I.

INTRODUCTION

The growing demand for reliability,

Integrated Active Fault-Tolerant maintainability, and survivability in safety critical
systems has drawn significant research in fault
Control Using IMM Approach detection and diagnosis (FDD) [4, 9, 10, 29, 34] over
the last two decades. Such efforts have led to the
development of many FDD techniques, unfortunately,
little attention has been paid to the related problem,
i.e., reconfigurable fault-tolerant control, until the
YOUMIN ZHANG, Member, IEEE mid-1980s [8, 17]. Only recently, the fault-tolerant
control problem has begun to receive more attention
JIN JIANG, Senior Member, IEEE
The University of Western Ontario [3, 14, 15, 27, 28, 35, 37].
A fault-tolerant control system (FTCS) is a control
system that possesses the ability to accommodate
system failures automatically, and to maintain
An integrated fault detection, diagnosis, and reconfigurable overall system stability and acceptable performance.
control scheme based on interacting multiple model (IMM)
Generally speaking, fault-tolerant control systems
can be classified as passive [15, 33] and active [17].
approach is proposed. Fault detection and diagnosis (FDD)
An active fault-tolerant control system compensates
is carried out using an IMM estimator. An eigenstructure
for faults either by selecting a precomputed control
assignment (EA) technique is used for reconfigurable
law (projection-based approaches) [19, 21] or by
feedback control law design. To achieve steady-state tracking, synthesizing a new control strategy on-line (on-line
reconfigurable feedforward controllers are also synthesized automatic controller redesign approaches) [17].
using input weighting approach. The developed scheme can deal Both approaches use FDD schemes to detect and
with not only actuator and sensor faults, but also failures in diagnose faults and to activate control reconfiguration
system components. To achieve fast and reliable fault detection, mechanisms. Such systems are also known as
diagnosis, and controller reconfiguration, new fault diagnosis and reconfigurable [11, 13, 19], restructurable [12, 17],
controller reconfiguration mechanisms have been developed by or self-repairing control systems [8]. Typically, an
a suitable combination of the information provided by the mode active FTCS consists of three parts: a reconfigurable
probabilities from the IMM algorithm and an index related to
controller, an FDD scheme, and a control law
the closed-loop system performance. The proposed approach is
reconfiguration mechanism. Key issues are how
to design: 1) a robust reconfigurable controller,
evaluated using an aircraft example, and excellent results have
2) an FDD scheme with high sensitivity to faults
been obtained.
and robustness to model uncertainties and external
disturbances, and 3) a reconfiguration mechanism
which can organize the reconfigured controller in such
a way that the prefault system performance can be
recovered to the maximum extent. This work focuses
on the development of a new integrated approach to
active fault-tolerant control system design.
In general, the existing active fault-tolerant control
system design methods can be categorized based on
the following approaches: linear quadratic regulator
(LQR) [17, 21]; eigenstructure assignment (EA)
[13]; multiple model (MM) [7, 19, 24, 31]; adaptive
Manuscript received December 20, 1999; revised and resubmitted control [6, 31]; pseudo-inverse [11]; model following
December 5, 2000; released for publication July 2, 2001. [12, 22]; and neural networks [23]. Among these,
IEEE Log No. T-AES/37/4/10988. the study on MM-based reconfigurable control
has drawn increasing attention recently [7, 19, 24,
Refereeing of this contribution was handled by A. Ray.
30, 31]. One of the approaches developed in [19],
This work was supported by the Natural Sciences and Engineering known as MM adaptive estimator/MM adaptive
Research Council of Canada. control (MMAE/MMAC), synthesizes each of the
Authors’ address: Department of Electrical and Computer reconfigurable controllers using LQR technique. In
Engineering, The University of Western Ontario, London, Ontario this approach, the system under the presumed failure
N6A 5B9, Canada, E-mail: (fymzhang, [email protected]). modes is represented by a set of models, and a bank
of Kalman filters is used to estimate the states of
0018-9251/01/$17.00 °
c 2001 IEEE the system based on the presumed failure modes.

IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 37, NO. 4 OCTOBER 2001 1221
Innovation sequence of each Kalman filter is used to
calculate the probabilities of individual failures, and
the aggregated state estimate and the control input
are the probability-weighted average of the signals
from each model. However, in the above scheme,
there is no interaction among Kalman filters. Such an
approach is effective in dealing with problems with an
unknown structure or parameter but without structural Fig. 1. General structure of proposed FDD and reconfigurable
or parametric changes. Obviously, the problem of fault control scheme.
diagnosis and reconfigurable control does not fit well
into such a framework because, in general, the system implementation issues. In Section V, performance
structure or parameters do change as a component or evaluation of the proposed scheme for an aircraft is
subsystem fails. To overcome this weakness and to presented. Conclusions are drawn in Section VI.
make MM algorithms more suitable for the current
problem, an interacting MM (IMM) FDD approach
has been proposed recently [36]. In this approach, the II. REPRESENTATION OF FAILURES IN SYSTEM
occurrence or the recovery of a failure in a dynamic COMPONENTS, ACTUATORS AND SENSORS
system has been explicitly modeled as a finite-state
A. Hidden Markov Chain Model for Systems with
Markov chain with known transition probabilities.
Since changes of the system are explicitly considered Failures
and effectively handled in the IMM algorithm, it has Suppose that a discrete-time process which
been shown that the IMM algorithm is much superior represents the possible system structural/parametric
in performing FDD than other existing methods [36]. changes due to failures is represented by a first-order
Typically, faults can occur in system components, Markov chain with state m(k) taking values in a finite
actuators, and sensors. Due to the inherent difficulties set S = 1, 2, : : : , s. At each time step, the transition
in considering all types of failures simultaneously,
probabilities of the chain can be defined by
existing single-model-based approaches consider these
faults separately. Using the MM approach, however, ¼ij (k) = Pfm(k) = j j m(k ¡ 1) = ig 8 i, j 2 S
the system component, actuator, and sensor faults
can be dealt with simultaneously. In this paper, FDD (1)
and reconfigurable control in the presence of all three with X
types of faults have been addressed. ¼ij (k) = 1, i = 1, : : : , s (2)
The objective of this work is to extend the j2S
proposed IMM-based FDD approach to the design
of an integrated FDD and reconfigurable control. In where Pf¢g denotes the probability; m(k) is the
the proposed scheme, the IMM estimator is utilized to discrete-valued modal state (i.e., the indicator of
provide the information on FDD, as well as the state the normal or the fault mode) at time k; ¼ij is the
estimation. The reconfiguration mechanism is based transition probability from the mode i to the mode
on a suitable combination of the information from the j; the event that mj is in effect at time k is denoted as
¢
mode probability in the IMM algorithm and a system mj (k) =fm(k) = jg.
performance index. Since the stability and dynamic
The model of the system with potential failures
behavior of the closed-loop system can be described
can be expressed as
by its eigenvalues and corresponding eigenvectors, i.e.
eigenstructure, recovery of the dynamic performance x(k + 1) = F(k, m(k + 1))x(k) + G(k, m(k + 1))u(k)
after the faults can be achieved via assigning the
eigenstructure of the reconfigured system as close to + ¡ (k, m(k + 1))»(k, m(k + 1)) (3)
that of the prefault system as possible if the redundant z(k) = H(k, m(k))x(k) + ´(k, m(k)) (4)
control is available [15]. A set of reconfigurable
control laws is synthesized using an EA technique. where x 2 Rn is the state vector; z 2 Rp is the
In addition, to achieve zero steady-state tracking measurement vector; u 2 Rl is the control input
even in the presence of faults, a set of feedforward vector; »(k) 2 Rn and ´(k) 2 Rp are independent
control laws is also designed using an input weighting discrete-time random process with mean »(k) and ´(k)
technique [13]. Fig. 1 depicts the structure of the and covariances Q(k) and R(k), representing system
scheme. and measurement noises, respectively. It is assumed
The paper is organized as follows. In Section II, that the initial state has a mean x0 and a covariance
modeling of system component, actuator, and sensor P 0 , and they are independent from »(k) and ´(k).
faults is presented. Based on IMM estimator, FDD The system (3)—(4) is known as a “jump linear
mechanisms are outlined in Section III. Section IV system” [16, 36]. It can be seen from (4) that the
discusses the reconfigurable controller design and state observations are noisy and mode dependent.

1222 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 37, NO. 4 OCTOBER 2001
Therefore, the mode information is imbedded in the of applications, e.g. maneuvering target tracking
measurement sequence. In other words, the system [16, 20], and FDD [36].
mode sequence is an indirectly observed (or hidden) The IMM algorithm is a recursive estimator with
Markov chain. the following steps in each iteration:
1) interaction of the model-conditional estimates,
B. Multiple-Model Representation of System Failures 2) model-conditional filtering,
3) mode probability update,
Based on the system model (3)—(4), it is possible 4) estimates combination.
to represent different failures in the system. Even
though the MM method is a suitable choice for FDD In the first step, the input to the filter matched to
and reconfigurable control, the way to choose a good a certain mode is obtained by mixing the estimates
model set to represent all possible fault situations of all filters from the previous iteration under the
(including fault types, number of faults, and fault assumption that this particular mode is in effect at
magnitude) is crucial for the success in FDD and the present time; a bank of filters corresponding to
reconfigurable control design. Generally speaking, different models is calculated in parallel in the second
the performance of an MM algorithm depends on the step; mode probability is then updated based on the
model set used. model-conditional innovations and the likelihood
Assume that a set of N models has been used functions; finally, the aggregated state estimate is
to represent the normal and N ¡ 1 different failure obtained as a probability-weighted sum of the updated
situations, then the system in (3)—(4) can further be state estimates from all the filters.
represented as The probability of the mode in effect plays a key
role in determining the weights in the combination of
x(k + 1) = (F(k) + ¢Fj (k))x(k) state estimates and covariances for aggregated state
estimate. It should be pointed out that in comparison
+ (G(k) + ¢Gj (k))u(k) + ¡j (k)»j (k)
with the existing noninteracting MM algorithms,
= Fj (k)x(k) + Gj (k)u(k) + ¡j (k)»j (k) (5) the step 1 is unique for the IMM estimator. It is
because of this mixture of the estimates that makes
z(k) = (H(k) + ¢Hj (k))x(k) + ´j (k) the estimation for the state and identification for the
= Hj (k)x(k) + ´j (k) j = 1, : : : , N (6) system mode more responsive to the system changes,
thus leads to significantly better FDD performance
where ¢Fj (k), ¢Gj (k) and ¢Hj (k) (j = 2, : : : , N) [36].
represent the fault-induced changes in the system Fig. 2 shows a block diagram of the IMM
components, actuators, and sensors, respectively. They estimator for FDD. From this diagram, inherent
should be null matrices when j = 1. The subscript j relations of the above four steps can be observed
denote quantities pertaining to the model mj 2 M. clearly.
M = [m1 , m2 , : : : , mN ] is a set of all system models
representing the normal system and the system with B. Fault Detection and Diagnosis Scheme
all considered faults. Matrices Fj (k), Gj (k), and Hj (k)
correspond to the jth postfault models of the system. In active FTCS, timely and correct detection and
The objective of the integrated FDD and diagnosis of a fault is crucial for good performance.
reconfigurable controller is to estimate the state and Using the IMM estimator, it is effective to use the
to identify the system mode in effect from a sequence mode probabilities to provide an indication of the
of noise-corrupted observations, and then to recover mode in effect at a given time. Hence, it can be used
the performance of the prefault system by selecting as an index for FDD. The fault detection decision can
an appropriate controller among a set of precomputed be made by the following rule:
½
controllers. > ¹T ) Hj : fault j occurred
¹j (k + 1) = max ¹i (k + 1)
i · ¹T ) H1 : no fault
III. FAULT DETECTION, DIAGNOSIS AND (7)
CONTROLLER RECONFIGURATION where ¹T , 0 < ¹T < 1, is the detection threshold.
MECHANISMS A complete cycle of the IMM-based FDD scheme
with Kalman filters as its mode-matching filters is
A. State Estimation and Fault Diagnosis Using IMM summarized in Table I.
Estimator
C. Integrated Fault Detection, Diagnosis, and
The IMM estimator [5] is generally considered to
Controller Reconfiguration Mechanisms
be one of the most cost-effective schemes for state
estimation involving both continuous and discrete It has been shown in [36] that, in comparison with
states [16]. It has been successfully used in a number noninteracting MM approaches, the IMM-based FDD

