32

IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICSPART B: CYBERNETICS, VOL. 35, NO. 1, FEBRUARY 2005

Fault Tolerant Control of Multivariable Processes

Using Auto-Tuning PID Controller
Ding-Li Yu, Member, IEEE, T. K. Chang, and Ding-Wen Yu

AbstractFault tolerant control of dynamic processes is investigated in this paper using an auto-tuning PID controller. A fault
tolerant control scheme is proposed composing an auto-tuning
PID controller based on an adaptive neural network model. The
model is trained online using the extended Kalman filter (EKF)
algorithm to learn system post-fault dynamics. Based on this model,
the PID controller adjusts its parameters to compensate the effects
of the faults, so that the control performance is recovered from
degradation. The auto-tuning algorithm for the PID controller is
derived with the Lyapunov method and therefore, the model predicted tracking error is guaranteed to converge asymptotically. The
method is applied to a simulated two-input two-output continuous
stirred tank reactor (CSTR) with various faults, which demonstrate
the applicability of the developed scheme to industrial processes.
Index TermsAdaptive NN models, auto-tuning PID, continuous stirred tank reactor (CSTR), extended Kalman filter (EKF),
fault tolerant control.

I. INTRODUCTION

ITH rapid increase in complexity of modern control systems, the importance of the fault tolerant control (FTC)
concept and technology has been realized and accepted by industry. Control system stability and reliability are not only critical for some projects where strict safety conditions apply, e.g.,
nuclear power stations and passenger airplanes, but also essential for large-scale productions, since most of modern industrial
plants are complex and often include a number of subsystems
which may compensate for the effects of sensor faults and component malfunction. This requires solutions that are very costly
in both hardware and development effort. Therefore, FTC is
very important from the viewpoint of safety, as well as reduced
production costs. FTC offers the capability to avoid inadvertent
process shut downs from simple faults, e.g. in instrumentation
and control loops that could develop into production stoppages
or plant failures.
Currently, FTC in most real industrial systems are realized by
hardware redundancy. For example, the majority-voting scheme
is used with redundant sensors to cope with sensor faults [1].
However, due to two main limitations of the hardware redundancy, high cost, and taking more space, solutions using analytical redundancy [1] have been investigated over the last two

Manuscript received February 7, 2004; revised June 3, 2004. This paper was
recommended by Associate Editor S. Phoha.
D.-L. Yu is with the Control Systems Research Group, School of Engineering,
Liverpool John Moores University, Liverpool, L3 3AF, U.K.
T. K. Chang is with the Control Systems Research Group, School of Engineering, Liverpool John Moores University, Liverpool, L3 3AF, U.K.
D.-W. Yu is with the Department of Automation, Northeast University, Qinhuangdao, China (e-mail: D.Yu@livjm.ac.uk).
Digital Object Identifier 10.1109/TSMCB.2004.839247

decades. There are generally two different approaches using analytical redundancy: 1) passive approaches, and 2) active approaches. Passive approaches use robust control techniques to
design closed-loop systems so that it is insensitive to certain
faults, e.g., [2]. Recently, an elegant design method of passive
approach was proposed by Chen et al. [3], in which the linear
matrix inequality (LMI) method was used to synthesis the reliable controller. Different faults were formulated as constraints
in the method and were considered in the optimal design using
LMI. While the design example showed the stability and maintenance of satisfactory system performance, a limitation is that
the method is based on an accurate linear state space model and
therefore, is not capable of controlling nonlinear processes for
which an accurate analytical model is usually unavailable. In addition, because the passive approaches consider fault tolerance
in only the stage of controller design without taking adaptation
when faults occur, the amplitude of the faults that can be tolerable is usually limited.
Active approaches use online fault accommodation information and reconfigurable controllers. When a fault is detected
using analytical or hardware redundancy, the controller is
reconfigured to guarantee the post-fault stability and maintain
acceptable performance. Active FTC has been investigated
using different methods including the feedback linearization
[4], control law rescheduling [5], and model following control
[6]. Reconfigurable control against plant component faults
has been studied using state feedback, where the feedback
gain matrix was designed using linear quadratic regulation
method [7], the pseudo inverse method [8] and eigenstructure
assignment method [9]. However, these studies were also based
on linear models so that they are not suitable for nonlinear
processes. Model predictive control (MPC) has been employed
in FTC [10], [11], where an adjustable objective function was
optimized based on a simple linear model. The model was
expected to learn the post-fault dynamics if the amplitude of
the fault is not too big. The research utilizing this method is
active and the models used are extended to nonlinear models.
More recently, it was reported that neural networks (NN) had
been employed to tackle FTC problems for nonlinear systems.
A pseudo-linear NN was used in a FTC system to form an additional feedback loop, which was used to compensate for the
degradation of system performance caused by component faults
[12]. A neural network was used in this method as a compensator that was adapted with a modified gradient-descent algorithm. The method guarantees the overall system stability for
linear systems, while it is still an open problem to ensure global
system stability for nonlinear systems. Besides, slow convergence of gradient-descent algorithms may result in instability

