You are on page 1of 14

Article

Transactions of the Institute of


Measurement and Control
Robust optimal Hi/HN fault detection 2018, Vol. 40(3) 719–732
Ó The Author(s) 2018

and its applications in a TGSCM Reprints and permissions:


sagepub.co.uk/journalsPermissions.nav
DOI: 10.1177/0142331216665687
control system journals.sagepub.com/home/tim

Maiying Zhong1 and Shusheng Li2

Abstract
This paper aims to deal with the problem of fault detection in a closed-loop mode for a three-axis gyro-stabilized camera mount under the consider-
ation of unknown disturbances. First, the influences of potential actuator and sensor faults are analysed and, based on this, the faults model as well as
an equivalent additive fault input are introduced. For the purpose of fault detection, an observer-based fault detection filter is considered to generate
the residual. The sensitivity of the residual to fault is evaluated by the non-zero singular values of the transfer function from fault to residual, whereas
the robustness to disturbance is represented by the HN-norm of the transfer function from disturbance to residual. To obtain a maximum sensitivity/
robustness ratio criterion, the design of fault detection filter is formulated as an Hi/HN optimization problem and the unified solutions for the optimiza-
tion criterion are given. For residual evaluation, a window evaluation function is developed and an analysis of threshold selection is given to reduce the
conservativeness of fault detection. Finally, experimental results are presented to demonstrate the effectiveness of the proposed method. It is demon-
strated that the proposed method can effectively detect the provided fault as soon as possible.

Keywords
Fault detection filter, Hi/HN optimization, three-axis gyro-stabilized camera mount.

Introduction However, observer-based FD has received much attention


during the past three decades and significant progress has
In order to improve the resolution and overlapping degree of been achieved (He et al., 2013a, 2013b; Hossein et al. 2013;
images for an airborne remote sensing system, three-axis Huang et al., 2014; Li and Zhong, 2013; Li ZL et al, 2015;
gyro-stabilized camera mounts (TGSCMs) has been widely Thumati and Halligan, 2013; Wang and Qi, 2015; Yin and Li,
used to maintain and stabilize the imaging loads (aerial cam- 2016; Yin and Huang, 2015; Yin et al., 2015a, 2015b, 2016;
era, lidar and spectrometer) (Leica Geosystems, 2014; Li Youssef et al., 2013; Zhang et al., 2012; and references
et al., 2014). Figure 1 shows the configuration of a typical air- therein). Among the results of observer-based FD, there are
borne earth observation system, which is composed of an air- two typical approaches. One scheme is fault estimation
craft, TGSCM, aerial camera, and position and orientation (Afshin, 2013; Chen and Saif, 2014; Liu and Duan, 2012).
system (POS). The TGSCM consists of three gimbals, which Based on the detection observer, the fault was reconstructed
are controlled to rotate, under the guidance of the POS, so to converge to the real value of the fault. This method can
that the imaging load moves smoothly along the course line achieve FD, isolation and estimation simultaneously. Another
and its line-of-sight (LOS) points at the earth stably (Fang scheme is residual-based FD, which concerns the design of the
et al., 2014; Fang and Liu, 2015; Li SS et al., 2015). However, FD filter (FDF) for residual generation, the determination of
the TGSCM usually works in a vibration environment during an evaluation function and threshold, and the decision of fault
flight and the reliability of the mechanical components meets occurrence by comparing the evaluation function with the
many challenges. Thus, the occurrence of fault is inevitable threshold. For systems subject to unknown disturbances, the
and it may lead to performance degeneration or even instabil-
ity of the TGSCM (Jia et al., 2014; Moradi and Fekih, 2014;
1
Wang et al., 2014). With increasing demands on system safety College of Electrical Engineering and Automation, Shandong University
and imaging quality, the issues of fault detection (FD) for of Science and Technology, Qingdao 266590, China
2
TGSCM control system has become increasingly important Department of Instrument Science and Opto-electronics Engineering,
(Zhong et al., 2016). Generally speaking, the potential faults Beijing University of Aeronautics and Astronautics, Beijing 100191,
China
of the TGSCM control system include actuator faults and
sensor faults. Recall that the POS is an INS/GPS integrated Corresponding author:
system and an independent subsystem of FD is usually Maiying Zhong, College of Electrical Engineering and Automation,
embedded in it. Without loss of generality, the problem of Shandong University of Science and Technology, Qingdao 266590, China.
FD for POS is not taken into account in the paper. Email: myzhong@buaa.edu.cn
720 Transactions of the Institute of Measurement and Control 40(3)

Figure 1. Configuration of an airborne earth observation system.

