reconfigurable of active fault tolerant control for nonlinier actuator and sensor fault - sami...
TRANSCRIPT
-
8/10/2019 Reconfigurable of Active Fault Tolerant Control for Nonlinier Actuator and Sensor Fault - Sami Patton
1/6
Reconfigurable Fault-Tolerant Control of Linear
System with Actuator and Sensor Faults
Katherin Indriawati, Trihastuti Agustinah, Achmad Jazidie
Department of Electrical Engineering
ITS
Surabaya, [email protected]
AbstractThis paper presents an active fault-tolerant control
for linear system in case of actuator and sensor faults where these
minor faults lead to degraded performance of the system. Three
steps are proposed to achieve fault tolerant control based on
simplified analytical redundancy. Firstly, a bank of linearobserver is proposed to estimate the actuator and sensor faults by
modeling a descriptor LTI system using the SVD technique.
Secondly, the estimated faults are used to design a fault decision
scheme to detect the faults correctly. Thirdly, a reconfigurable
fault-tolerant control scheme is designed by using the estimated
faults to compensate the fault effects on controller performance.
Simulation on the three tank system is given to illustrate the
performance of the proposed method.
Index TermsObserver, SVD, reconfigurable FTC.
INTRODUCTION
Control systems which have the ability to accommodate
component (actuator or/and sensor) failures automatically arecalled Fault Tolerant Control Systems. These systems are able
to maintain the stability and the desired performance of the
system in the presence of such failures [1]. FTCS is needed to
increase reliability and automation level in modern engineeringsystems. Generally, FTCS can be performed by passive
methods or by active methods. In passive methods, controller is
fixed and can be designed using robust control techniques to
ensure that a closed-loop system remains insensitive to certain
faults. This approach needs neither on-line fault information
nor controller reconfiguration, but it has limited fault-tolerantcapabilities [1]. On the other hand, in active methods, a new
control system is redesigned by using on line fault information
in order to maintain the stability and acceptable performance ofthe entire system, or in circumstances, to achieve accepted
degraded performance. Active FTCS are often referred to as
reconfigurable control. The design of an active FTCS requiresquick but effective fault detection and isolation (FDI) scheme
for adequate decision making that refers to the task of inferring
the occurrence of faults in a system.
A general approach of active FTCSis based on analytical
redundancy. Noura et al. in [2] has presented this approach for
discrete linear systems. They treat the sensor faults as the
actuator faults, then it used to estimate all faults by solving a
difference matrix equation. The estimated sensor faults are
used to modify the nominal control law to compensate for the
effects of the sensor faults. In [3], both sensor and actuator
faults are isolated and estimated by using of a unique structured
residual generator. The residual generator consist of a bank of
unknown input observer that each observer may be used to
detect a single fault. In [4] linear time invariant systems with
sensor faults are transformed into descriptor system and then
the proportional plus derivative observers are used to
simultaneously estimates the states of the system and the
sensor faults. However, all of those researches assume the
perfect condition of the plant regime and there is no
environmental noise in measurement system, which implies
that FDI algorithm detect faults instantaneously and always
correct [5].
To develop an active FTCS, it is required to examine
reconfigurable control and FDI to ensure that they can work inharmony. The kind of information needed from a FDI should
be examined to achieve a reasonable control strategy. An
imperfect FDI algorithm may not only result in loss of
performance, but also instability for the overall FTCS. This
paper focuses on active FTC based on analytical redundancy
which combines the functions of FDI and reconfigurablecontrol in noisy environment. Studies to this area is fewer than
other areas of fault-tolerant control research [6]. Furthermore,
many challenging issues still remain open for further research
and development for this area [1].
Generally reconfiguration scheme for the linear control
system that has been proposed in the scientific paper is only
detecting one sensor or actuator faults occur at a particulartime, as in [2] and [3]. This paper presents the results of a
simulation study for linear control system reconfiguration
scheme that tolerant of sensor and actuator faults that occur
sequentially. Furthermore, the measurement noise influences
are also considered in designing of FDI algorithm in order to
minimize the occurence of false alarm and missed alarm.
