A New Fault Tolerant Nonlinear Model Predictive Controller Incorporating An UKF-Based Centralized Measurement Fusion Scheme

Iran. J. Chem. Chem. Eng.
Archive of SID
Research Note
Vol. 30, No. 4, 2011
A New Fault Tolerant Nonlinear Model Predictive Controller

Incorporating an UKF-Based
Centralized Measurement Fusion Scheme
Salahshoor, Karim*+; Salehi, Shabnam
Department of Automation and Instrumentation, Petroleum University of Technology, Tehran, I.R. IRAN
ABSTRACT: A new Fault Tolerant Controller (FTC) has been presented in this research by
integrating a Fault Detection and Diagnosis (FDD) mechanism in a nonlinear model predictive
controller framework. The proposed FDD utilizes a Multi-Sensor Data Fusion (MSDF)
methodology to enhance its reliability and estimation accuracy. An augmented state-vector model
is developed to incorporate the occurred sensor faults and then a UKF algorithm is utilized
to estimate the augmented state vector including system states along with the fault terms using
a centralized measurement fusion scheme. The designed FDD architecture is then merged with
a conventional NMPC to form a Fault-Tolerant Control System (FTCS). A series of sensor fault
senarios is conducted on a Continuous Stirred Tank Reactor (CSTR) to comparatively illustrate
the superiority of the proposed FTCS in eliminating the miserable impacts of the induced sensor faults
against a conventional NMPC.
KEY WORDS: Fault Tolerant Control System (FTCS), Nonlinear Model Predictive Controller (NMPC),
Fault Detection and Diagnosis (FDD), Unscented Kalman Filter (UKF), Multi Sensor Data
Fusion (MSDF).
INTRODUCTION
The ever-increasing complexity of modern chemical
plants and the tightly environmental regulations are pushing
the process industries to optimize their production
systems against any process abnormally. Any malfunction
in these plants will cause great economic losses or may
even result into safety dangers [1]. The preventive action
should be done by the control mechanism, creating a Fault
Tolerant Control (FTC) system. If the designed mechanism
operates correctly, the system function is satisfied even
after occurrence of a fault, possibly after a short time
period of performance degradation.
* To whom correspondence should be addressed.
+ E-mail: salahshoor@put.ac.ir
1021-9986/11/4/107
6/$/2.60
Model Predictive Control (MPC) has found successful

applications, especially in the process industries. The novel
characteristics of MPC have encouraged the researchers to
merge it with the FDD methods to develop FTC systems.
Maciejowski [2] was among the first one to utilize MPC
in FTC systems. Pranatyasto et al. [3] introduced a PCA-based
FTC system with MPC to control a simulated Fluid Catalytic
Cracking (FCC) unit. Prakash et al. [4] have developed a
Fault Tolerant Control System (FTCS) based on the
Generalized Likelihood Ratio (GLR) and a standard MPC
controller for a simulated, nonisothermal CSTR system.
107
www.SID.ir
Salahshoor K. & Salehi S.
Archive of SID
In this work, a new automated FDD system has been

proposed based on an UKF estimator. The proposed
methodology utilizes a Multi-Sensor Data Fusion (MSDF)
technique to enhance the accuracy and reliability of state
estimation. The developed FDD system is then merged
with a Nonlinear Model Predictive Controller (NMPC)
to design a FTC system. A variety of simulation test studies
has been carried out to illustrate the performances of the proposed
FTC system in a CSTR against various sensor faults.
THEORETICAL DEVELOPMENT
FDD System
A deep review on the subject of FDD methods
is prsented in [5]. In this paper, a FDD approach is
proposed by incorporating a KF-based multi-sensor data
fusion methodology to detect and estimate the occurring
sensor faults in chemical process plants. For this objective,
a measurement fusion scheme is utilized in which the
sensor measurements are first fused to provide an
augmented state vector including the bias terms. A set of
most commonly appearing faults in the sensors,
i.e. calibration and aging or wearing biases, is examined and
an UKF-based Centralized Measurement Fusion (CMF)
technique is used to estimate these faults. The estimated
states, including the system states and the estimated
sensor faults, are finally incorporated in a FTC
framework utilizing the NMPC approach.
Unscented Kalman Filter
The Kalman filter addresses the general problem of
n
trying to estimate the state x of a discrete-time

controlled process, governed by the following linear
stochastic difference equation:
x k = Ax k 1 + Bu k 1 + w k 1
x k = f (x k 1 , u k , v k ) + w k 1
(3)
z k = h(x k , k) + v k
(4)
Where f and h indicate known nonlinear functions.

