Abstract—Fault tolerant control of dynamic processes is There are generally two different approaches using
investigated in this paper using an auto-tuning of Adaptive PID analytical redundancy: 1) passive approaches, and 2) active
controller. A fault tolerant control scheme is proposed composing approaches. Passive approaches use robust control techniques
an auto-tuning Adaptive PID controller based on an adaptive to design closed-loop systems so that it is numb to certain
neural network model. The model is trained online using the
Unscented Kalman filter (UKF) algorithm to learn system post-
faults, e.g., [2]. In recent times, an elegant design method of
fault dynamics. Based on this model, the Adaptive PID controller passive approach was proposed by Chen et al. [3], in which
adjusts its parameters to compensate the effects of the faults, so the linear matrix inequality (LMI) method was used to
that the control performance is improved. The auto-tuning synthesis the consistent controller. Different faults were
algorithm for the Adaptive PID controller is derived with the formulated as constraints in the method and were considered
Lyapunov method and therefore, the model predicted tracking in the optimal design using LMI.
error is guaranteed to desired point asymptotically.
While the design example showed the stability and
Keywords—Adaptive NN models; auto-tuning Adaptive PID; maintenance of suitable system performance, a limitation is
Unscented Kalman filter (UKF); fault tolerant control. that the method is based on an accurate linear state space
I. INTRODUCTION model and therefore, is not capable of controlling nonlinear
processes for which an accurate analytical model is usually not
With fast increase in difficulty of modern control systems, available. In addition, because the passive approaches consider
the importance of the fault tolerant control (FTC) concept and fault tolerance in only the stage of controller design without
technology has been appreciated and accepted by industry. taking adaptation when faults occur, the amplitude of the
Control system stability and reliability are not only vital for faults that can be allowable is usually limited.
some projects where stringent safety conditions apply, e.g.,
nuclear power stations and passenger airplanes, but also Active approaches use online fault adjustment information
essential for significant productions, since most of present and reconfigurable controllers. When a fault is identified using
industrial plants are complex and often include a number of analytical or hardware redundancy, the controller is
subsystems which may balance for the effects of sensor faults reconfigured to guarantee the post-fault stability and maintain
and element malfunction. acceptable performance. Active FTC has been explored using
different methods including the feedback linearization [4],
This requires solutions that are very costly in both control law rescheduling [5], and model following control [6].
hardware and development attempt. Therefore, FTC is very Reconfigurable control against plant component faults has
important from the viewpoint of safety, as well as reduced been studied using state feedback, where the feedback gain
production costs. FTC offers the ability to avoid accidental matrix was designed using linear quadratic regulation method
process shut downs from simple faults, e.g. in instrumentation [7], the pseudo inverse method [8] and eigen structure
and control loops that could develop into production loss or assignment method [9].
plant failures.
However, these studies were also based on linear models
Recently, FTC in most real industrial systems are so that they are not suitable for nonlinear processes. Model
appreciated by hardware redundancy. For example, the predictive control (MPC) has been employed in FTC [10],
majority-voting scheme is used with redundant sensors to cope [11], where an adjustable objective function was optimized
with sensor faults [1]. However, due to two main limitations based on a simple linear model. The model was expected to
of the hardware redundancy, high cost, and taking more space, learn the post-fault dynamics if the amplitude of the fault is
solutions using analytical redundancy [1] have been not too large. The research utilizing this method is active and
investigated over the last two decades. the models used are extended to nonlinear models.
II. STRUCTURE OF FTC SCHEME where y R p and u R m are the sampled process output and
The aim of the FTC is to obtain a control variable to force input vectors, and ny and nu are the output and input order, d is
the process to track the preferred trajectory when the process the input transmission delay, e is the measurement noise and
is not subject to any error, while to keep up the system g is the vector valued nonlinear function. Similarly, the
constancy and to make progress from the performance MLP network model used is :
degradation high-speed when fault occurs in the process.
Allowing for that element or organic processes are difficult
ŷk ĝ u k d 1 , ...,u k d n u , ...,y k 1,..., y k n y e k
and their mathematical models are usually unfamiliar, an (2)
active fault tolerant control approach is developed in this
section. where ŷ R p is the estimated process output by the NN model
The control design includes two parts, one is using a NN to and ĝ is an approximated nonlinear function of g . The
model the process and the model is made adaptive to hold the MLP network with one hidden layer of neurons is
dynamics change caused by the fault, another is an auto-tuning implemented.
PID controller based on the model. When the model captures
the post-fault dynamics, the PID controller is modified to ŷk W y hk (3)
balance the degradation of system stability and performance.
