Академический Документы
Профессиональный Документы
Культура Документы
+ 4C
2
l
2
2C
2
l z
1
+ 2C
2
l z
2
+ 4K
2
l
2
2K
2
lz
1
+ 2K
2
lz
2
= 0, (2)
J
+ 4C
2
b
2
2C
2
b
2
1
2C
2
b
2
2
+ (4K
2
b
2
+ 2K
) (2K
2
b
2
+ K
)
1
(2K
2
b
2
+ K
)
2
= 0, (3)
where z, z
1
and z
2
denote the vertical displacement of the carbody, the leading bogie and the
trailing bogie, respectively. denotes the pitch angle of the centre of gravity(c.g.). denotes
the roll angle of the c.g. for the masses. The parameters of the vertical vehicle suspension
system are given in Table 1.
D
o
w
n
l
o
a
d
e
d
b
y
[
M
c
M
a
s
t
e
r
U
n
i
v
e
r
s
i
t
y
]
a
t
0
4
:
1
8
1
3
M
a
r
c
h
2
0
1
3
Vehicle System Dynamics 703
Figure 1. The vertical suspension system of the rail vehicle.
Table 1. The parameters of the vehicle suspension system.
Description Unit
M Carbody mass kg
M
B
Bogie mass kg
J
1
+ 4C
1
l
2
1
1
C
1
l
1
d
1r
C
1
l
1
d
1l
+ C
1
l
1
d
2r
+ C
1
l
1
d
2l
+ 4K
1
l
2
1
1
K
1
l
1
d
1r
K
1
l
1
d
1l
+ K
1
l
1
d
2r
+ K
1
l
1
d
2l
= 0, (5)
J
B
1
2C
2
b
2
+ (2C
2
b
2
+ 4C
1
b
2
1
)
1
+ C
1
b
1
d
1r
C
1
b
1
d
1l
+ C
1
b
1
d
2r
C
1
b
1
d
2l
(2K
2
b
2
+ K
) + (2K
2
b
2
+ 4K
1
b
2
1
+ K
)
1
+ K
1
b
1
d
1r
K
1
b
1
d
1l
+ K
1
b
1
d
2r
K
1
b
1
d
2l
= 0, (6)
where d
1r
denotes the vertical displacement of the right wheel in the leading wheelset. d
2l
denotes the vertical displacement of the left wheel in the trailing wheelset. The meaning of
other symbols is dened in a similar way. To simplify the considered problem, this paper
assumes that the vertical displacements of the wheel are equal to the unevenness of the track.
D
o
w
n
l
o
a
d
e
d
b
y
[
M
c
M
a
s
t
e
r
U
n
i
v
e
r
s
i
t
y
]
a
t
0
4
:
1
8
1
3
M
a
r
c
h
2
0
1
3
704 X. Wei et al.
In a similar way, the model of the trailing bogie can be derived as follows:
M
B
z
2
2C
2
z + 2C
2
l
+ (4C
1
+ 2C
2
) z
2
C
1
d
3r
C
1
d
3l
C
1
d
4r
C
1
d
4l
2K
2
z + 2K
2
l + (4K
1
+ 2K
2
)z
2
K
1
d
3r
K
1
d
3l
K
1
d
4r
K
1
d
4l
= 0, (7)
J
B
2
+ 4C
1
l
2
1
2
C
1
l
1
d
3r
C
1
l
1
d
3l
+ C
1
l
1
d
4r
+ C
1
l
1
d
4l
+ 4K
1
l
2
1
2
K
1
l
1
d
3r
K
1
l
1
d
3l
+ K
1
l
1
d
4r
+ K
1
l
1
d
4l
= 0, (8)
J
b
2
2C
2
b
2
+ (2C
2
b
2
+ 4C
1
)b
2
1
2
+ C
1
b
1
d
3r
C
1
b
1
d
3l
+ C
1
b
1
d
4r
C
1
b
1
d
4l
(2K
2
b
2
+ K
) + (2K
2
b
2
+ 4K
1
b
2
1
+ K
)K
)
+ (2K
2
b
2
+ 4K
1
b
2
1
+ K
)d
4r
K
1
b
1
d
4l
= 0. (9)
The state-space description of the vertical suspension model can be derived as
x = Ax + B
d
d, (10)
y = Cx + D
d
d, (11)
where
x = [ z
z z
1
1
z
1
1
1
z
2
2
z
2
2
2
]
T
,
d = [
d
1r
d
1l
d
2r
d
2l
d
1r
d
1l
d
2r
d
2l
d
3r
d
3l
d
4r
d
4l
d
3r
d
3l
d
4r
d
4l
]
T
,
y = [z z
1
1
1
z
2
2
2
]
T
,
where matrixes A, B
d
, C and D
d
are derived from the previous differential equations. d is the
vertical track variation velocity and displacement due to track vertical irregularities.
