A MPC based Fault Tolerant Control Strategy for Actuator FaultMd Raihan Mallick, Syed A. ImtiazFaculty of Engineering and Applied Science
Memorial University of NewfoundlandSt. John’s, NL, A1B 3X5, Canada
Abstract—A Model Predictive Control (MPC) based fault tol-erant control application has been described. The controller wasapplied to control the concentration and level of a solid crystaldissolution tank. The system was frequently suffering processupset due to difficulty in solid discharge. The controller used anerror calculated from an intermittent signal as a feedforward tothe controller. The intermittent signal posed challenges in modeldevelopment and also for control. An iterative technique wasused to develop the model. The developed controller successfullyeliminated the operational problem.
Keywords: Solid Handling, Actuator Fault, Model PredictiveControl, Fault Tolerant Control
I. INTRODUCTION
Solid handling is often encountered in processing industries
ranging from chemical, materials and minerals to agricul-
tural, food and pharmaceutical. Examples include separation
of different petrochemicals and pharmaceutical products in
crystal form, Ni-Hydrometallurgey, oil sand processing etc.
These areas of manufacturing have a current value exceeding,
according to some estimates, two trillion dollars and a growth
factor of 5 to 10 over the next decade [1]. The vary nature
of solid makes it difficult to dispense and transport solid
between different units of the process. One common problem
is dispensing solid from storage in a controlled way. Solids are
typically stored in hoppers, which are circular or rectangular in
cross-section, with conical or tapering sections at the bottom.
Discharge from hopper takes place through an aperture at
the bottom of the cone. Due to factors such as, mechanical
interlocking, surface attraction, elctrostatic attraction, or pres-
ence of moisture, bridging of particles may take place which
leads to stable arches that impedes the smooth flow of solids.
Other known problems are rat-holing, shaft formation, arching,
piping etc [2]. If a stable arch is formed above the outlet,
difficulties commonly experienced are large fluctuation in flow
and in the extreme case no discharge at all from the hopper.
This often leads to operational problems in the downstream
processing units. Almost all of the solutions considered are
design changes in hopper, for example, changing the slope of
the conical section, different hopper geometry or attachment
of vibrators to the wall. Even after design change in many
cases the problem cannot be eliminated completely. Especially
the problem gets complicated when precise flow needs to be
maintained. It is important to manage this problem opera-
tionally as well. The motivation of this research comes from
0This work was financially supported by Research and DevelopmentCorporation (RDC).
a petrochemicals plant where solid crystals were dissolved
in a tank with water. The process used to run into frequent
operational problems when the crystals did not flow smoothly
from the hopper. In this paper we model the solid discharge
problem as an actuator problem propose a fault tolerant control
strategy that is able to deal with this problem. The proposed
fault tolerant strategy was implemented in the process which
effectively eliminated the problem.
A. Problem Description
A simplified process diagram for the system is shown in
Fig. 1. In this system solid crystals (for our purpose we will
call CRYSTAL-A) is dissolved in a tank with water. Water is
pumped into the tank under flow control. CRYSTAL-A is fed
to the dissolution tank from a hopper using a rotary feeder. The
feed rate of solid crystals to the mixing vessel is controlled
by the speed of the rotary feeder (RPM). The Water Level
in tank (LI) and the concentration of the liquid going out
of the tank (DI) are measured variables. The main control
objectives of the system are to prevent any overflow (maintain
the tank level) and maintain the concentration at desired value.
Under the existing control strategy, two PID controllers are
used to meet these objectives, the concentration of the outlet
stream is controlled by manipulating the rotary valve rpm,
while the flow controller under cascade control maintains the
tank level. However, the concentration at the outlet is subject
to frequent large disturbances when the operators have to take
control of the process and control the process manually during
this abnormal condition. The root cause of the disturbance
is the difficulty in solid dispensing. Occasionally because of
the variation in moisture content the solid gets lumped in
the rotary feeder. As a result solid does not dispense from
the feeder uniformly. After a while when the lump gets too
big it falls into the tank creating a big disturbance in the
concentration which causes further problem in the downstream
process. During these abnormal conditions the rotations of the
rotary feeder is not correlated with the solid discharge as a
result manipulating only rpm to control the concentration is
not effective. Under the existing control strategy this was a
frequent occurrence which caused severe operational problem.
Therefore it was important to design a controller that is able
to address this problem proactively. We designed a model
predictive controller based strategy to deal with operational
problem.
3777978-1-4244-8165-1/11/$26.00 ©2011 IEEE
Fig. 1. Process flow diagram with the existing control strategy
B. Why PID Failed?Given the overwhelming popularity of PID controller in
process industry, one relevant question is, why PID based
techniques were not used for managing the fault. Several
alternative PID control structures can be thought of which in
theory should have better success in rejecting the disturbance.