This paper is organized as follows. In section II, the
nominal linear control system and the reconfigurable control
probem dealing with actuator and sensor faults are presented.
In section III, the strategy of reconfigurable linear control
2013 IEEE International Conference on Control System, Computing and Engineering, 29 Nov. - 1 Dec. 2013, Penang, Malaysia
978-1-4799-1508-8/13/$31.00 2013 IEEE 22
-
8/10/2019 Reconfigurable of Active Fault Tolerant Control for Nonlinier Actuator and Sensor Fault - Sami Patton
2/6
based on a bank of observers is propo
example of a three-tank system and its sim
given in section IV. Finally, concluding resection V.
PROBLEM FORMULATIO
Nominal Control
Consider a discrete linear time invarigiven by the following state space represent
=
+=+
)()(
)()()1(
kk
kkk
Cxy
BuAxx
where xRn, uR
pand yR
mare the state
input, the output vector, respectively. ARnx
Rmxn
are the state, the control, and the
respectively. The number of outputs m t
reference input vector yrdo not exceed the
inputs due to controlability requirement.
The nominal control system for that plstructure with integrator as shown in Figu(hp) represents the vector of the masureme
required to follow the reference input vectorh)
represents the vector of the unmeasure
nominal control system take into account t
(U0, Y0). The state space representation of tshown in Figure 1 is
[ ]
=
+
+
=
+
+
)(
)(0)(
)(0
0)(
)(0
)1(
)1(
,
,
,1
,
kz
kxCky
kyIT
B
kz
kx
ICT
A
kz
kx
pq
rps
pn
mpps
pn
where C1 is row of matrix C related to the
is the sample period to be chosen properly
matrix of dimension pxp, and 0n,p is
dimension nxp. The nominal feedback co
system is computed by
[ ]
==
)()()()( 21
kzkxKKkXKku
K = [K1 K2] is the feedback gain matrix obta
techniques such as a pole placement t
quadratic optimization, and so on [8][9][10][
ed. A numerical
ulation results are
arks are given in
ant (LTI) systemtion
(1)
ector, the control
, BRnxp
, and Coutput matrices,
hat can track a
number of control
ant uses feedbacke 1,where y1R
h
nt outputs that are
yrwhile y2R(m-
ent output. The
e operating point
he control system
)(ku
(2)
ontrolled state, Ts
, Ip is an identity
a null matrix of
ntrol law of this
(3)
ined using several
echnique, linear-
11].
Fig. 1 Nominal tracking cont
Reconfigurable ControlThe designed control syst
control signal automatically
component faults so that the pl
used algorithm in designing of
the suitable model with the sim
To achieve a control syste
sensor faults, the proposed m
recalculation of the control si
type. The block diagram of th
is shown in Figure 2. The n
system is given by
)()()( ukukuku addaddan ++=
where un (k) represents the n
represents the additive cont
actuator faults, and uadds (k)
signal to compensate the senso
The control signal reconfi
process in order to detect
commonly known as fault
Here, the FDI proposed met
mathematical model (analyti
observer to generate residual.
make the observer in a simplhandle both actuator and senso
order to minimize the occure
alarm due to noise measure
modification algorithm of the
to the FDI threshold values.
RECONFIGURABLE LI
Residual is the differenc
measurement value and the s
operating condition. It is assu
taken by an actuator-sensorobserver needed for the FDI s
of actuator-sensor pairs con
observer then are used in the
about where may the faults be
technique. Based on that de
occured fault is estimated and
order to recalculate the control
ol with feedback structure [7]
m conducts reconfiguration of
in order to accommodate the
ant still operates as desired. The
that control system is based on
ple structure and techniques.
that tolerant from actuator and
ethod of this paper consist of
gnal based on the occured fault
reconfigurable control system
ew control law applied to the
)(k (4)
ominal control signal, uadda (k)
ol signal to compensate the
epresents the additive control
faults.