ZHANG & JIANG: INTEGRATED ACTIVE FAULT-TOLERANT CONTROL USING IMM APPROACH 1223
 Fig. 2. Block diagram of IMM-based FDD approach.

approach provides a faster and more reliable detection where "j is the 2-norm of the filter residual vector
and diagnosis for system failures. However, these corresponding to the jth model, defined by
conclusions were drawn only for open-loop systems.
When a control action is introduced and combined "j (k + 1) = kz(k + 1) ¡ Hj (k + 1)x̂j (k + 1 j k + 1)k2
with an MM estimator to form a closed-loop system, (9)
the effect of the feedback will force the residuals from
different models to be similar and this affects the where z(k + 1) is the measurement at time (k + 1),
discrimination property of the filter. Such a situation x̂j (k + 1 j k + 1) denotes the estimation of the state of
is even more serious in reconfigurable control design the closed-loop system at time (k + 1) for the model
because the objective of a reconfigurable control is to j. c1 ¸ 0 and c2 ¸ 0 are constants. They determine
make the performance of the reconfigured system as the relative weights given to the instantaneous and
close to that of the prefault system as possible. Such a accumulative measures. To consider the time-varying
design principle inherently leads to similar residuals in nature of the problem and to make the performance
each filter and may make the FDD provided only by measure more responsive to fault-induced changes,
the mode probability less reliable. a moving window of length M has been used in the
To overcome such a drawback, various heuristic second term in (8).
techniques have been investigated, e.g. the addition Based on the above analysis, for reliable fault
of probing signals, alternative computation of diagnosis and control system reconfiguration, the
probabilities, bounded conditional probabilities, following condition can be used to activate the
Kalman filter retuning, scalar penalty increase, reconfiguration process:
8
probability smoothing, and increased residual > ¹j (k + 1) = max ¹i (k + 1) > ¹T and
>
> i
propagation [2, 18, 19, 32]. These techniques >
< j = arg min J (k + 1) ) H : reconfiguration for fault j
may enhance the performance of FDD and lead to i i j
:
subsequent improvement in control to some extent. >
> ¹j (k + 1) = max ¹i (k + 1) · ¹T
>
> i
However, preliminary investigation has shown that :
) H1 : no reconfiguration
these techniques are not effective in the current
situation. This motivates us to develop a new (10)
method for more reliable FDD and reconfigurable
control. D. Structure of Integrated FDD and Reconfigurable
Inspired by the MMs’ switching and tuning Control
approach proposed in [25], a performance index is
used to provide additional information for reliable A block diagram of the combined IMM-based
fault diagnosis and controller reconfiguration. This FDD and the reconfigurable control scheme is shown
index is calculated as in Fig. 3. Based on the on-line FDD information, the
k+1
X real-time reconfiguration can be carried out.
Jj (k + 1) = c1 "2j (k + 1) + c2 "2j (i) (8) Similar to the aggregated state estimation in the
i=k¡M+2 IMM estimator, the aggregated control signal, u(k),

1224 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 37, NO. 4 OCTOBER 2001
 TABLE I
One Cycle of IMM-Based FDD Scheme

where ui (k) = Ki x̂i (k j k) is the feedback control law

designed for the ith model in the model set, x̂i (k j k)
is the corresponding state estimate obtained by the
ith Kalman filter. Ki and Li are the feedback and
the feedforward control gains corresponding to the
ith model. ¹i (k) denotes the corresponding mode
probability for the ith model.
Fig. 3. Block diagram of IMM-based FDD and reconfigurable
control. IV. RECONFIGURABLE CONTROLLER DESIGN AND
ON-LINE CONTROLLER RECONFIGURATION
can also be calculated as the probability-weighted
average of each controller output, which is given by A. Dynamic Performance Recovery-Feedback
0
u(k) = r (k) ¡ uf (k) Controller Design

N
X LQR and EA are among the most popular
= fLi r(k) ¡ ui (k)g ¢ ¹i (k) (11) controller design techniques for multiinput and
i=1 multioutput systems. The advantage of EA is that

ZHANG & JIANG: INTEGRATED ACTIVE FAULT-TOLERANT CONTROL USING IMM APPROACH 1225
when the performance specifications are given in as the desired eigenvector for vij : The best possible
terms of system eigenstructure, the eigenstructure closed-loop system eigenvector can be obtained by
can be achieved exactly for the stability and desired the projection of the desired eigenvector onto the
dynamic performance. The condition for the exact subspace spanned by the columns of (¸ij I ¡ Fj )¡1 Gj :
assignment is that there is a sufficient number of In the context of reconfigurable control system design,
actuators and measurements available and that the best choice of vij can be obtained by projecting the
the desired eigenvectors reside in the achievable corresponding vi1 onto Eji orthogonally, where Eji and
subspace [1]. The limitation of EA is that the a new vector wij can be defined as
system performance is not optimal in any measure.
The LQR could be used to optimize the controller Eji = (¸ij I ¡ Fj )¡1 Gj (16)
design by minimizing a quadratic function of the
system response and the control energy. In general, wij = Kj vij : (17)
the LQR-based control design will guarantee the
Consequently, (15) can be rewritten as
closed-loop system stability and certain degree
of robustness, but may not easily achieve the vij = Eji wij : (18)
specific system performance due to ambiguities
in the selection of the weighting matrices. For the The desired eigenvector vij in the achievable
above reasons, the EA is used for the design of the subspace can be found by the following least-squares
reconfigurable controller while LQR is used for the minimization
design of the nominal controller. min Ji (vij ) = minf(vij ¡ vi1 )T Wji (vij ¡ vi1 )g
1) Assignment for the Best Eigenvalues and
Eigenvectors: Suppose that the dynamics of the = minf(Eji wij ¡ vi1 )T Wji (Eji wij ¡ vi1 )g,
system have undergone some changes due to faults
in system components, actuators, and sensors, and i = 1, : : : , n, j = 2, : : : , N (19)
the system models with normal and different faulty and
conditions have become vij = Eji (EjiT WjiT Wji Eji )¡1 EjiT WjiT vi1 (20)
x(k + 1) = Fj x(k) + Gj u(k), j = 1, : : : , N: (12) where Wji 2 Rn£n is a positive definite weighting
matrix. Suggestions about how to choose weighting
The aim of reconfigurable control system design is
matrix Wji are given in [13].
to synthesize a set of new feedback gain matrices so
2) Computation of the Reconfigurable Control
that the closed-loop eigenvalues of the reconfigured
Gain Matrix: Consider the following closed-loop
system are the same as those of the prefault system,
reconfigured system equation
i.e.,
¸ij = ¸(Fj + Gj Kj ) = ¸i = ¸(F1 + G1 K1 ), xk+1 = (Fj + Gj Kj )xk : (21)