crucial issue in designing the FDF is improving simultane- discussions, this paper aims to apply the Hi/HN optimization-
ously the robustness of residual to unknown disturbances and based FD approach to closed-loop control system FD of the
the sensitivity to faults. As an ideal case, some techniques TGSCM. This paper is organized as follows. First, in the next
have been developed to design FDFs achieving unknown dis- section, a three-loop gimbal feedback control system model
turbances completely decoupling (Frank and Ding, 1997; Hou and the problem statement are introduced. Then, we describe
and Patton, 1996), but the decoupling conditions are usually the FD strategy. Experimental set-up and results are pre-
difficult to implement. For the general case of unknown dis- sented and finally, we conclude the paper.
turbance being coupled and energy bounded, one way is to
represent the robustness by the HN-norm of transfer function
from disturbance to residual, while the sensitivity by the HN- System modelling and problem statement
norm or a so-called H2 index of the transfer function from
fault to residual. Then the design of FDF can be formulated System description
as an HN/HN and/or H2/HN optimization problem, which As shown in Figure 1, the TGSCM is fixed to the aircraft
can achieve a best trade-off between the sensitivity to faults base, whereas the imaging load is mounted on the inner gim-
and the robustness to disturbances. It was shown that such an bal of the TGSCM and its LOS extends to the outside of the
optimal FDF can be obtained by applying a unified approach aircraft. The POS is installed on the imaging load to measure
to Hi/HN optimization in Ding et al. (2000b), where the ith
the load’s attitude angles with respect to the navigation frame
non-zero singular value of transfer function from fault to resi-
and then provides them to the TGSCM.
dual is used to evaluate the sensitivity to fault of a certain The schematic diagram of a TGSCM is shown in Figure 2.
direction, and a best trade-off between the sensitivity and the
From outside to inside, there are roll, pitch and drift gimbals
robustness is achieved at all fault directions. Because of this,
for compensation of aircraft three-axis angular movements,
the Hi/HN optimization-based FDF has been widely used in
respectively. Mx , My and Mz are three DC brushed torque
FD (Ding, 2013; Ding et al., 2000a; Li et al., 2013; Wang
motors and the gears are used to amplify the motors output
et al., 2007; Zhong et al., 2003, 2010).
torques. Gx , Gy and Gz are three angular rate gyroscopes to
The main focus of this paper is to deal with the problem
measure the gimbals’ angular rates with respect to the inertial
of FD for the TGSCM shown in Figure 1, where the roll,
frame. Ax and Ay are horizontal accelerometers to measure
pitch and drift gimbals are controlled independently. In
the aircraft accelerations. The encoders Ex , Ey and Ez are used
Zhong et al. (2016), an actuator fault was equivalent to addi-
to measure the gimbals’ relative angles. The three-axis direc-
tive signal and observer-based FDF was developed by apply-
tions are denoted x, y and z, respectively.
ing Hi/HN optimization approach. It should be noted that the
For the TGSCM system, each gimbal can be treated as a
design in Zhong et al. (2016) was carried out in the framework
rigid body and the kinematic model using the traditional
of open-loop FD. However, every gimbal of the TGSCM is a
Newton–Euler rotational equation can be described as
three-loop feedback control subsystem. Both the sensor fault
follows:
and actuator fault may be transferred in the closed-loop con-
trol system and the isolation of faults is very difficult for a 8
>
> dHx
closed-loop control system. In particular, once a fault is > + Hz vy  Hy vz = Tx
>
> dt
detected, the faulty case performance of TGSCM is usually >
<
dHy
unsatisfied for the application of high-resolution airborne + Hx vz  Hz vx = Ty ð1Þ
>
> dt
remote sensing and hence the first step is to set the operation >
>
>
> dHz + H v  H v = T
of TGSCM from ‘NORMAL’ mode to the ‘LOCKED’ mode : y x x y z
for protecting the imaging device. Motivated by the previous dt
Zhong and Li 721

Figure 2. Schematic diagram of a three-axis gyro-stabilized camera mount (TGSCM).

where H = ½ Hx Hy Hz T is inertial angular momentum, coupling effects Tcr , Tcp and Tcd are quite weak. Then, the
T T
v = ½ vx vy vz  is inertial angular rate and T = ½ Tx Ty Tz  kinematic equation (2) can be simplified to three independent
is external torque imposed to the gimbals’ body. equations:
Denote by Jrx , Jry , Jrz , Jpx , Jpy , Jpz , Jdx , Jdy and Jdz the iner-

tial moments of roll gimbal, pitch gimbal and drift gimbal Jx v_ r = Tx
along the x, y and z axes, respectively. ur , up and ud stand for roll gimbal:
u_ r = vr
the relative angles between the roll gimbal and base, the pitch 
Jy v_ p = Ty
gimbal and roll gimbal, and the drift gimbal and pitch gim- pitch gimbal: _ ð3Þ
bal, respectively. Then the kinematics of the TGSCM can be u p = vp

expressed as (Fang et al., 2015): Jz v_ d = Tz
drift gimbal:
u_ d = vd
8
> (Jrx + Jpx + Jdx )v_ r + Tcr = Tx
>
>
>
> (Jpy + Jdy )v_ p + Tcp = Ty where the inertial moments Jx =Jrx + Jpx + Jdx ; Jy = Jpy + Jdy
>
>
>
> and Jz = Jdz .
< Jdz v_ d + Tcd = Tz
ð2Þ Recall that the TGSCM is usually considered three single
>
> u_ r = vr gimbals for independent control. Here, a single gimbal system
>
>
>
> u_ p = vp is shown in Figure 3, which is composed of three parts: a DC
>
>
>
:_ brushed torque motor, gear transmission and gimbal system.
u d = vd
According to the typical motor voltage equation and the
simplified kinematic equation, the dynamic model of a single
where the gimbals’ coupling effects are
gimbal system for the TGSCM can be summarized as:

Tcr = Jdz u_ p u_ d  2(Jpx + Jdx  Jpz  Jdz )u_ r u_ p up , 8 di


>
> La dta + Ra ia = ua  NKe vl
>
>
< Tl = NKT ia
Tcp =  Jdz u_ d u_ r + (Jpx + Jdx  Jpz  Jdz )u_ 2r up , Tcd = Jdz u_ r up : ð4Þ
>
>
> (N 2 Jm + Jl ) dv
dt = Tl + Td
l

>
: dul
For the TGSCM, the angles ur and up are less than 5°, and dt = vl
angular velocities u_ r , u_ p and u_ d are less than 10°/s, so the prod-
uct of them is small. Therefore, comparing with other distur- where ua is the motor input voltage, ia the motor output cur-
bances like mass imbalance torque and friction torque, the rent, La the armature inductance, Ra the armature resistor, N
722 Transactions of the Institute of Measurement and Control 40(3)

Figure 3. Composition of a single gimbal system.

Figure 4. PID-based three-loop control block diagram.

the gear transmission ratio, Ke the back electromotive force the POS is crucial and such an INS/GPS integrated system
coefficient, KT the motor torque coefficient, Jm the motor iner- has an independent fault diagnosis system (Guo et al., 2014;
tia moment, Jl the single gimbal inertia moment, Tl the con- Zhong et al., 2015). Once a fault is detected in the POS, an
trol torque, Td the disturbance torque, vl the gimbal angular alarm signal will be delivered and the TGSCM will change
rate and ul the attitude angle. the working mode from POS combination mode to other
As shown in Figure 4, the single gimbal system usually modes (Li et al., 2014). In particular, under the assumption of
adopts a PID-based three-loop control strategy: current loop, the POS working well, the actuator fault and fault from the
stabilization loop and position loop (Li and Zhong, 2015). hall sensor as well as the stabilization gyroscope become main
uref , vref and Iref are three-loop reference inputs. Gu (s), Gv (s) factors affecting the safety and reliability of the TGSCM.
and Gi (s) are three-loop PID controllers. Therefore, as mentioned above, this paper focuses on devel-
In a normal aerial imaging mode, the attitude reference oping an FD strategy of actuator fault and sensor fault of
input uref = 0 and the pointing accuracy can reach 0.2° (1s) TGSCM in the framework of the closed-loop feedback con-
and angular rate steady accuracy can reach 0.3°/s (1s). trol system.
However, during a real flight, the TGSCM is affected seri-
ously by multi-source disturbances arising from external
atmospheric turbulences and internal structural vibrations. It Problem statement
is inevitable for the occurrence of faults, which can cause sys- First, the actuator fault is taken into consideration. For a sin-
tem performance deterioration and even lead to the system gle gimbal system, the actuator is constituted with the motor
instability. Moreover, if the TGSCM is not stable, then the stator and rotor, transmission shaft and gear, as shown in
mentioned accuracy indices are no longer significant. As one Figure 5 (Wang et al., 2016; Zhou et al., 2014). The stator is
of the most important sensors of TGSCM, the reliability of fixed on the mount body, whereas the rotor is connected to
Zhong and Li 723