1083-4419/$20.00 2005 IEEE

YU et al.: FAULT TOLERANT CONTROL OF MULTIVARIABLE PROCESSES

of overall system when the system is subject to the faults with

large amplitude.
To model system dynamics after fault occurring (post-fault
dynamics), neural network models are made adaptive with recursive training. For example, a hierarchically self-organizing
learning (HSOL) algorithm was developed to find a minimal
number of hidden units for radial basis function (RBF) networks starting from none of centers [13]. Platt [14] proposed
a sequential learning algorithm for resource allocating network
(RAN) as a growing RBF structure. This method was modified
by Kadirkamanathan and Niranjan [15] using the EKF to replace the least mean squares (LMS) to update the output layer
weights. The RAN was further augmented by Lu et al. [16] by
introducing center pruning to achieve a minimal network. Chen
et al. [17] developed a hybrid algorithm for recursive training of
RBF networks with fixed structure. Fabri and Kadirkamanathan
[18] and Liu et al. [19] proposed an adaptive RBF network with
variable structure. Some of these adaptive networks were used
for time-series prediction and the others for adaptive control.
A common feature of these adaptive networks is that they use
Gaussian RBF structure due to its linear output layer. However,
a limitation is that this network needs also to adapt its structure to copy with wide operating space of the system, especially when the operation point moves to a far region caused
by faults. The structure adaptation usually needs a period of online training before the modeling error caused by the structure
change is reduced.
It is well known that PID controllers have dominated industrial control for half of a century, and there has been a great
deal of research interest into the implementation of advanced
controllers. The reason is that the PID control has a simple
structure that is easy to be understood by field engineers, and
is robust to disturbance and system uncertainty. Relay feedback is a simple and reliable test that keeps the process output
under closed-loop control and makes it close to the operating
point. strm and Hgglund [20] combined the strengths of
both PID and relay control and invented the relay auto-tuner for
a single-input single-output (SISO) PID controller. The tuner
has been widely and successfully applied in industry and many
further researches to improve this technique followed [21][23].
A tutorial given by Hang et al. [24] outlined the recent development in this aspect. Some other techniques have also been
used in developing auto-tuning PID controllers, such as the gain
and phase margin based method [25], the stable auto-tuning PID
method designed using the Lyapunov method [26], etc.
Since many industrial processes are inherently of a multivariable nature, they need multivariable control to cope with interactions so as to enhance performance. Development of multivariable auto-tuning PID controllers has not been very active recently. The early work includes a method for tuning the integral
part of the multivariable PID controller proposed by Davidson
[27]. Pentinnen, and Koivo [28] proposed a method for tuning
the and parts of the multivariable PID controller. The limitation of these methods is that they require some experimental and
graphical procedures, which are rather time consuming. Therefore, such methods are not suitable for online tuning. Zgorzelski
et al. [29], Loh et al. [30] and Zhuang and Atherton [31] attempted to develop multivariable PID controllers based on the

33

method by strm and Hagglund [20]. For the early multivariable PID control, Koivo and Tanttu [32] gave a survey for its
tuning techniques. These techniques mainly aim at decoupling
the plant at certain frequencies.
Methods based on online parameter estimation have also
been proposed for the automatic tuning of PID regulators.
Some authors proposed auto-tuning regulators based on minimum variance, pole placement, or linear quadratic Gaussian
(LQG) design methods. Gawthrop, Radke, [33], and Isermann
[34] proposed auto-tuning PID using adaptive parameter estimation methods. Hang and his co-workers have proposed
auto-tuning PID regulators using different methods, including
a knowledge-based PID auto-tuner [35]. Based on the method
given in Nishikawa, et al. [36], Ruano, et al. [37] proposed a
connectionist approach to PID auto-tuning, which used integral
measures of the step response as the input to neural networks to
determine the required PID parameter values. However, most
of these methods are for SISO systems only.
In this paper, a new FTC approach for unknown nonlinear
dynamic systems is proposed. The approach combines an
auto-tuning PID control with an adaptive neural network
model. The neural network model is adapted to learn post-fault
dynamics when faults occur, while the PID controller tune its
parameters online based on the adapted model to compensate
the effects of the faults on the system stability and performance.
The multilayer perceptron (MLP) network is adopted because
this kind of network generally does not need structure tuning
as that for the Gaussian RBF networks. The EKF algorithm is
used to update the network weights online as this algorithm
can learn the process dynamics much faster than the steepest
descent method. The auto-tuning algorithm compute an optimal
control variable based on the adaptive model to minimize the
sum-squared error function of the predicted system tracking
error. The Lyapunuv method is used to derive the optimal
control so that the convergence of the predicted tracking error
is guaranteed. Consequently, the stability and the tracking
performance of the entire system are maintained after occurring
of the faults.
The proposed FTC scheme is evaluated by a simulated MIMO
CSTR process as an unknown system for its performance and
stability under various simulated actuator and component faults.
Both abrupt faults and incipient faults are tested. The rest of the
paper is arranged as follows. Section II presents the configuration of the FTC scheme. The adaptive NN model and its online
learning algorithm is described in Section III. The auto-tuning
PID controller is described in Section IV. Section V presents the
evaluation of the adaptive model, while Section VI describes the
application of the proposed method to the CSTR and shows the
simulation results. Finally, some conclusions are drawn in Section VII.
II. CONFIGURATION OF FTC SCHEME
The main objective of the FTC is to obtain a control variable to drive the process to track the desired trajectory when
the process is not subject to any fault, while to maintain the
system stability and to recover from the performance degradation as much and as fast as possible when fault occurs in the