guration needs fault diagnosis
nd isolate the occured fault,
detection and isolation (FDI).
od in this paper is based on
al redundancy), that is using
Thus the problem is how to
e manner that can be used tor faults at once. Furthermore, in
ce of false alarm and missed
ment, this paper propose the
Shewhart control chart related
NEAR CONTROL DESIGN
e of the considered quantity
ame quantity value in normal
ed that each controlled variable
air. Therefore, the number ofstem is equal with the number
sidered. The results of each
FDI system to make a decision
occured by means of statistical
cision, the magnitude of the
the estimation result is used in
signal.
2013 IEEE International Conference on Control System, Computing and Engineering, 29 Nov. - 1 Dec. 2013, Penang, Malaysia
23
-
8/10/2019 Reconfigurable of Active Fault Tolerant Control for Nonlinier Actuator and Sensor Fault - Sami Patton
3/6
Fig. 2. Reconfigurable fault-tolerant control scheme [7]
Actuator Fault Compensation
The state-space representation of a system that may be
affected by ith
actuator fault is
1 1 0, 0, 0, 0 0,
(5)
where is the magnitude of ithactuator fault and is the related fault matrix.
Detection, isolation even estimation of the actuator fault
magnitude is conducted by designing an observer that is able
to generate fault signal estimate of ith
actuator,: 0 if there is no actuator fault 0if there is actuator fault
The observers in this paper are developed by modifying (5)
so that
to be component of state vector. The
modification results is [7]: 1 (6)
where
0 0 0 0 0 ; 00 ;
0 0 00 0 0; 1;
00 00 ;
1
The estimation of the actuator fault magnitude isconducted by estimating the state vector , that is the lastcomponent ofas shown in (6).
The next problem is obtaining an equation of . Thisproblem can be solved by means of singular value
decomposition (SVD) technique toward matrix oncondition that it is of full column rank [12], with the result that
it can be declared in a product of three matrix:
0 (7)
TiandMirepresent orthonormal matrices, and Siis a diagonal
nonsingular matrix.
Substituting (7) to (6) and dividing matrix Ti to be two
parts, Ti= [Ti1 Ti2] leads to
1 0 (8)where
; ; ; ; (9) ; ;
is pseudo-inverse of matrix
.
The ith
actuator fault compensation observer works using
the first equation in (8). Therefore, the each observer produce
results of the state vector estimate , the integral error vectorestimate , and the ith actuator fault vector estimate , byusing the free-fault control signal u = ufsf, and the
measurement output signal y. Note that each is notsensitive to setpoint changes as well as to faults of the other
couples of actuator-sensor
The result of then is used in statistical test in order toproduce alarm. Based on the behavior of , the statistical testis done by adopting Shewhart control chart of Statistical
Process Control (SPC) method, that is evaluating each sample
ofto determine wheter it is in the in-control area or not. Thein-control area is the region that has two boundary limits: the
upper control law (UCL) and the lower control law (LCL)
defined by
(10a) (10b)where is mean value and is deviation standard value ofthe successive sample data set of
in a windowing. Length
and overlap of the windowing determine false alarm rate andmissed alarm rate. In this case, those both parameters of the
windowing is determined by trial and error by reference to the
value of signal-to-noise ratio (SNR) of the measurement. is a constant determined by reference to the ratio value
between fault magnitude and noise measurement. If the
sample value of at time instant k is inside the in-controlarea, then the indicating signal Iaiat that time is zero. On the
other hand, if the sample value of at time instant k isoutside the in-control area (out-of-control), then the
2013 IEEE International Conference on Control System, Computing and Engineering, 29 Nov. - 1 Dec. 2013, Penang, Malaysia
24
-
8/10/2019 Reconfigurable of Active Fault Tolerant Control for Nonlinier Actuator and Sensor Fault - Sami Patton
4/6
indicating signal Iai at that time is one. Effort to minimize the
occurrence of false alarms is done by way of the alarm signal
on if and only if there are four successive value ofIaiequal toone. Next, to each the actuator being considered, the alarm
signal displays 0 if there is no fault but 1 if there is a fault.