i = 1, : : : , n, j = 2, : : : , N: (13) To simplify the procedure in calculating the matrix

Kj , a linear transformation matrix
where K1 denotes a control gain matrix designed for .
the nominal (fault-free) system, and Kj , j = 2, : : : , N, −j = [Gj .. £j ] 2 Rn£n (22)
represents the new feedback control gain matrices
under different fault conditions. ¸(¢) denotes the is chosen, where £j 2 Rn£(n¡l) is an arbitrary matrix
eigenvalues of the system. such that rank(−j ) = n.
The closed-loop system eigenvectors of the Applying the linear transformation −j to (21), a
reconfigured system, fvij , i = 1, : : : , n, j = 2, : : : , Ng, new set of state variables xk 2 Rn can be obtained
with the feedback gain matrix Kj will satisfy
xk = −j¡1 xk : (23)
(Fj + Gj Kj )vij = ¸ij vij (14) Thus (21) is transformed to
or xk+1 = (F j + Gj Kj −j )xk (24)
vij = (¸ij I ¡ Fj )¡1 Gj Kj vij : (15) where
· ¸
Then, the objective of the reconfigurable control Il
Fj = −j¡1 Fj −j , Gj = −j¡1 Gj = :
system is to synthesize a feedback gain matrix Kj 0
such that the eigenvectors of reconfigured closed-loop
system vij is as close to the corresponding eigenvectors The corresponding eigenvectors under this
transformation are related by
of the prefault system vi1 as possible. Because of the
· i¸
variations in system dynamics, in general, vij does not s
vij = −j¡1 vij = i :
lie in the same subspace as vi1 , which may be viewed g

1226 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 37, NO. 4 OCTOBER 2001
 Clearly, the eigenvalues, eigenvectors and system pair of self-conjugate eigenvalues, the corresponding
matrices satisfy rows and columns in the ©j matrix need to be
assigned by the transformation matrix
(F j + Gj Kj −j )vij = ¸ij vij , i = 1, : : : , n, j = 2, : : : , N: · ¸
1=2 ¡j1=2
(25) ©0 =
1=2 j1=2
Equation (25) can be rearranged as follows
in place of unity matrix I.
(¸ij I ¡ F j )vij = Gj Kj −j vij , i = 1, : : : , n, j = 2, : : : , N: From (30) the desired feedback gain matrix can be
(26) calculated as
By exploiting the special structure of Gj , we can Kj = (Sj ¡ F 1 Vj )(−j Vj )¡1 , j = 2, : : : , N (33)
rewrite (26) as
" i #· ¸ · ¸ or from (32), the desired feedback gain matrix for
· i¸
¸j Il ¡ F 11 ¡F 12 si Il s the case of complex eigenvalues can be obtained
i
= K −
j j i as
¡F 21 i
¸j In¡l ¡ F 22 g 0 g
Kj = (Sj ©j ¡ F 1 Vj ©j )(−j Vj ©j )¡1 , j = 2, : : : , N:
(27)
where " #
F 11 F 12 (34)
Fj = = −j¡1 Fj −j :
F 21 F 22
B. Steady-State Performance Recovery-Feedforward
The first matrix equation in the partitioned form in Controller Design
(27) can be written as
Even though the dynamic performance of the
· i¸
i i i s prefault system can be recovered to the maximum
(¸j Il ¡ F 11 )s ¡ F 22 g = Kj −j i extent with EA, it is also important to consider the
g
steady-state performance. This can be accomplished
i = 1, : : : , n, j = 2, : : : , N. (28) by a set of properly designed feedforward control
gains Lj , j = 1, : : : , N [13]. These gains can be
Further, by letting F 1 = [F 11 F 12 ], (28) becomes
calculated as follows using a property of z-transform.
[F 1 + Kj −j ]vij = ¸ij si , i = 1, : : : , n, j = 2, : : : , N Suppose that the original reference input r(k) has
been scaled by a set of input weighting matrices Lj ,
(29) as r0 (k) for the jth model in the model set, then
or, in a compact form
we have
[F 1 + Kj −j ]Vj = Sj (30) r0j (k) = Lj ¢ r(k), j = 1, : : : , N: (35)
where Vj = [v1j v2j : : : vnj ] 2 Rn£n , and Sj = The steady-state output of the stable closed-loop
[¸1j s1 ¸2j s2 : : : ¸nj sn ] 2 Rl£n . system subject to a unit step input can be calculated
It should be noted that Vj and Sj are often using the final value theorem in z-transform:
complex. To alleviate the need for complex arithmetic, y(1) = lim y(k) = lim Hc x(k)
a transformation is needed to transform Vj and Sj k!1 k!1
¤ i
to real matrices. Assume that ¸ij = (¸i+1 j ) and vj = = lim(z ¡ 1)©j (z)Lj R(z)
i+1 ¤
(vj ) , and assuming all remaining eigenvalues are z!1
real, we can define a transformation matrix ½ ¾
¡1
2 3 = lim Hc (zI ¡ Fj + Gj Kj ) Gj Lj
I 0 0 0 z!1
6 0 1=2 ¡j1=2 0 7
6 7 = ªj Lj j = 1, : : : , N (36)
©j = 6 7 2 Rn£n : (31)
4 0 1=2 j1=2 0 5
where Kj , j = 1, : : : , N, represent the jth feedback
0 0 0 I control gains. Hc is a matrix such that the system
output y(k) = Hc x(k) tracks the reference input
Multiplying both sides of (30) by the
r(k). ªj = flimz!1 Hc (zI ¡ Fj + Gj Kj )¡1 Gj g is the
transformation matrix ©j , i.e.,
steady-state gain of the system before scaling.
The steady-state performance recovery problem
[F 1 + Kj −j ]Vj ©j = Sj ©j (32)
can then be stated as choosing appropriate matrices
will transform Vj and Sj to real matrices and at the Lj 2 Rl£l to minimize J(Lj )
same time not alter the calculation of the feedback min J(Lj ) = min kI ¡ ªj Lj k, j = 1, : : : , N: (37)
gain matrix. Note that for the case of more than one Lj Lj