34

IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICSPART B: CYBERNETICS, VOL. 35, NO. 1, FEBRUARY 2005

is the estimated process output by the NN model

where
is an approximated nonlinear function of
. The
and
commonly used structure of MLP network with one hidden
layer of neurons is adopted
(3)

Fig. 1.

where
is the weight matrix connecting hidden
is the hidden layer output
layer and output layer,
vector in which each entry is transformed from the correby a nonlinear activation function
.
sponding entry in
The sigmoid activation function is chosen in this study

Configuration of the NN-based auto-tuning PID control system.

process. Considering that chemical or biological processes are

complex and their mathematical models are usually unknown,
an active fault tolerant control approach is developed in this section. The control scheme includes two parts, one is using a NN
to model the process and the model is made adaptive to catch the
dynamics change caused by the fault, another is an auto-tuning
PID controller based on the model. When the model captures the
post-fault dynamics, the PID controller is adapted to compensate the degradation of system stability and performance. The
configuration of the FTC system is shown in Fig. 1.
In Fig. 1 the MLP model is online adapted with the model
prediction error using the EKF algorithm. The adapted model is
used to predict process output at next sample time. The prediction is used by the auto-tuning algorithm to derive an optimal
control variable. The framed part in Fig. 1 is the PID controller.
PID controller adapts its parameters in the way that the produced
control variable will drive the NN model output to track the desired reference. A recursive auto-tuning algorithm derived using
the Lyapunov method will generate the optimal control that is
guaranteed to minimize a squared tracking error. The thick line
between the model used for prediction and that to be adapted
indicates that the structure and weights are shared between the
two models.
III. ADAPTIVE NEURAL NETWORK MODEL

(4)
(5)
where
is the weight matrix connecting the
with
input layer and hidden layer,
is the network input vector with an extra unity input
augmented to introduce bias, and is defined by

(6)
Therefore,

is expressed as

with the last entry introducing bias to the hidden layer.

B. EKF-Based Training Algorithm
The EKF algorithm is chosen for online updating of the MLP
model since it is much faster than the back-propagation or other
gradient-based algorithm. The model parameters to be trained
and
. To enable these parameter
are weight matrices
matrices to be adjusted with the EKF algorithm, they are foras follows:
mulated as a parameter vector

A. NN Model

(7)

An adaptive MLP network model is developed to model the

system dynamics and also to capture the time-varying dynamics
of the process in online mode. For a multivariable nonlinear
system, the following NARX model is used:

with
..
.

..
.

(8)

(1)
and
are the sampled process input and
where
output vectors,
and
are the input order and output order
respectively, denotes input transmission delay, is measuredenotes a vector valued nonlinear function.
ment noise and
The MLP network model used in this work is represented by the
following:

where
is the
row vector in
and
is the
row
. In applying the EKF algorithm for model paramvector in
eter estimation, the following equations are formulated to describe the model
(9)
(10)
where

(2)

(11)

YU et al.: FAULT TOLERANT CONTROL OF MULTIVARIABLE PROCESSES

35

is defined in (3) and

is modeling error
is estimated by applying the
vector. The parameter vector
EKF algorithm to (10) and (11).

where

(12)
(13)
(14)
where

(21)
The second term,

, can be expressed as

(15)
is the model output with the weights estimated at the last sample
time, and
(16)
..
.

is an unknown priori error covariance

where
matrix. Iiguni, et al. [38] proposed an online estimation with the
following equations according to Ljung and Sderstrm [39]

(17)

(18)

model output

in (16) is the Jacobian matrix of

with respect to , and is expressed as

..

..
.

(22)

The procedure of applying the EKF algorithm to the adaptive

model is given as follows.
Step 1) At sample time , obtain the past process output,
and the past control variable to form the NN
in (4).
model input vector
Step 2) Obtain the current process measurement output,
which is used as the training target.
using
Step 3) Update the error covariance matrix
(17), (18).
Step 4) Implement the EKF-based training algorithm
(19)(22) to update the weight parameter vector,
with (12)(16).
The developed adaptive MLP model is evaluated by modeling
the simulated CSTR process and this is described in Section V.
IV. PID CONTROLLER AUTO-TUNING ALGORITHM

(19)
The first term in the matrix in right-hand side of (19),
, can be expressed as

..
.

..

..
.

..

..
.

In this section, a self-learning algorithm for the multivariable

PID controller is proposed based on an adaptive neural model.
The model prediction is used in evaluation of the PID controlled
tracking error, and an optimal set of PID parameter will be iteratively approached, which gives minimum squared tracking
error.
Consider the discrete-time multivariable PID controller of the
following form:

..
.

..
.

..