The estimated fault signals of the each actuator are
combined into an actuator fault vector:
(11)The additive control signal for compensating all actuator
faults is:
)()( 1 kfFBku aaadda
= (12)
where matrixFais the overall fault matrix (commonly equal to
matrix C1).
Sensor Fault Compensation
The output equation model with feedback control system
in (9) should be changed in case of sensor faults, i.e.:
0, (13)Thus, to detect and to estimate the magnitude of sensor
faults can be conducted by means of (13), using the
measurement output vector and the state vector estimate. As
mentioned above, the each developed observer produce results
of the state vector estimate
, the integral error vector
estimate , and the ith actuator fault vector estimate , byusing the free-fault control signal u = ufsf, and themeasurement output signal y. The state vector estimate
represent the free-fault condition state (the nominal condition
state). If there is an ith
sensor fault which is the couple of an ith
actuator, then the components value of is not anymoreequal with value of measurement results yi. The difference is
only occured at one time instant tsf, as a result of the nominal
control signal tries to bring the steady state error back to zero.
The difference turned out to be an estimate of the ith
sensor
fault magnitude,
. By using the assumption that the fault
sensor did not get better over time, but it may be gettingworse, then the sensor fault magnitude may not decrease with
increasing time. To minimize the possibility of false alarms
caused by outliers, it is usedjsuccessive samples ofythat are
compared with on hold for three step sample times from tsf.The determination ofcan be defined as: for 0 for (14)
where Ci is a null vector except ith
component equals to 1.
There is a trade off in determining the value of j, the delay
time detection and the false alarm. Note that each issensitive to setpoint changes, but is not senstivite to faults of
the other couples of actuator-sensor.
As with the actuator fault detection, the results in thenis used on satistical test in order to generate the sensor alarm.
Based on the characteristic of , the statistical test isconducted by using deviation standard of the windowingsample data set of . The two types of threshold used are: if
if (15)Di is detectability threshold that its value depends on the value
of SNR. The smaller the value ofDi, the higher occurrence of
false.
is a scalar that its value is influenced by the related
setpoint changes. If the sample value of at time instant kexceed the threshold Ti, the the indicating signalIsiat that timeis equal to 1. Next, to each the sensor being considered, thealarm signal displays 0 if there is no fault but 1 if there is a
fault.
The estimated fault signals of the each sensor are
combined into a vector of the sensor fault:
(16)The additive control signal for compensating all sensor
faults is:
)(~
)()( 21 kfKkfFKku sssadds += (17)
where sf~
is the integral of ssfF
Therefore, the free fault control signal is
ufsf(k) = un(k) + uadds(k) (18)
Note that vector ufsf is the control signal which is used by
the observer in the FDI system..
APPLICATION EXAMPLE
To illustrate the proposed method, a three tank benchmark
system is considered. The dynamical model of the system is
given by [7]
2013 IEEE International Conference on Control System, Computing and Engineering, 29 Nov. - 1 Dec. 2013, Penang, Malaysia
25
-
8/10/2019 Reconfigurable of Active Fault Tolerant Control for Nonlinier Actuator and Sensor Fault - Sami Patton
5/6
where
(m, n= 1,2,3 mn)
The variables l1, l2, l3 denote the level i
respectively; qmn represents the flow rate
while q20is the outflow rate at tank 2.The d
numerical values of the plant model para
Table 1. The controlled variables are l1
manipulated variables are q1and q2.The liplant can be derived in the equilibrium poin
0.325]T10
-4(m
3/s);[0.4 0.2 0.3]
T(m))
TABLEI.PARAMETER VALUES OF THE THRE
Parameter Symbol
Tank cross sectional areaInter tank cross sectional area
Inter tank outflow coefficient
Outflow coefficient at tank 2Maximum flow rate
Maximum level
SSp
13=3220qmax
lmax
0.05x1
0.5
0.61.2
0.6
TABLEII.