The MLP model is online modified with the model calculation where W y R p q1 is the weight matrix connecting output and
error using the EKF algorithm. The modified model is used to hidden layers, h k R q1 is the hidden layer output vector
calculate process output at next sample time. The prediction is where each entry is transformed from the corresponding entry
used by the auto-tuning algorithm to derive an optimal control in z(k) by using a sigmoid activation function as :
variable.
NN based Adaptive PID auto tuning control system
Fault, noise
eT (k) eT | k
PID Auto- u(i) u(k) y(k)
ei (i) tuning Process
yd (k)
algorithm
(k)
-1
Z NN
y (i) model
NN
model
em (k)
A. NN Model
To model the system dynamics and also to capture the (6)
time-varying dynamics of the process, an adaptive Multi Layer T
Perceptron (MLP) network model is developed. The NARX
h k ,1
model used for this multivariable nonlinear system is : 1 1
Z 1 K
, ...,
Z q K
(7)
yk g u k d 1 ,...,u k d n u ,...,y k 1,..., y k n y e k 1 e 1 e
(1)
with the last entry introducing bias to the hidden layer.
B. UKF – Based Training Algorithm the a priori known mathematical model f̂ x k , namely
UKF Training The UKF algorithm for NN training is e(k)=f(k)- f̂ x k . We can adjust the weighs of the NN through
similar to that of EKF; again, all the connecting weights are
the observation if and only if the observability condition is
organized as a state vector, but now the state is calculated
met. When e(k) g(x k wk ) 0 , the error is well
through unscented transformation [12], [13] and propagated
analytically through nonlinear system without the need to approximated, and the more accurate model becomes the sum
evaluate the Jacobian matrix. The generic UKF can be of f(xk)and the NN approximation, so we have :
summarized as in the following steps. x k 1 f x k gx k , w k
(14)
STEP 1) Initialization w k 1 w k
w o E w o
PO E w o ŵ o w o ŵ o
T
If we represent the augmented state vestors as x ak x TK w Tk , T
i, k 1 ŵ k 1 L k)P i
0.5
L k
Because the observation is only related with the state x, the
equation yk hx k vk remains unchanged. Now the
k 1 i
i L, k 1 ŵ k 1 L k)P i L
0.5
(9) estimation of x̂ ak based on the new model (15) and the noisy
L k observation yk hx k vk . Once given
k 1 i
yi, k|k 1 g i, k|k 1, x k o, k 1 ŵ k 1o
k
L k
ŷ k i yi, k|k 1
2L (11)
i 0 i , k 1 ŵ k 1 L k)P
k 1 i
i
0.5
L k
(18)
L k )P
STEP 4) Measurement update 0.5
i L, k 1 ŵ k 1 i L
k 1
L k
2L i
Sk i yi, k|k 1 ŷ k yi, k|k 1 ŷ k Rk
T
V i et i T et i
e T k y d k yk
(25)
where is a positive constant and thus, V (i) is positive
where yd t R p is the desired trajectory, T is the sampling definite. For V(i) calculation, define et i et i 1 et i .
time, K p R m p , Ki R m p and K d R m p are PID controller
In discrete time operation, ˆ pid i can be approximated by
parameter matrices, and m is the number of input. To estimate
the optimum PID parameters, K p k , Ki k and Kd k , a the derivative of ˆ p id i with respect to i . Thus
parameter vector pid k R 3mp is formulated as :
e t i ˆ pid i
et i
ˆ pid i
K p k K i k K d k T
1
i
1 1
pid k yd k ŷ i
ˆ pid i
(20)
K p m k K im k K dm k T
ˆ pid i
ŷ i
where k p j k , k i j k and k d j k denotes the jth row in K p k , K pid i et i (26)
ˆ i
Ki k and Kd k , j = 1, 2,…, m. pid
pid i
ˆ
where K pid i is the gain matrix, e t i R is the NN model p
T T
1
(23) by applying the obtained ˆ pid i to the PID
1 ŷ i ŷ i ŷ i
K pid i K pid i ˆ
ˆ pid i pid i ˆ pid i controller.
(28) Step 4) Calculate the model output, ŷ i and model
Then V i becomes, V i 2 1 eTt i e t i . If < 1 tracking error, e t i in (22) by applying the
is chosen, then V i is always negative. In the design of the
obtained û i to the NN model at each iteration
gain matrix, < 1 is chosen as the learning rate to adjust the
self-learning speed of the PID parameters. Due to this, the step, i.
predictive tracking error will converge to zero.
Step 5) Repeat Step 1 to Step 4 until the NN model
Thus, when the gain matrix K pid i is chosen according to
tracking error, is less than a pre-specified
(28), the PID parameter vector, ˆ p id k will asymptotically
threshold or a specified bound to the iterative
converge to the optimal value at each sampling time in the
sense of driving the model tracking error to minimum value. step is reached.