(23)

..
.

where

is the process tracking error defined as

(20)
(24)

36

IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICSPART B: CYBERNETICS, VOL. 35, NO. 1, FEBRUARY 2005

with
being the desired trajectory. is the sampling
,
and
are PID
time,
controller parameter matrices, and is the number of input. In
order to estimate the optimum PID parameters,
,
and
, a parameter vector,
is formulated
as
..
.

where is a positive constant. Obviously,

inite. Define

is positive def-

(31)
in (30). In discrete time operation,
for the calculation of
can be approximated by the derivative of
with
respect to . Thus

(25)

denotes the th row in

and it is the same
where
and
,
.
for
An iterative algorithm is to be developed and used to obtain
the optimal control variable in each sample period by minimizing an objective function of the predictive tracking error.
Such tuned PID parameters will be used at the end of sample
period to produce control variable. A new argument, is introduced to denote the iterative step within each sample period to
distinguish from the sample time, . To avoid confusion with the
iterative step , variable at sample time is change to
in the iterative process, such as
. When predicting
,
is used to denote the optimum
PID parameters in the iterative process, and the PID auto-tuning
algorithm is defined as

(32)

Then, the increment of the Lyapunov function,

pressed as

is ex-

(26)
is the gain matrix,
is the NN model
where
tracking error that is different from the process tracking error,

(33)

(27)
(28)
We design the gain matrix in (26) as
is the desired output at sample time and
is
the NN model output in iterative step within the sample peis a function of the predicted optimal control variable,
riod.

(34)
(29)

Then,

in (34) becomes
(35)

with

In order to derive the gain matrix

in (26) at each iterative
step, so that the convergence of the NN model output to the
desired process output is guaranteed, a discrete-time Lyapunov
function is chosen as follows:
(30)

in (35) is always negative if

is
It is evidenced that
chosen. Therefore, the predictive tracking error is guaranteed to
in (34),
converge to zero by the design of the gain matrix
can be chosen as a learning rate to adjust the
in which
self-learning speed of the PID parameters.
The Jacobian matrix in (34) can be expressed as

(36)

YU et al.: FAULT TOLERANT CONTROL OF MULTIVARIABLE PROCESSES

The first term in the right-hand side of (36),

the MLP network model is derived as follows:

..
.

..
.

..

..

..
.

37

for

..
.

..
.

..

Fig. 2.

..
.

CSTR process.

V. ADAPTIVE MODEL EVALUATION

..
.

..

..
.

(37)

where

The second term in (36) is given

..
.

..

..
.

(38)

To evaluate the learning performance of the developed adaptive network, the network is used to model a multivariable, nonlinear CSTR process.
The CSTR process used in this research is a typical nonlinear
dynamic process used in chemical and biochemical industry.
The process will also be used to evaluate the developed FTC
scheme. A second order endothermic chemical reaction 2A
B takes place in the tank. The schematic diagram of the process
is shown in Fig. 2.
The process works in the following way. The reactant A with
flows
constant concentration and variable temperature
into the tank with the flow rate
. An endothermic chemical
reaction takes place in the tank, which is based on the temperature, and the reaction absorbs heat energy. As a result, the reaction influences the temperature and concentration of outflow
liquid. The two inputs and two outputs are chosen as follows:

with

The following equations are derived to describe the process dynamics:

From the above analysis, it is clear that when the gain matrix
is chosen according to (34), the PID parameter vector,
will asymptotically converge to the optimal value at each
sampling time in the sense of driving the model tracking error to
minimum. The PID auto-tuning procedure is given as follows.
Step 1) At sample time , obtain the desired trajectory
, the past process output
and the past
to form the NN model input
control variable
.
vector
Step 2) Implement the PID auto-tuning algorithm given in
(27), (34), (36)(38) and (26) in iterative form to
.
predict the optimum PID parameter vector
is assigned to be the value
The initial value
.
at the last sample time,
to the PID controller to
Step 3) Apply the obtained
in (29).
calculate the optimal control variable
to the NN model to calcuStep 4) Apply the obtained
in (11) and NN model
late the model output
in (27) at iteration step .
tracking error,
Step 5) Repeat Step 1 to Step 4 until the NN model tracking
is less than a pre-specified threshold or
error,
a specified bound to the iterative step is reached.
to be equivalent to
, and then
Step 6) Set
apply it to the PID controller in the process.

(39)

(40)
Equation (39) is a mass balance equation for chemicals while
(40) is a heat energy balance equation. The operating ranges of
the input variables for the process are
Kelvin

(41)

This means that a physical limitation with the lower bound of

2 l/s and the upper bound 5 l/s is applied to the inlet server pump,
and
while a physical limitation with the lower bound of 273
the upper bound of 480 is applied to the heating facilities of the
inlet liquid. These limitations have been realized in the process
simulation model in SIMULINK with two saturation blocks. It

38

IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICSPART B: CYBERNETICS, VOL. 35, NO. 1, FEBRUARY 2005

Fig. 3. MLP model training structure.

can be seen that the process is MIMO, nonlinear and with strong
coupling among variables. The CSTR process is simulated as
an unknown nonlinear dynamic process in the research to evaluate the performance of the proposed adaptive NN model. Although the model of the CSTR process used in the simulation
is known, it is supposed that the process model is unknown and
these known parameters have not been used in the modeling and
control.
In practice, measurement noise is inevitable in process
control. Therefore, the white noise of zero mean and unity
variance with appropriate amplitude is superimposed on the
process output. The simulated CSTR process was used to
generate two input-output data sets when two different random
amplitude binary sequences were used as excitation signals.
One set containing 5000 samples is for network learning and
the other containing 1200 samples for network validation. The
,
sample period was chosen as 30 s. In this simulation,
,
and 10 hidden layer nodes are found to be most
appropriate for the MLP model. Hence, the MLP model has the
following structure:

where the subscript (8:10:2) denotes that the network has eight
input nodes, ten hidden layer nodes, and two output nodes. The
MLP model is trained in the online mode with the EKF algorithm using the training data set in the same way as that in online updating displayed in Fig. 3.
The initial value of vector is assigned to small random
and
values. The initial values of the parameters of EKF,
are assigned an identity matrix and a zero matrix respectively. All the initial values of the Jacobian matrix in the
learning calculation are assigned to be zero. After training the
model is evaluated using the test data set for one-step-ahead prediction, during which the online learning of the model is still
conducted after prediction. The validation results are shown in
Fig. 4.
As the error is very small in Fig. 4 so that the process output
and the model output are almost virtually identical. To numerically evaluate the modeling performance, an error index of the
normalized mean absolute modeling error (NMAE) as shown
below is used:
NMAE

(42)

,
].
The NMAE of Fig. 4 is [
To demonstrate the online learning ability of the developed
adaptive NN model, another data set is generated when a mal-

Fig. 4.

Process and MLP model outputs.

Fig. 5. Modeling performance of the fixed parameter MLP model.

Fig. 6.

Absolute modeling error of the fixed MLP model.

function of an actuator is simulated to the process. The fault is

simulated by a 100% abrupt change on the actuator output,
starting at the sample

where denotes the faulty data. In order to clearly show the

learning performance with a fault present, the CSTR is activated
by a constant control variable with noise between the sample
and
. This is to test if the EKF-based
instants
learning algorithm works for the data without fully persistent
exciting. For comparison, a fixed parameter MLP model with
the same structure as the adaptive one is also trained using the
same set of training data, and then is evaluated with the same
test data with fault. The modeling result by the fixed parameter
model is displayed in Fig. 5 with the absolute modeling error
in Fig. 6, while the modeling error by the adaptive model is
displayed in Fig. 7 for comparison.

YU et al.: FAULT TOLERANT CONTROL OF MULTIVARIABLE PROCESSES

Fig. 7. Absolute modeling error of the adaptive MLP model.

In Fig. 5 it is observed that an obvious mismatch between the

model output and the process output occurs after the fault oc. This indicates that, without online learning, the
curs at
MLP model is unable to track the change in system dynamics.
This error is shown more clearly in Fig. 6 and is easily to compare with Fig. 7. In contrast, Fig. 7 shows the EKF-based online learning performance of the adaptive model for the data set
, the modwith fault. When an abrupt fault occurs at
eling error caused by the abrupt change of the system dynamics
is reduced to normal quickly. The new dynamics is modeled
quickly by the adaptive model. These results indicate that the
online learning performs very well even when the CSTR process
is subjected to a fault. It is evident that the EKF-based online
learning algorithm is able to learn the fault dynamics quickly.
This is important for an adaptive model to be used in the FTC
scheme. The modeling error index of adaptive MLP trained by
,
].
EKF is [
VI. APPLICATION TO THE CSTR PROCESS
The developed FTC scheme shown in Fig. 1 is applied to the
simulated CSTR process in Fig. 2. The process has been described in Section V and the neural network model structure
selection is also given in Section V. The process model developed in Section V is used here for fault tolerant control simulation. The control parameters chosen are as follows. The ini,
tial values of the PID parameters are chosen as
,
according to experience and the learning
is chosen. Two actuator faults and two component
rate
faults in abrupt and incipient modes are simulated and evaluated
with the developed scheme. In the simulations, the white noise
with zero mean, unity covariance and appropriate amplitude is
added as measurement noise to simulate realistic processes.
A. Case 1: Healthy System
Fig. 8 shows the process tracking response under no fault
conditions. It is seen that the proposed FTC approach can adequately control the MIMO CSTR process under no fault condition.
The control variables associated to the process response in
Fig. 8 are shown in Fig. 9. The dot-dash line indicates the
bounded range of control variable.
It can be seen in Fig. 8 and Fig. 9 that the process tracks
the desired output with a fast response and without overshoot.

39

Fig. 8. Tracking performance of auto-tuning PID controller (without fault).

Fig. 9.

Control variables associated to Fig. 8 (without fault).

The process tracking response is good in different regions of

the entire operating space. The tracking error index in Fig. 8 is
,
].
[
B. Case 2: With Actuator Faults
After the test of the system for no fault condition, two actu, and
, are simulated on the process. 1)
is
ator faults,
is simsimulated for the malfunction of the inlet pump; 2)
ulated for a temperature change in the inlet feed. The faults are
superimposed on the normal values of the corresponding variables to form the faulty data. Starting from sample instant 251,
an abrupt change of of 40% and a ramp change of of 30%
are simulated respectively, as described by the following:

where denotes the faulty data. The simulation results for

process tracking are shown below.
It is observed in Fig. 10 that since the adaptive MLP model
has not yet learned the post-fault dynamics at sample instant
, the incorrect predicted control variables drive the process
output away from the desired trajectory. This tracking error is
quickly compensated when the adaptive NN learns the post-fault
dynamics. Hence, the tracking error is reduced to zero within
10 samples. The tracking error index in Fig. 10 is [
,
]. The associated control variables generated by the FTC are displayed in Fig. 11. The control variables

40

IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICSPART B: CYBERNETICS, VOL. 35, NO. 1, FEBRUARY 2005

Fig. 10.
q ).