Figure 5. Schematic diagram of the actuator composition.

the transmission shaft through fastening blocks and screws. Accordingly, let uFa represent the input voltage considering
The terminal of the transmission shaft is linked together with a faulty motor, ua represent the control voltage from the con-
the first stage gear, whereas the second-stage gear is fixed on troller’s output and f represent the fault voltage. Then we
the gimbal to be rotated. When the motor is powered, a con- have
trol torque is generated from the rotor coils and passed to the
transmission shaft through the spacer flange. Owing to the uFa = ua + f ð5Þ
meshing effect of the gears, the control torque is finally sent
to operate the rotating gimbal. First, we need to establish the state equation of the single
As indicated above, the spacer flange relies on the screws gimbal system. In Figure 4, the current controller Gi (s) is usu-
to connect with the rotor. During vibrations over a long time, ally designed to a proportional element with the coefficient
the screws can become loose and cause the spacer flange to KIP . Hence, letting control input u = Iref , state variable
slide arbitrarily. In this case, the motor rotor will be slipping x = ½ ia vl ul T and POS output as measured output y, the
or even free-wheeling with the gimbal not rotated. Then, the state space equation of the single gimbal can be expressed as
whole actuator experiences a loss of torque output effective- 
ness and thus the gimbal is uncontrollable. This circumstance x_ = Ax + Bu + Bf f + Bd d
ð6Þ
is similar to an opposite torque acting on the motor to pre- y = Cx + Dd d
vent its rotation.
In addition, the imaging load weights more than 100 kg. where
Derived from the pressure of the imaging load, the second- 2 3
stage gear has displacement changes and causes a degradation  KIP L+a Ra  NK
La
e
0
of the centring coaxiality. In this case, there are unexpected A=4 NKT
0 05
N 2 Jm + Jl
meshing forces in the gears and thus the whole actuator can 0 1 0
be stuck at unknown conditions. This circumstance is similar 2K 3
IP
to a faulty motor that cannot export a normal control torque. La
Above all, the actuator faults in both cases can lead to torque B=4 0 5
transmission failures including a loss of effectiveness and 0
sticking (Zhong et al., 2016). For a DC torque motor, the 213
La
output torque is controlled by the input voltage so that the Bf = 4 0 5
torque failure is similar to the input voltage failure. Thus, the 0
actuator fault can be regarded as an additive fault occurring
2 3
in the motor input voltage ua . 0 0
However, when the TGSCM system works in the control Bd = 4 N 2 Jm1+ Jl 0 5
scheme as indicated in Figure 4, the failure of the hall sensor 0 0
or gyroscope will be amplified and then transferred to the
plant in the framework of the closed-loop feedback mode. C = ½ 0 0 1 , Dd = ½ 0 0:1  and d = ½ Td nd T is an
Usually, the sensor fault possesses a characteristic of constant unknown input vector composed of disturbance torque Td
bias or constant scale, which can often deal with an additive and POS measurement noise nd . For the TGSCM, the distur-
fault in the sensor output. In the closed-loop feedback mode, bance torque includes a mass imbalance torque, friction tor-
as shown in Figure 4, the sensor fault will be coupled into the que and so on. These disturbance torques are difficult to
motor input voltage ua by the controllers’ output of Gv (s) model but they are usually bounded. Moreover, the system
and Gi (s). Thus, the sensor fault can also be considered as an noise arises from POS measurement noise. We have adopted
additive fault in the motor input voltage ua . a high-precision laser-gyro POS with the attitude accuracy of
724 Transactions of the Institute of Measurement and Control 40(3)


0.005° (1s) so that the noise nd is usually small. We take the J (r)\Jth ) fault-free case
Seu : ð10Þ
variance of nd as 0.1 for the noise matrix Dd considering a J (r) . Jth ) faulty case
worst case. Without loss of generality, both d and f are con-
sidered to be L2-norm bounded. Obviously, the threshold Jth is usually positive due to the exis-
For the purpose of FD, an observer-based FDF is often tence of unknown disturbance. The larger Jth is, the smaller
used to generate the residual (Ding, 2013) the false alarm rate of fault is, but the poorer the FD rate.
8 Now the main issues concerned with optimal FD for the
< ^x_ = A^x + Bu + H(y  ^y) TGSCM control system can be summarized as:
^y = C^x ð7Þ
:
r = R(y  ^y) A1) Finding an observer gain matrix H and a post-filter
R(s) such that A-HC is stable and (9) is satisfied;
where ^x and ^y represent the state and output estimation vec- A2) Choosing an evaluation function J (r) and a threshold
tors, respectively, r is the residual signal, H is the observer Jth so that the presence of fault is detected by
gain matrix and R is the post-filter. using (10).
Defining e = x  ^x, the dynamic of (7) is governed by