The Tracking performance of Auto –tuning PID controller Step 6) Set pid k to be equivalent to ̂pid i final , and then
is as shown in fig. 2
apply it to the PID controller in the process.
R
h( A
v
t)
The PID auto-tuning procedure is as follows: The process works in the following way as shown in fig. 3.
The reactant A with steady concentration and variable
Step 1) For each sample time k, form the NN model temperature flows into the tank with the flow rate. An
input vector X(k) by obtaining the desired endothermic chemical reaction takes place in the tank, which is
based on the temperature, and the reaction absorbs heat energy.
trajectory yd k 1 the past control variable As a result, the reaction influences the temperature and
concentration of outflow liquid. The two inputs and two
u (k) and the past process output y (k). outputs are chosen as follows:
Step 2) Implement the PID auto-tuning algorithm given
q c
in (22), (28), and (21) in iterative form to predict U= i , y = (29)
Ti Tr
the optimum PID parameter vector, ˆ pid k . The The following equations are derived to describe the process
dynamics:
value at the last sample time pid k 1 is
dc( t )
1 h EA
ci q i ( t ) c ( t ) 2 A h k o c2 ( t ) e xp
assigned to be the initial value ˆ p id 0 . dt Ah R
RTr( t )
v (30)
dTr ( t ) 1
h The initial value of vector is assigned to small random
r r q i ( t )Ti ( t ) Tr ( t ) H Ah k o c 2 ( t ) values. The initial values of the parameters of EKF, Ew(0) and
dt r r Ah R v
Rk are assigned an identity matrix and a zero matrix
EA
k 1
exp U1 Th Tr ( t ) U 2 [Tr ( t ) Tx ] (31) respectively. After training the model is evaluated using the
RTr( t )
test data set for one-step-ahead prediction, during which the
Equation (30) is a mass balance equation for chemicals online learning of the model is still conducted after prediction.
while (31) is a heat energy balance equation. The working The validation results are shown in Fig. 5.
ranges of the input variables for the process are
1
q i (t ) 2, 5 , Ti (t ) [273 480] (kelvin) (32)
s
This means that a material limitation with the lower bound
of 2 l/s and the upper bound 5 l/s is applied to the inlet server
pump, while a material constraint with the lower bound of 273
and the upper bound of 480 is applied to the heating facilities
of the inlet liquid. These limitations have been realize in the
process simulation model in SIMULINK with two saturation
blocks.
It can be seen that the process is MIMO, nonlinear and with
strong couple among variables. The CSTR process is virtual as
an unknown nonlinear dynamic process in the research to
evaluate the performance of the opportunity adaptive NN
model. Although the model of the CSTR process used in the
simulation is known, it is believed that the process model is
unknown and these known parameters have not been used in
the modeling and control. Fig. 5 Process and MLP model Output
The proposed system employs an adaptive PID controller to [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
compensate the fault effects. The convergence of the predicted Symp. SAFEPROCESS, Hull, U.K., 2007, pp. 1240–1245.
tracking error for auto-tuning algorithm is derived with
Lyapunov method. In UKF algorithm, the post-fault dynamics [11] J. M. Maciejowski, “Modeling and predictive control: enabling
technologies for reconfiguration,” in Proc. IFAC Symp. System
can be modeled in time and thus, the degradation in the process Structure Control, Oct. 1997, pp. 253–258.
tracking performance and in system relative stability is quickly
[12] J. Julier, K. Uhlmann, and F. Durrant-Whyte, “A new approach for
recovered. filtering nonlinear systems,” in Proc. American Control Conf., Seattle,
The adaptive NN model is online trained with measurement WA, Jun. 2015, pp. 1628–1632.
process input output data and consequently, the effects of [13] J. Julier and K. Uhlmann, “A new method for the nonlinear
sensor faults will also be modeled. Thus, the sensor faults are transformation of means and covariances in filters and estimators,” IEEE
Trans. Autom. Control, vol. 45, no. 3, pp. 477–482, Mar. 2000.
not tolerable with the developed method. This leads to a
process tracking error of the size of the occurred sensor fault. [14] R. N. Lobbia, S. C. Stubberud, and M. W. Owen, “Adaptive extended
The research work can be undergone on sensor fault tolerance Kalman filter using artificial neural networks,” Int. J. Smart Eng. Syst.
in processes with unknown dynamics. The potential
Des., vol. 1, pp. 207–221, 2015.
applications include different industrial processes with
multivariables and nonlinear dynamics. [15] Ronghui Zhan, J. Wan, “ Neural network-aided adaptive unscented
Kalman filter for nonlinear state estimation ,” IEEE Signal Processing
Letters , Vol 13, Issue: 7, July 2006.