Process response under auto-tuning PID control (with abrupt fault,

Fig. 11.

Control variables associated to Fig. 10 (with abrupt fault,

1q ).

for normal operation without faults are also generated and displayed in Fig. 11 for comparison.
We can see in Fig. 11 that the control variables change quickly
with a big amplitude to respond to the fault occurrence. In responding, the controller output changes and has a steady state
difference from that for no fault condition. This is because of the
step change of of 40%.
Fig. 12 shows the process tracking response to a 30% ramp
. Because the ramp fault is a fault with its amactuator fault
plitude increased gradually, the process tracking response has a
very small tracking error. The associated controller outputs to
Fig. 12 are displayed in Fig. 13.
is simulated on the CSTR process
The ramp fault
starting from sample instant
, therefore the predicted
control signal for 30% change of
actuator fault is gradually diverging away from the control signal under no fault
condition. The process tracking response show that the control
performance and stability are maintained after 40% change of
and 30% change of . The adaptive MLP model is able to
learn the post-fault dynamics quickly with the EKF algorithm.
The proposed FTC with auto-tuning PID controller drives the
CSTR process output back to the desired process trajectory
quickly when a fault occurs. The tracking error index in Fig. 12
,
].
is [
C. Case 3: With Component Faults
, and
, are simulated on the
Two component faults,
process, where
is simulated for the outflow pipe restriction

Fig. 12.
T ).

Process response under auto-tuning PID control (with ramp fault,

Fig. 13.

Control variables associated to Fig. 12 (with ramp fault,

1T ).

is simulated for a concentration change of the

change and
inflow liquid. A complete blockage of the outlet is regarded as a
nonrecoverable fault, which can be fixed only by replacing the
hardware. In this project the emphasis is on the FTC problem
with a recoverable fault in the process.
Starting from sample instant 251, an abrupt change of 50% of
and a ramp change of 30% of are simulated respectively
as described by the following:

Fig. 14 shows the process tracking response to the 50%

abrupt component fault
. The associated control variables
are shown in Fig. 15.
It can be seen in Fig. 14 that the deviation of the process response from the desired trajectory caused by the fault is quickly
recovered. This is due to the compensation effects of the FTC
scheme via the auto-tuning PID controller. Observe the sample
, the process output is unable to track
interval
the desired trajectory due to that the concerned control variable
is limited by the physical bound as can be seen in Fig. 15.
,
The tracking error index in Fig. 14 is [
]. The process tracking response and control variable to
is shown in Fig. 16 and
a 30% ramp component fault
Fig. 17.

YU et al.: FAULT TOLERANT CONTROL OF MULTIVARIABLE PROCESSES

41

Fig. 14. Process response under auto-tuning PID control (with abrupt fault,
R ).

Fig. 17.

Control variables associated to Fig. 16 (with ramp fault,

1c ).

TABLE I
PROCESS TRACKING ERROR INDICES IN DIFFERENT FAULT CONDITIONS

Fig. 15.

Control variables associated to Fig. 14 (with abrupt fault,

1R

).

the adaptive MLPN is able to learn the component fault quickly

with the EKF algorithm. The method still work for bigger amplitude of the simulated faults. They are not displayed because
the outputs are distorted by the actuator bounds. The tracking
,
].
error index in Fig. 16 is [
VII. CONCLUSION

Fig. 16.
c ).

Process response under auto-tuning PID control (with ramp fault,

It is seen in Fig. 16 that the ramp component fault causes a

much smaller tracking error in both outputs than that caused by
the abrupt fault. The tracking error is quickly compensated so
that the system stability is not degraded at all. Fig. 17 shows
the corresponding control variables that have a big change to
respond to the component fault of the process. It is also noticed
in Fig. 16 that the process output is unable to track the desired
and
.
trajectory in sample intervals
This is because the control variable is limited by the physical
bound of the actuator. The simulation results demonstrate that
the control performance and stability are maintained even for
50% change and 30% change of , and . This implies that

The proposed FTC scheme employs an auto-tuning PID

controller to compensate the fault effects. The auto-tuning
algorithm is derived with Lyapunov method to guarantee the
convergence of the predicted tracking error. The MLP model
is made adaptive online with the EKF algorithm, the post-fault
dynamics can be modeled in time and thus, the degradation
in the process tracking performance and in system relative
stability is quickly recovered.
The effectiveness of the method is demonstrated in the
application to the simulated CSTR process for different faults.
Actuator and component faults in abrupt and incipient modes
with large amplitude can be tolerated with the developed FTC
scheme. The tracking error is compensated within about 10
sample periods and the system stability is maintained. Table I
shows the process tracking error indices in the simulations
under different fault conditions.
It should be noted that for the last two error indices the
tracking errors caused by the control variable bound are included. Therefore, the error indices caused by the faults in
the last two conditions should be in a similar level to that in
condition two and three, respectively.
It is noted that sensor faults are not included in evaluations. In
fact, sensor faults are not tolerable with the developed method.