ZHANG & JIANG: INTEGRATED ACTIVE FAULT-TOLERANT CONTROL USING IMM APPROACH 1227
If ªj is invertible, then V. PERFORMANCE EVALUATION OF THE
½ ¾¡1 PROPOSED SCHEME FOR AN AIRCRAFT
¡1
Lj = lim Hc (zI ¡ Fj + Gj Kj ) Gj , j = 1, : : : , N: The effectiveness of the proposed scheme is
z!1
(38) demonstrated in this section through a longitudinal
It is important to point out that the feedforward vertical takeoff and landing (VTOL) aircraft model
control gains Lj are dependent upon the feedback [26].
control gain matrices Kj . It should be noted that if the
number of system output is less than or equal to the A. Aircraft Model
number of system input, i.e., m · l, the steady-state The linearized model of the aircraft can be
performance of the original system can be recovered described by
completely. If m > l, the chosen Lj is to minimize
J(Lj ) in a Frobenius-norm sense [13]. _ = Ax(t) + Bu(t) + »(t)
x(t)
(40)
z(t) = Cx(t) + ´(t)
C. Reconfigurable Control Signal Generation Strategies
where x = [Vh Vv q µ]T , u = [±c ±l ]T . The states and the
Even though in the above IMM scheme, there are inputs are: horizontal velocity Vh , vertical velocity Vv ,
a set of models and the corresponding feedback and pitch rate q, and pitch angle µ; collective pitch control
feedforward gains for each failure mode, there is still ±c , and longitudinal cyclic pitch control ±l . The model
a single control signal at any given time. In general, parameters are given as follows
there are two ways to generate this signal: one is 2 3
¡0:0366 0:0271 0:0188 ¡0:4555
based on a Bayesian scheme and the other is by the
6 0:0482 ¡1:01 0:0024 ¡4:0208 7
maximum a posteriori (MAP) approach [19, 32]. 6 7
A=6 7,
In the former, the control signal is obtained as a 4 0:1002 0:3681 ¡0:707 1:420 5
probability-weighted average of the signals generated
0:0 0:0 1:0 0:0
from all the models. The advantage of this approach 2 3 2 3
is that it is able to reduce the effect of incorrect model 0:4422 0:1761 1 0 0 0
selection during the early stage of reconfiguration. 6 3:5446 ¡7:5922 7 60 1 0 07
6 7 6 7
However, once the failure mode has been identified B=6 7, C=6 7:
4 ¡5:52 4:49 5 40 0 1 05
correctly, there is no need to continuously use the
probability-weighted approach, as the non-zero 0:0 0:0 0 1 1 1
probabilities (due to noise) can have adverse effects The zero hold equivalent system can be
on the control signal generation. represented by
In the MAP approach, the control signal is selected
from the model which has the highest probability. The x(k + 1) = Fx(k) + Gu(k) + »(k)
(41)
risk of this approach is that during the transient of z(k) = Hx(k) + ´(k)
reconfiguration process, the model with the highest RT
probability may not necessarily correspond to the where F = eAT , G = ( 0 eA¿ d¿ )B, H = C, and the
correct failure mode. Incorrectly selected control sampling period T = 0:1 s.
signals can introduce further transients into the Parameters are given as follows. Q =
system. diagf0:012 , 0:012 , 0:012 , 0:012 g, R =
Clearly, a combination of these two approaches diagf0:22 , 0:22 g, x0 = [250 50 10 8]T . The external
would be advantageous. To be more specific, during control input is selected as u = [100 100]T . Initial
the transient period, the Bayesian approach is used. parameters for the Kalman filters in the filter bank are
Once the failure has been detected and diagnosed x̂0j = x0 , Pj0 = 99I, Qj = Q, and Rj = 2R, 8j = 1, :::, N.
with some degree of certainty, the control signal The nominal controller is designed using LQR, the
from the model with the highest probability will be state and control weighting matrices are chosen as
used. In summary, this combined decision rule can be QLQR = diagf1, 1, 1, 1g and RLQR = diagf1, 1g.
described as follows: Since the controlled variables are horizontal
if ¹j (k) · ¹T velocity and vertical velocity, the command tracking
matrix Hc in (36) is chosen as
N
X · ¸
u(k) = fLi r(k) ¡ ui (k)g ¢ ¹i (k) 1 0 0 0
(39) Hc = :
i=1 0 1 0 0
otherwise
u(k) = Lj r(k) ¡ uj (k), j = arg max ¹i (k) B. Fault Modeling and Model Set Design
i

where ¹T denotes the threshold which takes the same In the following, we consider a simple situation
value as in (10). where there is only a single failure among system