Julier & Uhlmann [6,7] developed the UKF
algorithm which does not require to linearize the foregoing
general nonlinear system dynamics. The UKF algorithm
can be found in diffrent literatures such as [8].
KF-Based Measurement Fusion Approach
There are essentially two methods for measurement
fusion. The first simply merges the measurements into
an augmented observation vector and the second combines
the measurements using minimum mean square estimates [9].
In this paper, the first measurement fusion technique has
been used. In this method, the measurement vectors z1k

and z 2k , obtained from two (or more) sensors, are merged
into a new augmented measurement vector given by:
Zk = [(z1k )T
(z k2 )T ]T
(5)
Denoting:
H (xk , k ) =
Vk =
( h1 ( x k , k ) ) ( h 2 ( x k , k ))
( v1k ) ( v2k )
T T
T T
and
Eqs. (4) and (5) can then be formulated into a new

measurement equation, given by:
Zk = H(x k , k) + Vk
(6)
Based on the assumed statistical independence of

the two sensors, the covariance matrix Rk for the merged
measurement noise Vk can be defined as:
(1)
Rk =
with a measurement z that is:
z k = Hx k + v k
Vol. 30, No. 4, 2011
(2)
The random variables w k and vk represent the

process and measurement noise, respectively. They
are assumed to be independent of each other, modeled
by a white signal having normal probability distributions
with covariences Qk and Rk, respectively. In most
practical applications of interest, however, the process
dynamics and the measurement equations obey
the following non-linear relationships:
R1k
R 2k
(7)
Therefore, the estimate, xk|k , of the state vector can be

determined via the UKF using the above equations.
Fault Modelling Procedure

To carry out the sensor fault detection due to calibration,
a method is required to quantify the state deviation. For
this purpose, the following stochastic Auto-Regressive (AR)
model is employed to explain the time evolution of
the sensor bias as extra process state variables:
108
www.SID.ir
A New Fault Tolerant Nonlinear Model Predictive Controller ...
Archive of SID
b k i = bk 1i + N k bi
i = 1,
, nb
Vol. 30, No. 4, 2011
(8)
Where nb denotes the number of faulty sensors, and

indicates a zero-mean white Gaussian noise with
covariance matrices Qkbi. Using the above model for
sensor bias faults, a new augmented state variable vector
Nkbi
x*k = [bTk
xTk ]T is developed by considering the
sensor faults as additional state variables.This assumption

changes the nonlinear model formulations in Eqs. (3) & (6)
to the following nonlinear augmented state model:
x*k = F(x*k 1 , u k , v k ) + w *k 1
z*k = H(x*k , k) + v k
(9)
(10)
where
F(x*k 1 , u k , v k ) = [bTk 1
w *k 1 = [N bki
f (x k 1 , u k , vk )]
w k 1 ]T
H(x*k , k) = [0 H(x k , k)]
(11)
(12)
(13)
NMPC Approach
Since most of the chemical processes are highly nonlinear,
nonlinear extensions to MPC (NMPC) must be applied in
order to provide satisfactory control results. To alleviate
the derivation of control law, the following objective function,
formulated in vector notation, can be considered for
computing the desired control moves [10]:
N
|| ysp (k + p | k) y(k + p | k) ||2 +
J(k) =
p = N1
Fig. 1: Continuous Stirred Tank Reactor.
or Newton-related methods should be considered as

proper candidates [10]. A full Newton method, however,
involves the calculation of the Hessian matrix which is
difficult when the control and prediction horizons become
larger than one [10]. The quasi-Newton algorithm is
suitable when the Hessian matrix is difficult to be derived. In
fact, a positive-definite approximation to the inverse Hessian
matrix can be constructed in quasi-Newton approach [11].
Quasi-Newton Approach
In order to determine the minimum, it is necessary to
apply an iterative search method, specified as follows:
U (i +1) = U (i) + (i) f (i)
(15)
U(i) defines the current iterate of the sequence of

future control inputs, (i) indicates the step size, and f(i)
represents the search direction. The Newton search
direction is given by:
N u 1
|| u(k + p | k) ||2
(14)
p=0
Where ysp denotes the reference or the desired output