42

IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICSPART B: CYBERNETICS, VOL. 35, NO. 1, FEBRUARY 2005

This is because that the adaptive NN model is online trained

with measurement process input output data and consequently,
the effects of sensor faults will also be modeled. This leads to
a process tracking error of the size of the occurred sensor fault.
Sensor fault tolerance in processes with unknown dynamics
is the research being undertaken. Due to lack of a test rig for
which the fault simulation mechanism is equipped, the developed approach is evaluated only by simulated process with the
simulated faults. The potential applications include different
industrial processes with multivariables, nonlinear dynamics,
and a considerable complexity, such as chemical or biological
processes.
REFERENCES
[1] R. J. Patton, P. M. Frank, and R. Clark, Fault Diagnosis in Dynamic Systems: Theory and Application. Englewood Cliffs, NJ: Prentice-Hall,
1989.
[2] M. Tyler and M. Morari, Optimal and robust design of integrated control and diagnostic modules, in Proc. American Control Conf., 1994.
[3] J. Chen, R. J. Patton, and Z. Chen, Active fault tolerant flight control
systems design using the linear matrix inequality method, Trans. Inst.
Meas. Control, vol. 21, no. 2/3, pp. 7784, 1999.
[4] Y. Ochi, Application of feedback linearization method in a digital restructurable flight control system, J. Guid. Control Dyn., vol. 16, no. 1,
pp. 111117, 1993.
[5] I. Kaminer, A. M. Pascoal, P. P. Khargonekar, and E. E. Coleman, A velocity algorithm for the implementation of gain-scheduled controllers,
Automatica, vol. 31, no. 8, pp. 11851192, 1995.
[6] W. D. Morse and K. A. Ossman, Model-following reconfigurable flight
control system for the AFTI/F-16, J. Guid. Control Dyn., vol. 13, no.
6, pp. 969976, 1990.
[7] D. P. Looze, J. L. Weiss, J. S. Eterno, and N. M. Barrett, An automatic
redesign approach for restructurable control systems, IEEE Contr. Syst.
Mag., vol. 5, no. 2, pp. 1622, May 1985.
[8] Z. Gao and P. J. Antsaklis, Stability of the pseudo-inverse method for
reconfigurable control systems, Int. J. Control, vol. 53, pp. 717729,
1991.
[9] J. Jiang, Design of reconfigurable control systems using eigenstructure
assignments, Int. J. Control, vol. 59, no. 2, pp. 395410, 1994.
[10] W. K. Son, O. K. Kwon, and M. E. Lee, Fault-tolerant model-based predictive control with application to boiler systems, in Proc. IFAC Symp.
SAFEPROCESS, Hull, U.K., 1997, pp. 12401245.
[11] J. M. Maciejowski, Modeling and predictive control: enabling technologies for reconfiguration, in Proc. IFAC Symp. System Structure
Control, Oct. 1997, pp. 253258.
[12] H. Wang and Y. Wang, Neural-network-based fault-tolerant control of
unknown nonlinear systems, Inst. Elec. Eng.Proc. Part D: Control
Theory Applications, vol. 146, no. 5, pp. 389398, 1999.
[13] S. Lee and R. M. Kil, A Gaussian potential function network with hierarchically self-organizing learning, Neural Netw., vol. 4, no. 2, pp.
207224, 1991.
[14] J. Platt, A resource allocating network for function interpolation,
Neural Comput., vol. 3, pp. 213225, 1991.
[15] V. Kadirkamanathan and M. Niranjan, A function estimation approach
to sequential learning with neural network, Neural Comput., vol. 5, pp.
954975, 1993.
[16] Y. W. Lu, N. Sundararajan, and P. Saratchandran, A sequential learning
scheme for function approximation using minimal radial basis function
neural networks, Neural Comput., vol. 9, no. 2, pp. 461478, 1997.
[17] S. Chen, S. A. Billings, and P. M. Grant, Recursive hybrid algorithm
for nonlinear system identification using radial basis function networks,
Int. J. Control, vol. 55, pp. 10511070, 1992.
[18] S. Fabri and V. Kadirkamanathan, Dynamic structure neural networks
for stable adaptive control of nonlinear systems, IEEE Trans. Neural
Netw., vol. 7, no. 5, pp. 11511167, Sep. 1996.
[19] G. P. Liu, V. Kadirkamanathan, and S. A. Billings, Variable neural
networks for adaptive control of nonlinear systems, IEEE Trans. Syst.
Man., Cybern. C, Appl. Rev., vol. 29, no. 1, pp. 3443, Feb. 1999.
[20] K. J. strom and T. Hagglund, Automatic tuning of simple regulators
with specifications on phase and amplitude margins, Automatica, vol.
20, pp. 645651, 1984.