1228 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 37, NO. 4 OCTOBER 2001
 TABLE II
System Matrices for Normal and Fault Modes

component, actuator, or sensor in any given test run. 1) One CDI is counted if the model that is closest
Faults are simulated at t = 5 s in each scenario. There to the system mode (normal or fault mode) in effect
are 4 possible modes in total. Table II represents the at a given time has a probability higher than the
system, control, and measurement matrices for each threshold ¹T = 0:9.
mode, with the parameter changes highlighted. 2) One IFI is counted if the model with a
The following model transition probability matrix probability over ¹T is not the one closest to the fault
is used mode in effect at the given time.
2 3 3) One FA is counted if the model with a
87=90 1=90 1=90 1=90
6 0:1 probability over ¹T is not the normal mode while the
6 0:9 0 0 77
¦ =6 7: normal mode is in effect at the given time.
4 0:1 0 0:9 0 5 4) One MD is counted if the normal model has the
0:1 0 0 0:9 highest probability which exceeds ¹T while the system
has a fault.
C. Indices for Performance Evaluation 5) One NMD is counted if no model has a
probability above ¹T .
In order to evaluate the performance, the following 6) DCID is the time delay that FDD takes to
measure is used correctly detected and diagnosed fault.
e(k) = kHc (xnormal (k) ¡ xreconfigured (k))k2 (42) Obviously, it is desirable to have higher CDI and
lower FA, IFI, MD, and NMD. The more detailed
where xnormal (k) denotes the nominal closed-loop
discussions on the performance evaluation and indices
system state when there were no failures, while
for FDD can be found in [36].
xreconfigured (k) denotes the reconfigured closed-loop
Similarly, to evaluate the reconfiguration
system state at time k.
mechanism, percentages of correct reconfiguration
In addition, the mean and the maximum value
(CR), incorrect reconfiguration (IR), and the
of e(k), 8k 2 f1, kmax g are also used as overall
reconfiguration delay (RD) are also used in this work.
performance indicators, i.e.,
kmax
1 X D. Results
e= e(k) (43)
kmax 1) Eigenstructures and Controller Gains: The
k=1

emax = maxfe(k)g: (44) eigenvalues and eigenvectors of the closed-loop

system in fault-free mode and different fault modes
In addition to the conventional performance are presented in Tables III and IV, respectively. The
indices, such as false alarm (FA) and missed detection designed controller gains are shown in Table V. It is
(MD), the following performance measures have also clear that the eigenvalues under different faults are
been used in this work: average percentages of correct assigned exactly to those of the fault-free system,
detection and identification (CDI), incorrect fault and the corresponding eigenvectors are assigned as
identification (IFI), no mode detection (NMD), and close to those of the fault-free system as possible.
detection and correct identification delay (DCID). To evaluate quantitatively how close the assigned

ZHANG & JIANG: INTEGRATED ACTIVE FAULT-TOLERANT CONTROL USING IMM APPROACH 1229
 TABLE III
Eigenvalues of Designed Reconfigurable Control System

TABLE IV
Eigenvectors of Designed Reconfigurable Control System

TABLE V
Controller Gains for Fault-Free and Reconfigurable Control

eigenvectors are to those of the prefault system, system is recovered by using the proposed approach
misalignment of the eigenvectors are calculated in after a short transient. However, without the controller
terms of distances and angles between them which reconfiguration, the system becomes unstable. The
are shown in Table IV. mode probability and the control switching sequence
As can be seen from the Table IV, the new for FDD and the controller reconfiguration are shown
eigenvectors are very close to those of the prefault in Fig. 5. It can be observed that even though there
system. are a few mis-switchings at the beginning stage of
2) Responses of the System Under Component the reconfiguration (around t = 5 s), satisfactory
Failure: The plant and reconfigured output responses performance has eventually been achieved. Fig. 6
in the presence of the system component faults demonstrates the closed-loop control signals in the
are illustrated in Fig. 4. For comparison purpose, two control channels, from which one can easily
the plant output responses without the controller examine how the closed-loop control signals have
reconfiguration are also shown. It can be seen that been adjusted accordingly to compensate for the
the output responses of the fault-free closed-loop fault.

1230 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 37, NO. 4 OCTOBER 2001
 Fig. 4. Output responses under system component fault.

Fig. 5. Mode probability and control switching sequence under system component fault.

TABLE VI
Performance of Proposed Scheme

FDD and reconfiguration, satisfactory performance

has been obtained via the proposed integrated active
fault-tolerant control scheme.
3) Responses of the System Under Actuator Failure:
The plant and reconfigured output responses in the
presence of the actuator fault are illustrated in Fig. 7.
It can be seen that the output responses of the original
closed-loop system are completely recovered. The
corresponding mode probability and the control
switching sequence are given in Fig. 8. However,
without the controller reconfiguration, the closed-loop
system becomes unstable. Fig. 9 demonstrates the
Fig. 6. Closed-loop controller output under system component
corresponding control signals. Similar to the case of
fault.
the system component fault, performance indices are
In addition, the quantitative performance indices also given in Tables VI and VII.
described previously are given in Table VI. Table VII 4) Responses of the System Under Sensor Failure:
presents the results for the case with perfect FDD for Figs. 10—12 show the behavior of the system in
reconfiguration. All results in the Tables VI and VII the presence of the sensor fault. Satisfactory FDD
are the average of 100 Monte Carlo simulation runs. and reconfiguration results have been obtained. It
Compared with the performance with correct (perfect) should be noted that even though the design of the

ZHANG & JIANG: INTEGRATED ACTIVE FAULT-TOLERANT CONTROL USING IMM APPROACH 1231
 Fig. 7. Output responses under actuator fault.

Fig. 8. Mode probability and control switching sequence under actuator fault.