e_ = (A  HC)e + (Bd  HDd )d + Bf f Remark 1. It should be noted that si (Grf (jv)) is not a norm
ð8Þ
r = R(Ce + Dd d) except for the largest one. In (9), the si (Grf (jv)) is used to
measure the sensitivity of the residual to a fault at direction
Recall that observer-based FD consists of two stages: resi- of the ith eigenvector. The HN norm of transferfunction 
dual generation and residual evaluation. It is seen from (8) Grf (s) represents the best-case sensitivity, whereas Grf (s)
that the residual r is affected by not only the fault f but also is the worst-case sensitivity. In Ding (2013), a unified solution
the unknown input d. Moreover, the TGSCM suffers from of (9) is given for all the possible sensitivity indices denoted
multi-source disturbances that are non-linear, have large mag- by non-zero singular value si (Grf (jv)). Moreover, the optimal
nitude and varying, which directly affect the performance of solution of (9) is not unique and the static post-filter given
FD. Therefore, the first key issue of improving FD perfor- later in Lemma 1 is a special case of the optimal dynamic
mance is to design an FDF so that an optimal trade-off can post-filter (Zhong et al., 2010).
be achieved between the robustness of residual r to unknown
input d and the sensitivity to fault f.
Denote the robustness of residual r to unknown input d by Fault detection strategy
kGrd (s)k‘ and the sensitivity of residual r to fault f by
si (Grf (jv)). Then the design of FDF can be formulated as the Residual generation
following optimization problem (Ding, 2013): The first step of FD is to generate the residual signal. In this
work, a traditional FDF by (7) is applied as a detection
si (Grf (jv)) observer:
max JHi =H‘ = , v 2 ½0, ‘Þ ð9Þ
H, R(s) kGrd (s)k‘
8
where si (Grf (jv)) stands for the ith non-zero singular value of
>
>
> ^i_a =  KIP + Ra ^ia  NKe v ^l +
KIP
Iref + H1 (y  ^ul )
>
> La La La
Grf (jv) and JHi =H‘ is the sensitivity/robustness ratio criterion >
>
>
<_ NKT ^
of robust FD. As the residual generator satisfying (9) achieves ^ ^l = 2
v ia + H2 (y  ^ul )
Sob : N J m + Jl
an optimal trade-off between the sensitivity to fault and >
>
> ^_
>
robustness to disturbance, the Hi/HN optimization based FD >
>
> ^ l + H3 (y  ^ul )
ul = v
>
:
provides us with an efficient approach to robust FD for a sys- ^
^y = ul
tem subject to L2-norm bounded  unknown disturbances. ð11Þ
Moreover, Grf (s)‘ and Grf (s) = inf smin (Grf (jv))
v2½0, ‘Þ
As a result, the observer S ^ ob performs an estimation of the
are often used to respectively measure the best-case and
worst-case sensitivity of the residual to a fault and that attitude information ^ul based on POS attitude measurement y
satisfying and software calculation value Iref . The residual signal r is
derived from the output y and estimation ^ul as follows:
   
Grf (s) Grf (s)
‘ 
max JH‘ =H‘ = , max JH =H = r(s) = R(s)e(s), e(s) = y(s)  ^ul (s) ð12Þ
H , R(s) kGrd (s)k‘ H, R(s)  ‘ kGrd (s)k‘
Combining (6), (11) and (12), gives
is the well-known optimal HN/HN-FDF and H2/HN-FDF,
where smin (Grf (jv)) denotes the smallest non-zero singular
r(s) = R(s)(Gef (s)f (s) + Ged (s)d(s)) ð13Þ
value of Grf (jv).
After design a residual generator, the another key issue where the transfer function matrices are
of improving FD performance is the residual evaluation,
which concerns the choice of evaluation function J(r) and Gef (s) = C(sI  A + HC)1 Bf ,
threshold Jth. Then the occurrence of a fault can be detected
by using Ged (s) = C(sI  A + HC)1 (Bd  HDd ) + Dd :
Zhong and Li 725

Based on this, the sensitivity of the residual to a fault can be calculate the observer gain H and post-filter R(s). Using
evaluated by the non-zero singular values si (R(jv)Gef (jv)), Lemma 1, a best trade-off between the sensitivity to fault and
whereas the robustness of residual to unknown input can be robustness to unknown input can be achieved. However, the
evaluated by kR(s)Ged (s)k‘ . Thus, a sensitivity/robustness optimal solutions are not unique. That given in Lemma 1 is
ratio criterion is given by just a special case. For any observer gain matrix H ensuring
the stability of A-HC, there exists a dynamic post-filter R(s)
si (R(jv)Gef (jv)) such that (H, R(s)) is still an optimal solution to the optimiza-
max JHi =H‘ , where JHi =H‘ = , v 2 ½0, ‘Þ
H, R(s) kR(s)Ged (s)k‘ tion index (14). The corresponding proofs have been given in
ð14Þ Ding (2013) and here we omit them.
Remark 2. Differently from Zhong et al. (2016), this paper
Substituting Gef (s) and Ged (s) into (14), the criterion can be
deals with the problem of FD for the PID-based three-loop
implemented as
feedback control system of TGSCM, whereas the FDF in
si (R(jv)C(jvI  A + HC)1 Bf ) Zhong et al. (2016) is designed only for the open current loop
JHi =H‘ =   , v 2 ½0, ‘Þ system. Note that the mentioned fault can be modelled as an
R(s)(C(sI  A + HC)1 (Bd  HDd ) + Dd )
‘ additive signal f and the developed FD strategy for closed-
loop system model (6) can be applied to detect the occurrence
To solve this optimization problem, a unified solution was of actuator fault, hall sensor fault and gyroscope fault in the
proposed in Ding (2013), which is shown in the following stabilization loop.
Lemma 1.
Lemma 1. Considering a nominal system Residual evaluation

x_ = Ax + Bu + Bf f + Bd d
with the following detectable con- After residual generation, the remaining task is the residual
y = Cx + Df f + Dd d
ditions satisfied: evaluation stage, which concerns the choice of evaluation
function and threshold. As an evaluation over the whole time
B1) (C, A)is detectable, is usually unrealistic, the residual evaluation can be performed
  by using a finite-time window norm as (Ding, 2013; Zhong
A  jvI Bd
B2) has full row rank for all v. et al., 2010)
C Dd
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ð t2 ffi
Then a unified solution of (18) can be given by J (r)=krk2, T , wherekrk2, T = T
r rdt and T = t2  t1 ð15Þ
t1
 
H = Bd DTd + YC T Q1 , R(s) = Q1=2
On the other hand, multi-source unknown disturbances exist
where Q = Dd DTd and Y  0 is a solution of the following alge- in the TGSCM so that the evaluation function J (r) will not be
braic Riccati equation: zero even under a no-fault scenario. Let
 T    
Y A  Bd DTd Q1 C + A  Bd DTd Q1 C Y  YC T Q1 CY J (rd ) = krd k2, T , J (rf ) = rf 2, T
 2
+ Bd I  DTd Q1 Dd BTd = 0 where rd (t) = r(t)jf = 0 and rf (t) = r(t)jd = 0 . Note that, in the
fault-free case, we have
Consequently, for the TGSCM system, the ranks of the
following matrices are: J (r) = J (rd ), rd (s) = R(s)Ged (s)d(s) ð16Þ
2 3
KIP + Ra NKe Hence,
s+ 0
  6 La La 7
6 7
sI  A 6 NKT 7 J (rd ) = kR(s)Ged (s)d(s)k2, T  kR(s)Ged (s)k‘ kd k2, T ð17Þ
rank = rank 6  2 s 0 7 = 3, 8s
C 6 N Jm + Jl 7
4 0 1 s5 By using the unified solution in Lemma 1, the robustness
0 0 1 index can be calculated as (Ding, 2013)
 