2.2. The dynamical suspension model under the new sensor conguration
In the developed model (10) and (11), the vertical displacement, pitch angle displacement and
roll angle displacement of the cardody and the bogie are selected as the system outputs. This
means that the displacement sensor and angle displacement sensors are required to measure
these signals for the purpose of fault detection. However, displacement sensor and angle dis-
placement sensors have some problems in reliability, maintenance and installation. Compared
with these sensors, acceleration sensors have the merits such as cheapness and reliability. In
addition, it does not need to be maintained for a long time period. Acceleration sensors are
widely used in the health condition monitoring of railway systems.Considering all the reasons
stated above, a novel sensor conguration is proposed as shown in Figure 1. The vehicle sus-
pension system is only equipped with acceleration sensors in the four corners of the carbody,
the leading and trailing bogies. Carbody sensors are equipped in the four corners on the oor-
board, and the bogie sensors are equipped in the four corners on the upside of the bogie. The
acceleration signal can be transformed to displacement signal by applying double integral to
the acceleration signal, that is,
z =
__
a dt dt, (12)
where a is the acceleration value and z is the displacement.
Remark 2.1 In principle, the displacement signal can be obtained by double integrating
directly the acceleration signal. However, in reality, the output of acceleration sensors always
D
o
w
n
l
o
a
d
e
d
b
y
[
M
c
M
a
s
t
e
r
U
n
i
v
e
r
s
i
t
y
]
a
t
0
4
:
1
8
1
3
M
a
r
c
h
2
0
1
3
Vehicle System Dynamics 705
Figure 2. Relationship among the displacement of the four points.
Figure 3. The pitch motion.
contains the direct current (DC) component. The DCcomponent must be ltered by a high-pass
lter. The numerical integral algorithm is also critical to achieve a high accuracy.
In the following, the new output equation under the new sensor conguration framework is
derived. The problemneeds to be solved is mainly howto build the relation matrix between the
outputs in Equation (11) and the new outputs, the displacements at the corners of the carbody,
the leading bogie and the trailing bogie.
Here, the relationship between the four vertical displacements {z
, z
fr
, z
rl
, z
rr
} of the four
carbody oor corners and the carbody displacements, the roll angle and the pitch angle {z, , }
is derived. A simplied carbody oor is depicted in Figure 2. Dene two variables z
f
and z
r
,which are the vertical displacement of the middle point of the front edge and the middle point
of the right edge, respectively, then one obtains
z + z
fr
= z
f
+ z
r
. (13)
The displacement z
f
can be replaced by the following equation:
z
f
= z + l
c
sin() z + l
c
, (14)
which is trivially obtained by using the pitch motion of the carbody oor depicted in Figure 3.
Similarly, we have the following equation:
z
r
z w
c
, (15)
which is derived by using the roll motion of the carbody oor depicted in Figure 4.
D
o
w
n
l
o
a
d
e
d
b
y
[
M
c
M
a
s
t
e
r
U
n
i
v
e
r
s
i
t
y
]
a
t
0
4
:
1
8
1
3
M
a
r
c
h
2
0
1
3
706 X. Wei et al.
Figure 4. The roll motion.
In terms of Equations (13)(15), one obtains
z
fr
z + l
c
w
c
. (16)
Following the same procedure, one obtains
z
z + l
c
+ w
c
, (17)
z
rl
z l
c
+ w
c
, (18)
z
rr
z l
c
w
c
. (19)
The transformation matrix between {z
, z
fr
, z
rl
, z
rr
} and {z, , } is built, which is
z
fr
z
rl
z
rr
1 l
c
w
c
1 l
c
w
c
1 l
c
w
c
1 l
c
w
c
. (20)
The relation matrix between the four corner displacements of the leading bogie and the
trailing bogie and their three DOF variables can be derived in a similar way. The following
transformation matrix is obtained:
T =
1 l
c
w
c
0 0 0 0 0 0
1 l
c
w
c
0 0 0 0 0 0
1 l
c
w
c
0 0 0 0 0 0
1 l
c
w
c
0 0 0 0 0 0
0 0 0 1 l
b
w
b
0 0 0
0 0 0 1 l
b
w
b
0 0 0
0 0 0 1 l
b
w
b
0 0 0
0 0 0 1 l
b
w
b
0 0 0
0 0 0 0 0 0 1 l
b
w
b
0 0 0 0 0 0 1 l
b
w
b
0 0 0 0 0 0 1 l
b
w
b
0 0 0 0 0 0 1 l
b
w
b
,
one yields
y = Ty, (21)
where
y = [z
z
fr
z
rl
z
rr
z
1_fl
z
1_fr
z
1_rl
z
1_rr
z
2_fl
z
2_fr
z
2_rl
z
2_rr
]
T
is the output under the new sensor conguration. {z
1_fl
, z
1_fr
, z
1_rl
, z
1_rr
} represent the four ver-
tical displacements of the leading bogie corners, respectively. {z
2_fl
, z
2_fr
, z
2_rl
, z
2_rr
} represent
D
o
w
n
l
o
a
d
e
d
b
y
[
M
c
M
a
s
t
e
r
U
n
i
v
e
r
s
i
t
y
]
a
t
0
4
:
1
8
1
3
M
a
r
c
h
2
0
1
3
Vehicle System Dynamics 707
the four vertical displacements of the trailing bogie corners. Then, the state-space description
under the new measurement sensor conguration can be obtained
x = Ax + B
d
d, (22)
y = Ty,
= TCx + TD
d
d,
=
Cx +
D
d
d, (23)
where
C = TC and
D
d
= TD
d
.