ISOLATION TIME OF TH
Faults on Noise standard
deviation of 10-4m
Nois
deviat
Occurence Isolation Occurenc
Pump 1 100 s 107 s 100 s
Sensor 1 50 s 51 s 50 s
Pump 2 400 s 407 s 400 s
Sensor 2 500 s 501 s 500 s
The simulation is intended to determine
developed control system to overcome pu
and level sensor faults as well, by
reconfiguration scheme. Therefore, the simu
not the component functional failure, but the
with small severity. The actuator faults ar
loss of effectiveness of the actuator pumps
degradation of 0.3 on the output control s
faults are simulated as a constant offset or
the piezoresistif level sensor so that the fa
used by the controller is equal to l+ 0.03.
(19)
tank 1, 2, and 3
rom tank m to n
escription and the
eters are listed in
and l2 while the
ear model of thes (U0;Y0) = ([0.35
TANK SYSTEM
Value
154 m2
0-5m2
75x10-4m3/s
2 m
FDI
e standard
ion of 10-3m
Isolation
119 s
51 s
427 s
501 s
the ability of the
p actuator faults
means of the
lated fault types is
component faults
simulated as the
, by using a gain
ignal. The sensor
biasof 0.03 m on
ulty measurement
The measurement noise
Gaussian distribution. The st
distribution that used in the si
The simulation results are
concluded that the proposed F
environment. Note that the lar
the larger the delay time of th
fault detection time is largertime. This is because the actua
on the dynamics of the syste
fault detection process is don
direct measurements and the s
the controller from reacting.
Figure 3 shows the syste
change in the reference value
faults, i.e the pump 1 at t = 10
It is noticed that the respons
system (with FTC) is better t
system (without FTC). Altho
still able to track the referenc
and the overshoot of its outp
system outputs with FTC. Th
additional control signal uaddasystem capables of compensate
In the second simulation
faults appears in the tank 1 at i
tank 2 at instant 500 s. The
illustrated in Fig. 4. It can be
the real level follows the set
classical control law. It is ca
used to generate the control s
the real situation. The fault is i
for sensor 1 and sensor 2reconfiguration approach pres
the system in the presence of s
To know the ability of t
dealing with more than one
faults occurrence is simulated
occured before the actuator fa
the tank 2, the actuator fault is
fault. Figure 5 illustrates the
that the developed control
compensate more than one fau
first occured. The analysis of t
also emphasizes the better perfto the classical control, i.e. 4
FTC for the level of tank 1; as
without FTC for the level of ta
is assumed to be zero-mean
andard deviation of the noise
ulations is 10-3
m and 10-4
m.
illustrated in Table 2. It is
I is able to work well in noisy
ger the noise standar deviation,
e FDI. In addition, the actuator
than the sensor fault detectionor fault detection process relies
m response. While, the sensor
e by comparing the results of
ate estimate thus be preventing
m responses when there is a
and followed with the actuator
0 s and the pump 2 at t = 400 s.
of the reconfigurable control
an that of the classical control
gh the system without FTC is
e value, but the time response
uts are larger than that of the
s, it has been proved that the
rom the reconfiguration control
the actuator faults properly.
experiment, an abrupt sensor
nstant 50 s, and followed in the
results of this experiment is
noticed that with FTC method
oint; it is not the case for the
sed the tank level information
gnal is not in accordance with
solated at instant 51 s and 501 s
respectively. Therefore, therves the dynamical behavior of
nsor faults.
e proposed control system in
fault, the sensor and actuator
sequentially. The sensor fault is
ult in the tank 1. In contrast to
occured first before the sensor
simulation results which prove
system has the ability to
lts, no matter what type of fault
he integral absolute error (IAE)
ormance of the FTCS compared01 with FTC and 428 without
well as 200 with FTC and 214
k 2.