A  jvI Bd
rank kR(s)Ged (s)k‘ = 1 ð18Þ
C Dd
2 3
KIP + Ra NKe Thus we obtain
6 La
 jv 
La
0 0 0 7
6 7
6 NKT 1 7
6
= rank 6 jv 0 0 77 = 4, 8v
J (rd )  kd k2, T  d, with d . 0 ð19Þ
6 N 2 Jm + Jl N 2 Jm + Jl 7
6 7
4 0 1 jv 0 0 5 where the parameter d is called a supremum of the L2-norm of
0 0 1 0 0:1 d. With a given zero false alarm rate, the threshold Jth can be
selected as
Therefore, the detectable conditions B1 and B2 are easily sat-
isfied so that we can use the solutions in Lemma 1 to Jth = d ð20Þ
726 Transactions of the Institute of Measurement and Control 40(3)

Table 1. Relation between the threshold and false alarm rate. tolerable within 5%, the threshold can be set to 0.03417. On
some occasions, the missed alarm rate should be limited in a
Threshold False alarm rate certain range so that a smaller threshold is appropriate
(Jth = 0:03378 with a false alarm rate 20%; or Jth = 0:03362
Jth = 0:03494 0%
with a false alarm rate 30%).
Jth = 0:03475 0.1%
Jth = 0:03465 0.5% Remark 3. The selection of threshold Jth is a trade-off between
Jth = 0:03450 1% the false alarm rate and missed alarm rate. In practice, the
Jth = 0:03428 3% bound of the disturbance can often be obtained through experi-
Jth = 0:03417 5%
mental tests (Zhong et al., 2016). Therefore, the relationship
Jth = 0:03398 10%
between the threshold and false alarm rate can be determined
Jth = 0:03390 13%
Jth = 0:03386 15% considering the worst disturbance case. Meanwhile, we can
Jth = 0:03378 20% select an appropriate small threshold to reduce the missed
Jth = 0:03369 25% alarm rate based on an acceptable false alarm rate.
Jth = 0:03362 30%
Jth = 0:03359 32%
Remark 4. For residual evaluation, the window length T can
Jth = 0:03354 35% accommodate the data smoothness to noise and system
response speed. A larger T can reduce the influence of noise
and increase the detection time, whereas a smaller value can
make the system response faster but can extend the data tur-
bulence. In this paper, we think T = 1 s is appropriate with a
sampling frequency 1 kHz.

Fault detection algorithm


Based on (10), (11), (15) and the threshold selection in
Table 1, an equivalent actuator failure can be detected effec-
tively. To this point, we give the complete strategy for Hi/HN
optimization based FD as shown in Figure 7. The overall FD
algorithm can be summarized in Algorithm 1.

Algorithm 1. Given the original system (6) and a sampling


period Ts, set computational number k = 0 and initial time
t1 = t2 = 0. The observer gain matrix H and post-filter R(s)
have been calculated based on Lemma 1.

Figure 6. Residual evaluation J(rd) and supremum d with a zero fault. Step 1. Sample POS to refresh attitude measurement y. The
value of Iref is the stabilization loop controller’s out-
It should be noted that such a threshold may lead to poorer put. Substitute y and Iref into (11) to obtain the
attitude estimation ^ul .
missed alarm rate, especially for an incipient fault. In order to
Step 2. Substitute y and ^ul into (12) to calculate the resi-
reduce the conservativeness of FD, a smaller threshold can be
dual signal r.
selected based on an acceptable false alarm rate. For this
Step 3. Put k=k + 1 and t2 = kTs . If kTs  T , accumulate
problem, some real experiments have been developed under a
r into J(r) according to (15) betweent 2 ðt1 , t2  and
zero fault scenario. The residual evaluation J (rd ) and supre-
then set t1 = t1 + Ts for next calculation.
mum d is shown in Figure 6.
Step 4. Select an appropriate threshold Jth based on an
In Figure 6, the data interval period is 0.01 s and there are
acceptable false alarm rate in Table 1.
30,000 points during the whole 300 s. It can be seen that, on
Step 5. Run the FD system to refresh J(r) based on Step 1
rare occasions, the value of J (rd ) can reach the supremum
to Step 3. Then evaluate the fault occurrence by
d = 0:03494 and in most instances, J (rd ) is much smaller
(10) with the real-time J(r) and preselected thresh-
than d. Hence, if the threshold Jth is set to d, it is much con-
old Jth .
servative. According to the sampling data of J (rd ), we have
established a corresponding relation between the threshold
and false alarm rate. The results are shown in Table 1. Remark 5. With the aid of online fault isolation and estima-
It is indicated that based on an acceptable false alarm rate tion, the integrated design of fault diagnosis and fault toler-
(maybe 1%, 3% or 5%), the threshold can be set to a smaller ant control is a useful choice of TGSCM. It should be noted
value and then the missed alarm rate can be further reduced. that, however, the isolation of fault for closed-loop control
Although the data in Table 1 is not statistic, it can provide an system remains a challenging problem. The focus of this
approximate relation between the threshold and false alarm paper is on FD for TGSCM with an L2-norm bounded
rate. Because of this, if a zero false alarm rate is required, the unknown disturbance and additive fault from both actuator
threshold can be set to 0.03494. If the false alarm rate is and sensors. From the viewpoint of reliability in practice,
Zhong and Li 727

Figure 7. Actuator fault detection strategy.

Table 2. Roll gimbal parameters.

Symbol Description Value

Vp Motor peak voltage 50.0 V


Ip Motor peak current 6.25 A
vNL Motor no load speed 46 rad/s
Ra Motor armature resistor 8O
La Motor armature inductance 17 mH
N Gear transmission ratio 6.70
KT Motor torque coefficient 1.08 Nm/A
Ke Motor back EMF coefficient 1.08 Vs/rad
KIP Current loop PID controller 4.05
Jm Motor inertia moment 5.43e-4 kgm2
Jl Gimbal inertia moment 1.00 kgm2

Figure 8. Experimental set-up.