TABLE VII
Performance of FDD and FTCS with Correct FDD and Reconfiguration

to the fact that the same control law has been used
as the normal controller, and that the satisfactory
state estimation for both fault-free and fault cases
have been obtained. There are no obvious
changes in the corresponding closed-loop control
signals. This fact can also be seen from Tables VI
and VII.
The results from the simulation have indicated
that the proposed integrated FDD and reconfigurable
control scheme can deal with system component,
actuator and sensor faults effectively.
Fig. 9. Closed-loop control input under actuator fault.
VI. CONCLUSIONS

reconfigurable feedback controller is not influenced An integrated MM fault detection, diagnosis, and
directly by sensor faults due to the feedback of reconfigurable control scheme has been proposed.
the estimated states, sensor faults do affect the FDD has been carried out using the IMM approach.
implementation of the closed-loop control. This can Steady-state tracking of a constant reference command
be observed from Fig. 11. Fig. 12 demonstrates the has been achieved using a feedforward control
control signals in the presence of the sensor fault. Due technique, and the reconfigurable controller is

1232 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 37, NO. 4 OCTOBER 2001
 Fig. 10. Output responses under sensor fault.

Fig. 11. Mode probability and control switching sequence under sensor fault.

REFERENCES

[1] Andry, A. N., Jr., Sharipo, E. Y., and Chung, J. C. (1983)

Eigenstructure assignment for linear systems.
IEEE Transactions on Aerospace and Electronic Systems,
19 (Sept. 1983), 711—729.
[2] Athans, M., Castanon, D., Dunn, K., Greene, C. S., Lee,
W. H., Sandell, N. R., Jr., and Willsky, A. S. (1977)
The stochastic control of the F-8C aircraft using a
multiple model adaptive control (MMAC) method–Part
1: Equilibrium flight.
IEEE Transactions on Automatic Control, 22 (May 1977),
768—780.
Fig. 12. Closed-loop control input under sensor fault. [3] Balle, P., Fischera, M., Fussel, D., Nells, O., and
Isermann, R. (1998)
Integrated control, diagnosis and reconfiguration of a heat
designed via eigenstructure assignment. To achieve exchanger.
IEEE Control Systems (June 1998), 52—63.
fast and reliable fault detection, diagnosis, and
[4] Basseville, M. (1988)
controller reconfiguration, an index related to the Detecting changes in signals and systems–A survey.
closed-loop performance has been combined with the Automatica, 24 (1988), 309—326.
information of the mode probability. New strategy for [5] Blom, H. A. P., and Bar-Shalom, Y. (1988)
generation of the aggregated reconfigurable control The interacting multiple model algorithm for systems
signal has also been proposed. Several performance with Markovian switching coefficients.
indices have been introduced for the evaluation of the IEEE Transactions on Automatic Control, 33 (Aug. 1988),
780—783.
integrated FDD and reconfigurable control scheme.
Simulation results for an aircraft example in the [6] Bodson, M., and Groszkiewicz, J. (1997)
Multivariable adaptive algorithms for reconfigurable flight
presence of system component, actuator and sensor control.
faults have shown the effectiveness of the proposed IEEE Transactions on Control System Technology, 5 (Mar.
approach. 1997), 217—229.

ZHANG & JIANG: INTEGRATED ACTIVE FAULT-TOLERANT CONTROL USING IMM APPROACH 1233
 [7] Boskovic, J., and Mehra, R. K. (1998) [21] Moerder, D. D., Halyo, N., Broussard, J. R., and
A multiple model-based reconfigurable flight control Caglayan, A. K. (1989)
system design. Application of precomputed control laws in a
In Proceedings of the 37th IEEE Conference on Decision reconfigurable aircraft flight control system.
and Control, Tampa, FL, Dec. 1998, 4503—4508. Journal of Guidance, Control, and Dynamics, 12
[8] Chandler, P. R. (1984) (May—June 1989), 325—333.
Self-repairing flight control system reliability and [22] Morse, W. D., and Ossman, K. A. (1990)
maintainability program executive overview. Model following reconfigurable flight control systems for
In Proceedings of the IEEE National Aerospace and the AFTI/F-16.
Electronics Conference, Dayton, OH, 1984, 586—590. Journal of Guidance, Control, and Dynamics, 13
(Nov.—Dec. 1990), 969—976.
[9] Chen, J., and Patton, R. J. (1999)
Robust Model-based Fault Diagnosis for Dynamic Systems. [23] Napolitano, M. R., Naylor, S., Neppach, C., and
Norwell, MS: Kluwer Academic Publishers, 1999. Casdorph, V. (1995)
On-line learning nonlinear direct neurocontrollers for
[10] Frank, P. M. (1990)
restructurable control systems.
Fault diagnosis in dynamic systems using analytical and
Journal of Guidance, Control, and Dynamics, 18 (Jan.—Feb.
knowledge-based redundancy–A survey and some new
1995), 170—176.
results.
[24] Napolitano, M. R., and Swaim, R. L. (1991)
Automatica, 26 (1990), 459—474.
New technique for aircraft flight control reconfiguration.
[11] Gao, Z., and Antsaklis, P. J. (1991) Journal of Guidance, Control, and Dynamics, 14 (Jan.—Feb.
Stability of the pseudo-inverse method for reconfigurable 1991), 184—190.
control systems. [25] Narendra, K. S. Balakrishnan, J., and Coliz, M. K. (1995)
International Journal of Control, 53 (1991), 717—729. Adaptation and learning using multiple models,
[12] Huang, C. Y., and Stengel, R. F. (1990) switching, and tuning.
Restructurable control using proportional-integral implicit IEEE Control System Magazine (June 1995), 37—51.
model following. [26] Narendra, K. S., and Tripathi, S. S. (1973)
Journal of Guidance, Control, and Dynamics, 13 Identification and optimization of aircraft dynamics.
(Mar.—Apr. 1990), 303—309. Journal of Aircraft, 10 (Jan. 1973), 193—199.
[13] Jiang, J. (1994) [27] Noura, H., Sauter, D., Hamelin, F., and Theilliol, D. (2000)
Design of reconfigurable control systems using Fault-tolerant control in dynamic systems: Application to
eigenstructure assignment. a winding machine.
International Journal of Control, 59 (1994), 395—410. IEEE Control System Magazine (Feb. 2000), 33—49.
[14] Jiang, J., and Zhao, Q. (1998) [28] Patton, R. J. (1997)
Fault tolerant contol system synthesis using imprecise Fault-tolerant control: The 1997 situation.
fault identification and reconfiguration control. In Proceedings of IFAC Symposium SAFEPROCESS’97,
In Proceedings of the IEEE International Symposium Hull, UK, Aug. 1997, 1033—1055.
on Intelligent Control, Gaithersburg, MD, Sept. 1998, [29] Patton, R. J., Frank, P. M., and Clark, R. N. (1989)
169—174. Fault Diagnosis in Dynamic Systems, Theory and
[15] Jiang, J., and Zhao, Q. (2000) Applications.
Design of reliable control systems possessing actuator Englewood Cliffs, NJ: Prentice-Hall, 1989.
redundancies. [30] Rago, C., Prasanth, R., Mehra, R. K., and Fortenbaugh, R.
Journal of Guidance, Control, and Dynamics, 23 (1998)
(July—Aug. 2000), 709—718. Failure detection and identification and fault tolerant
control using the IMM-KF with applications to the
[16] Li, X. R. (1996)
eagle-eye UAV.
Hybrid estimation techniques.
In Proceedings of the 37th IEEE Conference on Decision
In C. T. Leondes (Ed.), Control and Dynamic Systems, 76.
and Control, Tampa, FA, Dec. 1998, 4208—4213.
New York: Academic Press, 1996, 213—287.
[31] Rauch, H. E. (1995)
[17] Looze, D. P., Weiss, J. L., Eterno, J. S., and Barrett, N. M. Autonomous control reconfiguration.
(1985) IEEE Control System Magazine (Dec. 1995), 37—48.
An automatic redesign approach for restructurable control
[32] Schott, K. D., and Bequette, B. W. (1997)
systems.
Multiple model adaptive control.
IEEE Control System Magazine (May 1985), 16—22.
In R. Murray-Smith and T. A. Johansen (Eds.), Multiple
[18] Maybeck, P. S., and Hanlon, P. D. (1995) Model Approaches to Modeling and Control.
Performance enhancement of a multiple model adaptive London, UK: Taylor and Francis, 1997, 269—291.
estimator. [33] Veillette, R. J., Medanic, J. V., and Perkins, W. R. (1992)
IEEE Transactions on Aerospace and Electronic Systems, Design of reliable control systems.
31 (Oct. 1995), 1240—1254. IEEE Transactions on Automatic Control, 37 (Mar. 1992),
[19] Maybeck, P. S., and Stevens, R. D. (1991) 290—300.
Reconfigurable flight control via multiple model adaptive [34] Willsky, A. S. (1976)
control methods. A survey of design methods for failure detection in
IEEE Transactions on Aerospace and Electronic Systems, dynamic systems.
27 (May 1991), 470—479. Automatica, 12 (1976), 601—611.
[20] Mazor, E., Averbuch, A., and Bar-Shalom, Y. (1998) [35] Zhang, Y. M., and Jiang, J. (2001)
Interacting multiple model methods in target tracking: A Integrated design of reconfigurable fault-tolerant control
survey. systems.
IEEE Transactions on Aerospace and Electronic Systems, Journal of Guidance, Control, and Dynamics, 24 (Jan.—Feb.
34 (Jan. 1998), 103—123. 2001), 133—136.