One possible control structure is to use a cascade control in
the solid feed. Instead of directly using the rpm to control
the density, in the slave layer rpm can be used to control
the solid flowrate, while the setpoint of the solid flowrate
can be cascaded to the concentration. This should allow early
rejection of the any disturbance in the solid flow. Another PID
based alternative structure can be to use the solid flowrate as a
feedforward to the controller. However, in practice these PID
control structures could not be used for several reasons. First,
there was no continuous measurement for the solid flowrate
to the tank. The solid flow rate is an intermittently calculated
signal from the ‘loss in weight’ of the hopper shown in Fig.
2. The hopper is mounted on a load cell which gives the
‘loss in weight’ of the hopper. Once the sensed weight reaches
the hopper refill level (low weight) the hopper is recharged,
during the recharge period the ‘loss in weight’ calculation is
not valid. Therefore the signal has missing periods of data for
the recharge events. The ‘mass flowrate’ signal after applying
‘zero order hold’ for each refilling periods is shown in Fig.
3. PID controller does not have a built in mechanism to deal
with this type of intermittent signal. Essentially during ‘zero
order hold’ period it is a constant signal and under cascade
control the controller will not take any action for the changesin density. Therefore the controller will be inactive for half of
the time. Secondly, prediction error of solid discharge could
constitute a good feedforward signal. However, PID controller
does not have the prediction capability therefore it requires
Fig. 2. Solid flow rate calculated from ‘loss in weight’ of the hopper
Fig. 3. Fault in the solid discharge under PID control
additional customization in the Distributed Control System
(DCS) to calculate dynamic prediction error. In addition to
that, the Single Input Single Output (SISO) nature of the PID
controller is also not ideal for rejecting the disturbance due
to actuator fault. Once the actuator (rotary control valve) is at
fault it has limited effectiveness. Therefore it is important to
use other handles which can quickly bring the concentration
within the limit. However, because of the SISO nature of PID
controller it cannot use other handles to reject the disturbance.
Therefore we used a model based controller, Dynamic Matrix
Controller (DMC) with additional fault handling capability to
manage this fault.
II. FAULT HANDLING IN DYNAMIC MATRIX CONTROL
There are several variants of DMC: ‘original’ DMC [3],
DMC with least squares satisfaction of input constraints
[4]. DMC with constrained linear programming optimization
(LDMC) [5], DMC with constrained quadratic programming
optimization (QDMC) [6], discrete state space version of
DMC [7]. A fault tolerant variant of DMC able to handle
sensor faults is proposed by [8]. These variants use different
optimizers and vary in terms of the structure of the prediction
models, however use the same prediction and minimization
of error concept. In our application we used Dynamic Matrix
Controller (DMC) and modified it for fault rejection. Dynamic
Matrix Control is a variant of Model Predictive Control that
uses a step response model to predict the future behavior of
the plant. Excellent reviews on MPC can be found in [9], [10].
The control input is calculated in a way that it minimizes the
error between the prediction and the desired set point. In this
application we used the following optimization at each time
step.
3778
Fig. 4. Step response model for DMC
minimize J(Δu) = (r− y)TQ(r− y) +ΔuTRΔu. (1)
y is the predicted value given by,
y = y∗ +A Δu. (2)
Minimization of the above objective function gives the
following control law,
Δu = (ATQA+R)−1ATQT (r − y∗). (3)
In the current application we modeled the solid discharge
problem as an actuator fault. In order to detect the fault
prediction error was calculated from the difference of the
predicted solid flow based on the rpm setpoint of the rotary
feeder and the solid flowrate calculated from ‘loss in weight’
of the hopper. Here we will give the modified DMC for
the error signal and defer the application problem due to
the intermittent nature of the error signal in the following
section. In this case, the uncontrolled output predictions have
an additional term of this disturbance,
y∗t+p = Zt+p +n∑
i=p+1
aiΔut+p−i + νt+p. (4)
The disturbance in this case is a multiple of input signal
that goes through the process,
νt+p =n∑
i=p+1
aiδt+p−i. (5)
where δt+p−i is the prediction error calculated from the
difference of predicted solid discharge and the calculated solid
flowrate.
III. CONTROLLER DESIGN AND MODELING
The core of the Dynamic Matrix Controller (DMC) design
is the 2×2 matrix with two Manipulated Variables (MVs) (i.e.,
Water Flowrate and RPM of rotary valve) and two Controlled
Variables (CVs) (i.e., Concentration and Tank Level). The
design philosophy for the controller is to manipulate Water
Flowrate setpoint as the primary handle for the level control. In
order to control concentration the RPM setpoint will be used as
the primary handle and Water Flowrate as a secondary handle.
The step response model for DMC is shown in Fig. 4. In an
ideal situation this structure should be sufficient for controlling
these two variables. However, the system frequently suffers
large disturbance in concentration due to the solid discharge
problem discussed in the previous section. This warrants
special fault tolerant configuration of the controller. The error
between the measured solid flowrate and the predicted solid
flowrate was used to early detect the fault in solid discharge.
