. Yu and S.J. Qin. Multiway gaussian mixture model based multiphase batch process monitoring. Ind. Eng. Chem. Res., 48:85858594, 2009. P. Zhang and S. X. Ding. On fault detection in linear discrete-time, periodic, and sampled-data systems (survey). Journal of control science and engineering, pages 118, 2008. Q. Zhang. Adaptive observer for multiple-input-multipleoutput (MIMO) linear time-varying systems. IEEE Trans. Autom. Contr., 47:525529, 2002. K. Zhou, J.C. Doyle, and K. Glover. Robust and Optimal Control. Prentice-Hall, Upper Saddle River, New Jersey, 1996.

847