[21] C. C. Hang, K. J. Astrom, and W. K. Ho, Relay auto-tuning in the presence of static load disturbance, Automatica, vol. 29, no. 2, pp. 563564,
1993.
[22] J. H. Park, S. W. Sung, and I. B. Lee, Improved relay auto-tuning with
static load disturbance, Automatica, vol. 33, no. 4, pp. 711715, 1997.
[23] W. K. Ho, Y. Hong, A. Hansson, H. Hjalmarsson, and J. W. Deng, Relay
auto-tuning of PID controllers using iterative feedback tuning, Automatica, vol. 39, no. 1, pp. 149157, 2003.
[24] C. C. Hang, K. J. Astrom, and Q. G. Wang, Relay feedback auto-tuning
of process controllersa tutorial review, J. Process Control, vol. 12, no.
1, pp. 143162, 2002.
[25] W. K. Ho, C. C. Hang, and L. S. Cao, Tuning of PID controllers based
on gain and phase margin specifications, Automatica, vol. 31, no. 3, pp.
497502, 1995.
[26] W. D. Chang, R. C. Hwang, and J. G. Hsieh, An auto-tuning PID control for a class of nonlinear systems based on Lyapunov approach, J.
Process Control, vol. 12, pp. 233242, 2002.
[27] E. J. Davidson, Multivariable tuning regulators: the feedforward and robust control of general servomechanism problem, IEEE Trans. Autom.
Control, vol. 21, no. 1, pp. 3547, Feb. 1976.
[28] J. Pentinnen and H. N. Koivo, Multivariable tuning regulators for unknown systems, Automatica, vol. 16, pp. 393398, 1980.
[29] P. Zgorzelski, H. Unbehauen, and A. Niederlinski, A new simple decentralized adaptive multivariable regulator and its application to multivariable plants, in Proc. 11th IFAC World Congr., Tallin, Estonia, 1990,
pp. 226231.
[30] A. P. Loh, C. C. Hang, C. K. Quek, and V. U. Vasnani, Auto-tuning of
multi-loop proportional-integral controllers using relay feedback, Ind.
Emg. Chem. Proc. Des. Dev., vol. 25, pp. 654660, 1993.
[31] M. Zhuang and D. P. Atherton, PID controller design for a TITO
system, Inst. Elect. Eng.Proc. Pt. D: Control Theory Applications,
vol. 141, pp. 111120, 1994.
[32] H. N. Koivo and J. T. Tanttu, Tuning of PID controllersurvey of SISO
and MIMO, in Proc. Preprints IFAC Int. Symp. Intelligent Tuning Adaptive Control, Singapore, 1991.
[33] P. J. Gawthrop, Self-tuning PID controller: algorithms and implementation, IEEE Trans. Automatic Control, vol. 31, no. 3, pp. 201209, Mar.
1986.
[34] F. Radke and R. Isermann, A parameter adaptive PID controller with
stepwise parameter optimization, Automatica, vol. 23, no. 4, pp.
449457, 1987.
[35] T. H. Lee, C. C. Hang, W. K. Ho, and P. K. Yue, Implementation of
a knowledge-based PID auto-tuner, Automatica, vol. 29, no. 4, pp.
11071113, 1993.
[36] Y. Nishikawa, N. Sannomya, T. Ohta, and H. Tanaka, A method for
auto-tuning of PID parameters, Automatica, vol. 20, no. 3, pp. 321332,
1984.
[37] A. E. B. Ruano, P. J. Fleming, and D. I. Jones, Connectionist approach
to PID auto-tuning, Instit. Elect. Eng.Proc. Part D: Control Theory
Applications, vol. 139, no. 3, pp. 279285, 1992.
[38] Y. Iiguni, H. Sakai, and H. Tokumaru, A real-time algorithm for a
multi-layered neural network based on the extended Kalman filter,
IEEE Trans. Signal Process., vol. 40, no. 4, pp. 959966, Apr. 1992.
[39] L. Ljung and T. Soderstrom, Theory and Practice of Recursive Identification. Cambridge, MA: MIT Press, 1983.

Ding-Li Yu (M00) received the B.Eng. degree from

Harbin University of Civil Engineering, Harbin,
China, in 1982, the M.Sc. degree from Jilin University of Technology (JUT), Jilin, China, in 1986,
and the PhD.. degree from Coventry University,
Coventry U.K. in 1995, all in electrical engineering.
He was a Lecturer at JUT from 1986 to 1990, before he came to University of Salford, Salford, U.K.,
as a Visiting Researcher in 1991. He then worked at
Liverpool John Moores University, Liverpool, U.K.,
as a Postdoctoral Researcher since 1995 and became
a Lecturer in 1998. He is currently a Reader in process control. His current research interests include fault detection and fault tolerant control of bilinear and
nonlinear systems, adaptive neural networks and their control applications, and
model predictive control for chemical processes and engine systems.

YU et al.: FAULT TOLERANT CONTROL OF MULTIVARIABLE PROCESSES

T. K. Chang received the B.Eng and Ph.D. degrees in electrical engineering

from Liverpool John Moores University, Liverpool, U.K., in 1998 and 2002,
respectively.
He was appointed Research Fellow in the Control Systems Research Group
at Liverpool John Moores University. His main research interests are fault detection, isolation and fault tolerant control for dynamic systems, neural network
modelling, and control applications.

43

Ding-Wen Yu received the B.Eng. and M.Sc. degrees from Beijing Chemical
Engineering Institute (BCEI), China, in 1982 and 1987, respectively, and the
Ph.D. degree from Nottingham Trent University, Nottingham, U.K., in 2000, all
in control engineering.
He was a Lecturer at BCEI from 1987 to 1994, before he came to University
of Salford as a Visiting Researcher in 1994. He then worked at Liverpool John
Moores University, Liverpool, U.K., as a Postdoctoral Research Fellow between
2001 and 2003. Then, he joined the Northeast University, Qinhuangdao, China,
in 2004 as a Professor. His current research interests include modeling and dynamic simulations of chemical processes, model predictive control of industrial
systems, and neural network applications in control and signal processing.