The dynamic ‘Solid flow vs. RPM’ model was used to predict
the solid flow rate at anytime based on the RPM of the rotary
valve. Subsequently the predicted solid flowrate was subtracted
from the solid flow signal from the load cell, this gives the
δt+p−i in Equation 5. The calculated error signal was used
as a feedforward to the concentration control. As soon as the
fault initiates the error, signal magnitude (δi) keeps increasing,
the effect of the fault gets reflected on the concentration
prediction. The controller will first use the primary handle
(RPM) in order to keep the concentration within limit, if the
error keeps growing it will use the secondary handle (Water
Flowrate), which has a faster dynamics with respect to the
concentration.
A. Modeling using intermittent signalThe above fault tolerant formulation of the controller re-
quires two additional dynamic models ‘Solid Flow vs. RPM’
and ‘Concentration vs. Solid Flowrate’. However, these models
cannot be identified directly because of the intermittent nature
of the solid flow signal. The intermittent signal poses problem
in the modeling phase as well as in the control phase. We used
the following iterative technique to estimate the ‘Solid Flow
vs. RPM’ model parameters.
• Step 1 The solid flowrate signal is linearly interpolated
for the periods the signal values are missing; a model
between the rpm and the solid flowrate is estimated.
• Step 2 The model calculated from Step 1 is used to
predict solid flow. The missing values in the solid flow
signal are replaced by the predicted values.
• Step 3 Steps 1 and 2 are repeated until there is no
significant change in the model.
Subsequently the reconstructed ‘Solid Flowrate’ signal was
used to estimate the ‘Concentration vs. Solid Flowrate’ model.
The solid prediction error will have the same relationship with
concentration as the ‘Solid Flowrate’. The estimated model
was used to predict the disturbance in concentration for error
in the solid flowIV. RESULTS AND DISCUSSION
The controller was successfully implemented on the system.
The unique feature of the controller is the calculation of
the error in solid flow. However, the intermittent nature of
solid flow signal posed difficulty in online error calculation.
Under the current control strategy when the measurements are
unavailable, error value is frozen to its previous value. The
strategy worked well for the controller. The controller showed
3779
Fig. 5. Concentration control after the implementation of DMC
good performance in controlling the concentration and the
level. Since the implementation of the controller, operators
never had to intervene for the solid discharge problem. The
concentration signal under current control strategy is shown
in Fig. 5. Two instances when the solid flow was not smooth
are marked in the figure. In the first event the disturbance
is reflected in the concentration, however the controller was
able to keep it within the acceptable limit. In the second case,
the error was detected early and the controller took proactive
action. As a result of the proactive action the fault had very
little effect on the concentration.
V. CONCLUSIONS
The application of a DMC based fault tolerant control strat-
egy is described. The novelty of the controller is it is simple
to apply in process yet effective. The other contribution of the
research is the modeling and control strategy when a signal has
missing values. The proposed controller successfully managed
the actuator fault in a solid handling process.
VI. ACKNOWLEDGMENT
The authors would like to thank Research and Development
Corporation (RDC) for financial support.
REFERENCES
[1] A. Gani, P. Mhaskar, and P. D. Christofides, “Handling sensor malfunc-tions in control of particulate processes,” Chemical Engineering Science,vol. 63, no. 5, pp. 1217 – 1229, 2008, control of Particulate Processes.
[2] A. S. Foust, L. A. Wenzel, C. W. Clump, L. Maus, and L. B. Adersen,Principles of Unit Operations. John Wiley and Sons Inc, NY, 1980.
[3] C. R. Cutler and B. L. Ramaker, Dynamic matrix control- A computercontrol algorithm. Paper WP5-B, 1980.
[4] D. M. Prett and C. E. Garcia, Fundamental Process Control.Butterworths-Heinemann, Boston, MA, 1988.
[5] A. M. Morshedi, C. R. Cutler, and T. A. Skrovanek, “Optimal solutionof dynamic matrix control with linear programing techniques (ldmc),”in American Control Conference, 1985, june 1985, pp. 199 –208.
[6] C. E. GARCIA and A. MORSHEDI, “Quadratic programming solutionof dynamic matrix control (qdmc),” Chemical Engineering Communica-tions, vol. 46, no. 1-3, pp. 73–87, 1986.
[7] P. D. M. and R. D. Gillette, Proc. Joint Automatic Control Conj. SanFrancisco,CA, 1980.
[8] P. M. Marusak, “Disturbance measurement utilization in easily re-configurable fuzzy predictive controllers: sensor fault tolerance andother benefits,” in Proceedings of the 7th international conference onRough sets and current trends in computing, ser. RSCTC’10. Berlin,Heidelberg: Springer-Verlag, 2010, pp. 551–559.
[9] S. J. Qin and T. A. Badgwell, “A survey of industrial model predictivecontrol technology,” Control Engineering Practice, vol. 11, no. 7, pp.733 – 764, 2003.
[10] K. R. Muske and J. B. Rawlings, “Model predictive control with linearmodels,” Aiche Journal, vol. 39, pp. 262–287, 1993.
3780