1234 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 37, NO. 4 OCTOBER 2001
[36] Zhang, Y. M., and Li, X. R. (1998) [37] Zhou, D. H., and Frank, P. M. (1998)
Detection and diagnosis of sensor and actuator failures Fault diagnosis and fault tolerant control.
using IMM estimator. IEEE Transactions on Aerospace and Electronic Systems,
IEEE Transactions on Aerospace and Electronic Systems, 34 (Apr. 1998), 420—427.
34 (Oct. 1998), 1293—1313.

Youmin Zhang (M’99) received the B.S., M.S. and Ph.D. degrees in electrical
engineering from Northwestern Polytechnical University, Xian, P. R. China, in
1983, 1986, and 1995, respectively.
Since 1986, he has worked in the Department of Automatic Control,
Northwestern Polytechnical University, Xian, P. R. China, where he was promoted
as an Associate Professor in 1992 and was appointed as an Associate Chair of
the Department in 1994. During Sept. 1995 and Aug. 2000, he held several
research positions in the University of New Orleans, Louisiana State University,
State University of New York at Binghamton, and The University of Western
Ontario, respectively. He is currently an Assistant Professor in the Department
of Electrical and Computer Engineering at The University of Western Ontario,
London, Ontario, Canada. His main research interests include fault detection,
diagnosis and fault-tolerant (control) systems; machinery condition monitoring;
signal and data processing; estimation, identification theory and algorithms for
signal processing, communications, and control; neural networks and applications
to modeling, identification, diagnosis and control.
Dr. Zhang has published 34 refereed journal papers and more than 50
international conference papers, edited one monograph and coauthored a graduate
textbook. He received an award from Natural Sciences and Engineering Research
Council (NSERC) of Canada for a Visiting Fellowship in Canadian Government
Laboratories. He was the recipient of several research achievement awards from
National Education Commission of China, Aeronautics and Astronauts Industry
Ministry of China, and Shaanxi Province Education Commission; an outstanding
teaching achievement award from Shaanxi Province; several outstanding teaching
and research awards from Northwestern Polytechnical University. He served
as the General Chair of the 10th Youth Conference on Automatic Control and
Automation in China in 1994 and an international program committee member of
several international conferences. He served as an Associate Editor of the 2001
IEEE Conference on Control Applications. He serves also as a reviewer of several
refereed international journals and conferences.

Jin Jiang (M’85–SM’94) obtained his Ph.D. in electrical engineering from the
University of New Brunswick, Fredericton, New Brunswick, Canada, in 1989.
Currently, he is a Professor in the Department of Electrical and Computer
Engineering at The University of Western Ontario, London, Ontario, Canada.
His research interests are in the areas of fault-tolerant control of safety-critical
systems, power system dynamics and controls, and advanced signal processing.

ZHANG & JIANG: INTEGRATED ACTIVE FAULT-TOLERANT CONTROL USING IMM APPROACH 1235