Accordingly, the single gimbal system (4) can be governed
once a fault alarm is triggered, the operation of the TGSCM with the following matrices:
system will be set from the ‘NORMAL’ mode to the 2 3 2 3 2 3
‘LOCKED’ mode for protecting the device. 708:82 425:64 0 238:23 58:82
A = 4 7:06 0 0 5, B = 4 0 5, Bf = 4 0 5,
0 1 0 0 0
Experimental results 2 3
0 0
Experimental set-up Bd = 4 0:97 0 5, C = ½ 0 0 1 , Dd = ½ 0 0:1 :
0 0
The experimental set-up is shown in Figure 8. The TGSCM
consists of a blue roll gimbal, a yellow pitch gimbal and a Based on Lemma 1, the observer gain matrix and post-filter
black drift gimbal. The simulated payload (about 30 kg) is can be calculated as
used to stand for the imaging load. The laser-gyro POS is
made in our laboratory with the attitude accuracy 0.005° H = ½ 1:0733 1:7767 1:8851 T , R(s) = 10
(1s). The POS is composed of two parts: IMU and PCS. The
IMU is an inertial measurement unit and the PCS is called
POS computer system. The POS is not only the attitude angle
sensor but also the accuracy evaluation benchmark. In these
Fault detection results
experiments, we just take the roll gimbal of the TGSCM as For simplicity, only an actuator failure is considered in the
the subject to indicate the FD results. The type of the roll experiment, which is potentially dangerous and may cause the
motor is QT2603A, which is the main constituent part of the TGSCM to be out of control. With the safety requirement,
actuator. All of the required parameters are given in Table 2. some experiments are carried out to create a faulty condition
728 Transactions of the Institute of Measurement and Control 40(3)

Figure 10. Case 1: Jth=0.03494 and J(r) with a 0.3-V stepwise fault.
Figure 9. 0.3-V stepwise fault voltage.

in the actuator. In the experiments, the basic three-loop con-


trol strategy is applied to stabilize the roll gimbal and the atti-
tude reference input is set to zero. When the system works
normally, some fault signals are added into the motor input
voltage and the proposed algorithm is used to detect the pro-
vided fault.
First, in order to describe a sudden actuator failure due to
the components looseness, a stepwise fault voltage is consid-
ered. In Table 2, the motor peak voltage Vp=50.0 V. There is
no doubt that a large fault signal can easily be detected. To
demonstrate the effectiveness, this paper only focuses on the
circumstance of a small fault signal. Figure 9 shows the step-
wise fault signal (using real experiment data). At t=(100,
120) s, the fault is set to f=0.3 V, whereas at t=(180, 200) s,
f=20.3 V and otherwise f=0. The evaluation window is Figure 11. Case 2: Jth=0.03475 and J(r) with a 0.3-V stepwise fault.
initialized to T=1 s with a sampling frequency of 1 kHz. In
order to compare the detection performance with different
Table 3. Fault detection results with a 0.3-V stepwise fault.
thresholds, two cases are experimented:
Case 1: Selecting Jth=0.03494, the residual evaluation Threshold Detection time False alarm rate
function J(r) (blue sold line) and threshold Jth (red dashed
line) are shown in Figure 10. It can be seen that the evalua- Jth = 0:03494 1.9 s, 2.7 s 0%
tion function J(r) is below the threshold level in the fault-free Jth = 0:03475 1.0 s, 1.8 s 0.1%
case and zero false alarm rate is achieved. When the positive
fault occurs, the evaluation function goes above the threshold
level after 1.9 s of fault occurrence and therefore, the fault evaluation function J(r) is below the threshold with a small
can be detected in a hurry. Similarly, after 2.7 s of the nega- rate of fault false alarm occurring at t=(165, 165.3) s. The
tive fault occurrence, J(r) . Jth and a fault alarm is triggered. false alarm rate is about 0.1%, which is near to the result in
Note that the detection time of a negative fault is larger Table 1. Moreover, the FD time is shortened, the values of
than the positive fault. The fact is that there is a disturbance which are reduced to positive 1.0 s and negative 1.8 s,
torque Td in the TGSCM varying in different positions or respectively.
rotation directions. We experimented to show that Td has a The FD results with the stepwise fault in both cases have
positive bias. When a negative fault occurs, the actions of Td been given in Table 3. Obviously, if the threshold Jth is set to
and f will be weakened reciprocally. In this case, the control be large, some small amplitude faults might be missed,
signal becomes smaller, which is more easily affected by noise whereas if Jth is set to be small, then the number of fault false
and the system response speed becomes slower than the case alarms will increase. Therefore, the selection of Jth is a trade-
of a positive fault. off between the fault false alarm rate and missed alarm rate.
Case 2: Selecting Jth=0.03475, the residual evaluation However, in the experiments we show that by running the
function J(r) (blue sold line) and threshold Jth (red dashed system in the fault-free condition, the evaluation function J(r)
line) are shown in Figure 11. In this case, the number of is no more than 0.03494. Even with a maximum threshold,
missed alarm can be reduced. Similarly, in most instances, the the provided fault can be detected effectively.
Zhong and Li 729

Figure 12. 0.3-V sine fault voltage. Figure 14. Case 2: Jth=0.03475 and J(r) with a 0.3-V sine fault.

Table 4. Fault detection results with a 0.3-V sine fault.

Threshold Detection time False alarm rate

Jth = 0:03494 3.2 s, 4.8 s 0%


Jth = 0:03475 2.1 s, 3.4 s 0.12%

Figure 13. Case 1: Jth=0.03494 and J(r) with a 0.3-V sine fault.