vector and y(k + p | k) is the vector of output predictions.
The tuning parameters of the controller are N1, N, Nu,
and . N1 is called the minimum cost horizon, N the prediction
or maximum cost horizon, and Nu the maximum control
horizon. denotes a weighting factor, penalizing changes
in the control inputs.
There are several well-known algorithms to solve
the foregoing NMPC optimization problem. A gradient descent
method doesnt generally have fast convergence and hence
cannot be used in real time applications. Thus, only Newton
f (i) = B(i) G ( U (i) (k) )
(16)
Where B(i) specifies the inverse Hessian and G(U(i) (k))

is the gradient of the cost function with respect to the
unknown control inputs. The quasi-Newton method
approximates the full Newton search direction, B(i),
according to the information embedded in the previous
evaluations of gradient and criterion [10].
CSTR Plant Description
The CSTR plant, represented schematically in Fig. 1,
works under atmospheric pressure [12]. It is, in fact,
a cooling water-jacketed reactor which involves
an irreversible and liquid phase exothermic reaction A B
109
www.SID.ir
Archive of SID
Vol. 30, No. 4, 2011
Table 1: Fault scenario.

The magnitude of the fault
Sensor fault parameter
Bias
Ca
0.05
0.06
Bas in Ca sensor
Bas in V sensor
6.5
5.5
10
Discrete time
0.04
0.02
0
0
10
Discrete time
Fig. 2: Estimating the introduced calibration biases via the CMF-based FDI mechanism.
taking place inside the reactor tank. Negligible heat losses,

constant densities, perfect mixing inside the tank and
uniform temperature in the jacket are assumed.
An NMPC controller is used to regulate the reactor
outlet temperature (T), the liquid volume (V) inside
the reactor tank and the output concentration (Ca) via jacket
flow rate (u1=Fj), outlet flow rate (u2=Fo) and the
inlet flow rate (u3=Fi), respectively. The dynamic equations
describing the system along with the values of process
parameters and steady state conditions are given in [12].
SIMULATION STUDY
MATLAB together with its SIMULINK facilities
have been employed to perform the organized simulation
test studies. The basic time unit is hours (h) and the
sampling time is taken to be equal to 0.0025 h.
A fault test scenario is conducted on the process
in which two CSTR key measured variables (V, Ca)
are corrupted with the corresponding sensor faults. First,
the effectiveness of the designed FDD in estimating sensor
calibration biases is studied. The introduced calibration
biases are summarized in Table 1.
The designed FDD utilizes an Unscented Kalman
Filter (UKF) to estimate the fused measurement state
vector including the bias terms. The controller uses these
estimates to obtain the compensated sensor measurements
and hence keep the controlled variable near its
desired trajectory. Fig. 2 illustrates the ability of the

proposed FDD in estimating the induced sensor biases
in the process, while Fig. 3 compares the performance
of the designed FTC system against a conventional
NMPC controller.
In the next study, aging biases are considered.
The following fault scenario is introduced artificially
in the plant to investigate the ability of the designed
FDD mechanism in estimating the occurred aging biases.
These biases have a zero value in the beginning of
the simulation run but at t=5 and 6 they start increasing with
a constant slope until they reach their final values.
The organized test studies were conducted similar to
the calibration bias case using the proposed FDD
procedure. Fig. 4 illustrates the satisfactory performance
of the designed FDD architecture in detecting and
identifying the introduced sensor biases in the plant. Fig.
5 comparatively demonstrates the performance of the
proposed FT-NMPC against a conventional one in the
presence of aging biases.
The elapsed time during the calculation of control
action in each time step is an important issue which
should be considered in the design of a FTCS.
The average calculation time in the case of calibration
and aging biases is equal to 0.0115 s and 0.0107 s with
a standard deviation of 0.0056 and 0.0056, respectively,
which is much less than the sample time (0.0025 h =9 s).
110
www.SID.ir
Archive of SID
A New Fault Tolerant Nonlinear Model Predictive Controller ...
C NMPC
FT NMPC
Tank level
48
46
44
42
40
0
Set point
0.25
Output concentration
50
Vol. 30, No. 4, 2011
0.2
0.15
0.1
0
10
Discrete time
10
Discrete time
Fig. 3: Performance of the proposed FT NMPC against a conventional NMPC.
Estimated bias
Aging bias
0.06
0.04
Bias in CA sensor
Bias in V sensor
0
-1
-2
-3
-4
0
0.02
0
-0.02
-0.04
-0.06
0
10
10
Discrete time
Discrete time
Fig. 4: Estimating the introduced aging biases via the CMF-based FDD mechanism.
FT NMPC
52
Set point
0.25
Output concentration
Tank level
51
C NMPC
50
49
48
47
46
0
10
Discrete time
0.2
0.15
0.1
0.05
0
0
10
Discrete time
Fig. 5: Performance of the proposed FT- NMPC against a conventional NMPC.
CONCLUSIONS
In this paper, a new fault tolerant NMPC has been
presented which incorporates a FDD mechanism. The
proposed mechanism utilizes a MSDF architecture which
fuses the sensor measurements in augmented state vectors
including bias terms to obtain a combined measurement