2013 IEEE International Conference on Control System, Computing and Engineering, 29 Nov. - 1 Dec. 2013, Penang, Malaysia
26
-
8/10/2019 Reconfigurable of Active Fault Tolerant Control for Nonlinier Actuator and Sensor Fault - Sami Patton
6/6
Fig. 3. The output measurement responses when the actuator faults occurred
in the tank 1 and the tank 2
Fig. 4. The real output responses when the sensor faults occurred in the tank 1
and the tank 2
Fig. 5. The system responses when the actuator and sensor faults occured
sequential (sensor fault then actuator fault for tank 1; actuator fault thensensor fault for tank 2)
CONCLUSION
After The simulation example of the fault tolerant control
of linear system with more than one fault has been conducted.
The three tank system is used to illustrate the abilities of the
proposed method to compensate for both sensor and actuator
faults. A bank of observers has been developed to detect,
isolate, and estimate faults of the sensor-actuator pairs. Based
on the simulation results, it is concluded that the control system
with FTC has the output responses which are closer to the
nominal outputs rather than that of the system with the classical
control law.
REFERENCES
Y. Zhang, J. Jiang, "Bibliographical review on reconfigurable fault-tolerant control systems", Annual Reviews in Control, vol. 32,issue 2, pp. 229-252, December 2008.
H. Noura, D. Sauter, F. Hamelin, D. Theilliol, Fault-tolerant control
in dynamic systems: Application to a winding machine, IEEEControl Syst. Mag.,vol.20, pp. 33-49, 2000.
D. Theilliol, H. Noura, J.C. Ponsart, "Fault diagnosis andaccommodation of a three-tank system based on analyticalredundancy", ISA Transactions, vol. 41, no. 3, pp.365382,2002.
Z. Gao, H. Wang, Descriptor observer approaches for multivariablesystem with measurement noises and application in faultdetection and diagnosis, Systems & Control Letters, vol. 55, pp.304313, 2006.
M. Mahmoud, J. Jiang, Y.M. Zhang, "Active fault tolerant controlsystems: Stochastic analysis and synthesis", Lecture notes incontrol and information sciences, vol. 287, Berlin, Germany:Springer, 2003.
R.J. Patton, Fault-tolerant control: The 1997 situation (survey),Proseding IFAC SAFEPROCESS'97, Hull, U.K., vol.2, 1033-1055, 1997
H. Noura, D. Theilliol, J.C. Ponsart, A. Chamseddine,Fault-tolerantControl Systems: Design and Practical Applications, Springer-Verlag London, 2009.
K. Ogata,Modern Control Engineering- 4thed., Prentice Hall, 2006.
R.L. Williams-II and D.A. Lawrence, Linear state-space controlsystems, John Wiley & Sons, Inc., 2007.
K.J. Astrom, R.M. Murray, Feedback systems: an indtroduction forscientists and engineers, Princenton University Press, 2008.
D. Xue, Y. Chen, D.P. Atherton, Linear feedback control: analysisand design with MATLAB (Advances ind design and control),Society for Industrial Mathematics, first ed., 2008.
A. Bassong-Onana, M. Darouach, G. Krzakala, "Optimal estimationof state and inputs for stochastic dynamical systems withunknown inputs", Proceedings of International Conference onFault Diagnosis, pages 267275, Toulouse, France, 1993.
0 100 200 300 400 500 600 700 800 900 1000
0.2
0.25
0.3
0.35
0.4
0.45
0.5
time (s)
level(m)
Tank 1
Tank 3
Tank 2
without FTC
with FTC
without FTC
with FTC
0 100 200 300 400 500 600 700 800 900 10000.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
0.5
time (s)
level(m)
Tank 1
Tank 2
Tank 3
with FTC
without FTC
with FTC
without FTC
0 100 200 300 400 500 600 700 800 900 1000
0.2
0.25
0.3
0.35
0.4
0.45
0.5
time (s)
level(m)
measured
measured
real
Tank 2
Tank 1
Tank 3
real
2013 IEEE International Conference on Control System, Computing and Engineering, 29 Nov. - 1 Dec. 2013, Penang, Malaysia
27