Second, to simulate a slowly changing actuator failure due


to the aircraft disturbing angular movement, a periodic sine
fault is considered. The period is set to 40 s and magnitude is
0.3 V. Figure 12 shows the sine fault voltage (using real
experiment data). At t=(100, 120) s, a positive half-wave sine
fault is generated, whereas at t=(180, 200) s, a negative half- Figure 15. 0.25-V stepwise fault voltage.
wave fault is appended and otherwise f=0. Similar to the
stepwise fault, two cases with different thresholds are intro- whereas the false alarm rate is about 0.12%. Moreover, the
duced as follows: FD time is shortened, reduced to positive 2.1 s and negative
Case 1: Selecting Jth=0.03494, the residual evaluation 3.4 s, respectively. The FD results with the sine fault in both
function J(r) (blue sold line) and threshold Jth (red dashed cases have been given in Table 4. It is shown that a slowly
line) are shown in Figure 13. It is shown that J(r) is also a changing fault can also be detected effectively even with a
slowly changing signal. With a maximum threshold, there is maximum threshold.
no fault false alarm. At about t=103.2 s after 3.2 s of the pos- Based on the analysis above, it can be concluded that the
itive fault occurrence, the evaluation function goes above the presented FD strategy can detect a provided fault signal effec-
threshold and the fault is detected quickly. Likewise, after 4.8 tively. Additionally, comparing with a normal control voltage
s of the negative half-wave fault occurrence, J(r) . Jth and a (more than 0.5 V), the magnitude of fault voltage is smaller
fault alarm is triggered. (f=0.3 V). Even so, the detection performance is also accep-
Case 2: Selecting Jth=0.03475, the residual evaluation table in both cases of a stepwise or sine fault.
function J(r) (blue sold line) and threshold Jth (red dashed Following this, to indicate the detection results with a
line) are shown in Figure 14. In this case, a small rate of fault smaller fault, we take a stepwise fault for example. The fault
false alarm occurs at t=(74.5, 74.7) s and t=(76.3, 76.5) s, magnitude is reduced to 0.25 V, as shown in Figure 15 (using
730 Transactions of the Institute of Measurement and Control 40(3)

Figure 16. Jth=0.03494 and J(r) with a 0.25-V stepwise fault. Figure 18. Jth=0.03494 and J(r) with a 0.2-V stepwise fault.

detected. However, any closed-loop system has a certain


robustness and fault tolerance capability. If the fault is too
small, the effects of the fault on the system are much weaker
and the closed-loop system can still work correctly. For a real
TGSCM system, the focus is on detecting the fault, which has
an obvious effect on the system normal operation.

Conclusion
In this paper, the problem of FD in a closed-loop mode has
been studied for the TGSCM. Through analysis of the poten-
tial faults, the actuator or sensor fault results in a failure of
torque transmission, which can be summarized as an additive
fault existing in the motor voltage, and based on this, the
fault model was given. For the purpose of FD, an observer-
Figure 17. 0.2-V stepwise fault voltage.
based FDF was used for residual generation. The sensitivity
of the residual to a fault is evaluated by the non-zero singu-
lar values of the transfer function from the fault to the resi-
dual, whereas the robustness to disturbance is represented
real experimental data). The initial sets are the same as those by the HN-norm of the transfer function from disturbance
above and the threshold is set to Jth=0.03494. The evaluation to the residual. Based on this, the design of FDF was formu-
function J(r) and threshold Jth are shown in Figure 16. It can lated as an Hi/HN optimization problem such that a sensi-
be seen that when there is no fault, J(r)\Jth and a zero fault tivity/robustness ratio criterion was maximized and unified
false alarm rate is achieved. In the faulty cases at t=(100, solutions for the optimization problem were given. After
120) s or t=(180, 200) s, the evaluation function goes above that, a finite-time window evaluation function is developed
the threshold quickly and the fault can still be detected. to evaluate the residual. The window length is selected to
Because the peak magnitude of J(r) is close to the threshold, balance a trade-off between data smoothness against noise
the fault alarms are not continuous due to data turbulences. and system response speed. To determine the threshold, an
Next, the fault magnitude is further reduced to 0.2 V. analysis of false alarm rate and threshold selection was
Figure 17 shows the fault signal and Figure 18 shows the given. Finally, the overall FD strategy was summarized.
detection results. The threshold is also set to Jth=0.03494. It Experimental results were presented to demonstrate the
can be seen from Figure 18 that the evaluation function J(r) effectiveness. It was shown that the designed FD system
has further decreased in the faulty case. Because the fault is may detect the provided fault in the form of stepwise or sine
too small, the changes in the residual generated by the fault effectively. By reducing the evaluation threshold, the detection
become unnoticeable and then the fault cannot be detected. time is shortened and only a small rate of fault false alarm
However, it is also shown experimentally that the effect of the occurs. However, due to the influences of disturbances or noise,
fault on the system operation is quite small and in this case, if the fault signal is too small, the fault alarms are not continu-
the control system can still work normally. ous or the fault cannot even be detected. In the further, we will
The presented method has limitations in the detection of adopt a high-precision disturbance compensation strategy to
smaller faults. In some occasions, the fault cannot even be improve the performance of FD.
Zhong and Li 731