vector. Then, a single UKF is used to obtain the final
state estimates based upon the fused observations.
The designed NMPC controller then uses the estimated
fault information to compensate for the ill-measurements
111
www.SID.ir
Archive of SID
and keep the controlled variables near their desired

trajectories. The set of conducted sensor fault scenarios
has illustrated the capabilities of the proposed controller
to cope with the induced faults due to calibration and
aging biases in a CSTR benchmark plant.
Vol. 30, No. 4, 2011
[12] Luyben W.L., Process Modeling Simulation and

Control for Chemical Engineers, 2nd Ed.,
McGraw-Hill, (1989).
Received : Feb. 12, 2010 ; Accepted :Jan. 31, 2011
REFERENCES
[1] Visscher, R., Chernobyl Nuclear Disaster.
http://www.chernobul.co.ul/hoef.html ,(1990).
[2] Maciejowski J. M., Modeling and Predictive Control:
Enabling Technologies for Reconfiguration, Annual
Reviews in Control, 23, p. 13 (1999).
[3] Pranatyasto T.N., Qin S.J., Sensor Validation and
Process Fault Diagnosis for FCC Units under MPC
Feedback, Control Engineering Practice, 9, p. 877
(2001).
[4] Prakash J., Patwardhan S.C., Narasimhan S.A.,
Integrating Model Based Fault Diagnosis with
Model Predictive Control, Ind. Eng. Chem. Res., 44,
p. 4344 (2005).
[5] Venkatasubramanian V., Rengaswamy R., Yin K.,
Kavuri S.N., A Review of Process Fault Detection
and Diagnosis Part I: Quantitative Model-Based
Methods, Computers and Chemical Engineering, 27,
p. 293 (2003).
[6] Julier S.J., Uhlmann J.K., Durrant-Whyte H., A New
Approach for Filtering Non-Linear Systems, in
"Proceedings of the American Control Conference",
p. 1628 (1995).
[7] Julier S.J., Uhlmann J.K., A New Extension of the
Kalman Filter to Non-linear Systems, in "Proc. Of
AeroSense: The 11th Int. Symp. A.D.S.S.C", (1997).
[8] Grewal M., Andrews A.P., Kalman Filtering, Theory
and Practice Using Matlab, Third Edition, John
Wiley & Sons, (2008).
[9] Harris C., Bailey A., Dodd T., Multisensor Data
Fusion in Defense and Aerospace, Journal of Royal
Aerospace Society, 162, p. 229 (1998).
[10] Srensen H., Nrgaard M., Ravn O., Oulsen N.K.,
Implementation of Neural Network Based Non-Linear
Predictive Control, Neurocomputing, 28, p. 37 (1999).
[11] Dennis J.E, Schnabel R.B., Numerical Methods for
Unconstrained
Optimization
and
Nonlinear
Equations, Prentice-Hall, (1997).
112
www.SID.ir

A New Fault Tolerant Nonlinear Model Predictive Controller Incorporating An UKF-Based Centralized Measurement Fusion Scheme

Uploaded by

Copyright:

Available Formats

A New Fault Tolerant Nonlinear Model Predictive Controller Incorporating An UKF-Based Centralized Measurement Fusion Scheme

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

A New Fault Tolerant Nonlinear Model Predictive Controller Incorporating An UKF-Based Centralized Measurement Fusion Scheme

Uploaded by

Copyright:

Available Formats

Iran. J. Chem. Chem. Eng.

Vol. 30, No. 4, 2011