Declaration of Conflicting Interests loss of actuator effectiveness. IET Control Theory & Applications
8(1): 42–50.
The authors declare that there is no conflict of interest.
Leica Geosystems GIS (2014) Leica PAV100-gyro-stabilized camera
mount technical reference manual. Available at: http://
Funding www.gis.leicageosystems.com.
Li SS and Zhong MY (2015) High-precision disturbance compensa-
This work was supported in part by the National Science tion for a three-axis gyro-stabilized camera mount. IEEE/ASME
Foundation of China (Nos. 61333005 and 61421063) and Transactions on Mechatronics 20(6): 3135–3147.
Research Fund for the Taishan Scholar Project of Shandong Li SS, Zhong MY and Zhao Y (2014) Accelerometer error estimation
Province of China. and compensation for three-axis gyro-stabilized camera mount
based on proportional multiple-integral observer. Science China
Technological Sciences 57(12): 2387–2395.
References Li SS, Zhong MY and Zhao Y (2015) Estimation and compensation
Afshin I (2013) Self-tuning fault diagnosis of MEMS. Mechatronics of unknown disturbance in three-axis gyro-stabilized camera
23: 1094–1099. mount. Transactions of the Institute of Measurement and Control
Chen W and Saif M (2014) Simultaneous fault isolation and estima- 37(6): 732–745.
tion of lithium-ion batteries via synthesized design of Luenberger Li W, Zhu Z, Zhou G, et al. (2013) Optimal Hi/HN fault-detection fil-
and learning observers. IEEE Transactions on Control System ter design for uncertain linear time-invariant systems: an iterative
Technology 22(1): 290–298. linear matrix inequality approach. IET Control Theory & Applica-
Ding SX (2013) Model-based Fault Diagnosis Techniques. New York: tions 7(6): 1160–1167.
Springer-Verlag. Li YY and Zhong MY (2013) Optimal fault detection for a class of
Ding SX, Ding EL and Jeinsch T (2000a) A new optimization discrete-time switch linear systems. Journal of Systems Engineering
approach to the design of fault detection filters. In: Proceedings of and Electronics 24(3): 512–518.
IFAC Symposium on Fault Detection, Supervision and Safety of Li ZL, Rachid O, Stefan G, et al. (2015) Diagnosis for PEMFC sys-
Technical Processes, pp. 255–260. tems: a data-driven approach with the capabilities of online adap-
Ding SX, Jeinsch T and Frank PM (2000b) A unified approach to the tation and novel fault detection. IEEE Transactions on Industrial
optimization of fault detection systems. International Journal of Electronics 62(8): 5164–5174.
Adaptive Control Signal Processes 14(7): 725–745. Liu H and Duan Z (2012) Actuator fault estimation using direct
Fang JC and Liu ZC (2015) In-flight alignment of POS based on reconstruction approach for linear multivariable systems. IET
state-transition matrix. IEEE Sensor Journal 15(6): 3258–3264. Control Theory & Applications 6(1): 141–148.
Fang JC, Wang CE and Wen T (2014) Design and optimization of a Moradi A and Fekih A (2014) Adaptive PID-sliding-mode fault-
radial hybrid magnetic bearing with separate poles for magneti- tolerant control approach for vehicle suspension systems subject
cally suspended inertially stabilized platform. IEEE Transactions to actuator faults. IEEE Transactions on Vehicle Technology 63:
on Magnetics 50(5): 1–11. 1041–1054.
Fang JC, Yin R and Lei XS (2015) An adaptive decoupling control Thumati BT and Halligan GR (2013) A novel fault diagnostics and
for three-axis gyro stabilized platform based on neural networks. prediction scheme using a nonlinear observer with artificial
Mechatronics 27: 38–46. immune system as an online approximator. IEEE Transactions on
Frank PM and Ding SX (1997) Survey of robust residual generation Control System Technology 21(3): 569–578.
and evaluation methods in observer-based fault detection systems. Wang J, Yang GH and Liu J (2007) An LMI approach to H_/HN
Journal of Process Control 7: 403–424. fault detection observer design. Automatica 43: 1656–1665.
Guo J, Zhang H and Zhong MY (2014) Online evaluation of the iner- Wang JC and Qi XH (2015) Fault diagnosis for flight control systems
tial sensor qualities for the INS/GPS systems. In: Proceedings of using subspace method and adaptive two-stage Kalman filter.
IEEE Chinese Guidance, Navigation and Control Conference, pp. Transactions of the Institute of Measurement and Control, DOI:
1806–1810. 01423312115596805.
He X, Wang ZD, Wang XF, et al. (2013a) Least-squares fault detec- Wang YQ, Zhang SY, Li YB, et al. (2014) Fault detection for
tion and diagnosis for networked sensing systems using direct state singular networked control systems with randomly occurring
estimation approach. IEEE Transactions on Industrial Electronics delays and partially known distribution transmission delays.
9(3): 1670–1679. Transactions of the Institute of Measurement and Control 37(4):
He X, Wang ZD, Wang XF, et al. (2013b) Networked strong track- 446–456.
ing filtering with multiple packet dropouts: algorithms and appli- Wang ZP, Luo Y and Yu B (2016) Optimization of matching layer
cations. IEEE Transactions on Industrial Electronics 61(3): structure for piezoelectric fiber composite ultrasonic actuator and
1454–1463. performance testing. Transactions of the Institute of Measurement
Hossein B, Mostafa A and Mehran H (2013) Fault detection, isola- and Control 230(1): 37–43.
tion and accommodation for attitude control system of a three- Yin S and Huang ZH (2015) Performance monitoring for vehicle sus-
axis satellite using interval linear parametric varying observers and pension system via fuzzy positivistic c-means clustering based on
fault tree analysis. Transactions of the Institute of Measurement accelerometer measurements. IEEE/ASME Transactions on
and Control 228(8): 1403–1424. Mechatronics 20(5): 2613–2620.
Hou M and Patton RJ (1996) An LMI approach to fault detection Yin S, Wang G and Gao HJ (2015a) Data-driven process monitoring
observers. In: Proceedings of UKACC International Conference on based on modified orthogonal projections to latent structures.
Control, pp. 305–310. IEEE Transactions on Control System Technology, DOI: 10.1109/
Huang J, Yu L and Zhu F (2014) Actuator fault detection and estima- TCST.2015.2481318.
tion for the Lur’e differential inclusion system. Applied Mathemat- Yin S, Xie XC, Lam J, et al. (2015b) An improved incremental
ical Modelling 38: 2090–2100. learning approach for KPI prognosis of dynamic fuel cell system.
Jia Q, Chen W and Zhang Y (2014) Robust fault reconstruction via IEEE Transactions on Cybernetics, DOI: 10.1109/
learning observers in linear parameter-varying systems subject to TCYB.2015.2498194
732 Transactions of the Institute of Measurement and Control 40(3)

Yin S, Zhu XP, Qiu JB, et al. (2016) State estimation in nonlinear sys- Zhong MY, Ding SX and Ding EL (2010) Optimal fault detection for
tem using sequential evolutionary filter. IEEE Transactions on linear discrete time-varying systems. Automatica 46: 1395–1400.
Industrial Electronics, DOI: 10.1109/TIE.2016.2522382. Zhong MY, Ding SX and Lam J (2003) An LMI approach to design
Yin X and Li ZJ (2016) Decentralized fault prognosis of discrete robust fault detection filter for uncertain LTI systems. Automatica
event systems with guaranteed performance bound. Automatica 39(3): 543–550.
69: 375–379. Zhong MY, Guo J and Cao Q (2015) On designing PMI Kalman fil-
Youssef AB, Khil SKE and Belkhodja IS (2013) State observer-based ter for INS/GPS integrated systems with unknown sensor errors.
sensor fault detection and isolation and fault tolerant control of a IEEE Sensors Journal 15(1): 535–544.
single-phase PWM rectifier for electric railway traction. IEEE Zhong MY, Li SS and Zhao Y (2016) Optimal actuator fault detec-
Transactions on Power Electronics 28(12): 5842–5853. tion for a TGSCM system based on disturbance compensation.
Zhang K, Staroswiecki M and Jiang B (2012) Reduced-order IEEE Transactions on Industrial Electronics 63(5): 3205–3215.
observer-based fault estimation design for multiple input-multiple Zhou XY, Zhang HY and Yu R (2014) Decoupling control for two-
output discrete-time systems. Transactions of the Institute of Mea- axis inertially stabilized platform based on an inverse system and
surement and Control 226(1): 1037–1059. internal model control. Mechatronics 24: 1203–1213.

You might also like