The discrete model of the vertical suspension system is easily derived as follows:
x
k+1
= Gx
k
+ H
d
d
k
, (24)
y
k
=
Cx
k
+
D
d
d
k
, (25)
where G = e
AT
, H =
_
T
0
e
A
dB and T is the sampling time.
Remark 2.2 High-integrity data are very critical for the data-driven fault detection methods.
The data should contain rich enough information of the system dynamics and the fault infor-
mation when a fault occurs in the system. In this paper, only acceleration sensors are used
for the vehicle suspension fault detection systems. From the above observation, the dynamics
of rail vehicle suspension systems are contained in the 12 displacements of the carbody, the
leading and trailing bogies, which are measured by the acceleration sensors. That is to say, the
sensor conguration presented before can provide enough information for the fault detection.
3. A brief review of the fault detection methods
In this section, the model-based fault detection methods, the robust observer-based method
and the Kalman lter-based approach, and data-driven fault detection methods, the DPCAand
CVA, are briey reviewed.
3.1. Robust fault detection observer design and MCUMSUM
The discrete model of the suspension system with faults is described by
:=
_
x
k+1
= Gx
k
+ H
d
d
k
+ H
f
f
k
,
y
k
=
Cx
k
+
D
d
d
k
+
D
f
f
k
,
(26)
where f
k
R
n
f
is the fault vector. The robust fault detection observer design objective here is
to design an observer O, which has the following formulation:
O :=
x
k+1
= G x
k
+ L( y
k
y
k
),
y =
C x
k
,
r
k
= y
k
y
k
(27)
to maximise the sensitivity of the fault to the residual r
k
and also maximise the robustness of
the disturbance to the residual.
D
o
w
n
l
o
a
d
e
d
b
y
[
M
c
M
a
s
t
e
r
U
n
i
v
e
r
s
i
t
y
]
a
t
0
4
:
1
8
1
3
M
a
r
c
h
2
0
1
3
708 X. Wei et al.
Dene e = x
k
x
k
, the state estimation error dynamic equations can be described by
e
k+1
= G
e
e
k
+ H
d
e
d
k
+ H
f
e
f
k
,
r
k
=
Ce
k
+
D
d
d
k
+ D
f
f
k
, (28)
where G
e
= G L
C, H
d
e
= H
d
L
D
d
, H
f
e
= H
f
L
D
f
.
The transfer function of the state estimation error dynamic systemof the observer is given by
r = G
rd
(z)d + G
rf
(z)f , (29)
where
G
rd
(z) =
C(zI G
e
)
1
H
d
e
+
D
d
, (30)
G
rf
(z) =
C(zI G
e
)
1
H
f
e
+
D
f
. (31)
The fault detection observer must be robust to the disturbances (the robustness conditions)
and sensitive to the faults (the sensitivity conditions). The observer design can be transformed
into an linear matrix inequality optimisation problem as follows:
max
s.t. G
rd
(z)G
rd
(z) <
2
I
G
rf
(z)G
rf
(z) >
2
I
G
e
is stable.
Please refer to our previous work [17] for details of the observer design. In this paper, the
well-known multivariate CUMSUM [18] is adopted for the residual r
k
change detection. At
each time k, we calculate statistic Q
k
as
Q
k
= (Q
k1
+ r
k
u)
_
1
q
C
k
_
if C
k
> q, (32)
where u represents the mean of the residual r
k
and q is a predetermined statistical distance,
C
k
=
_
(Q
k1
+ r
k
u)
1
(Q
k1
+ r
k
u) (33)
and is the covariance matrix of the observation data. If C
k
q, the process resets Q
k
= 0.
The MCUMSUM starts with Q
0
= 0 and triggers an alarm when S
k
=
_
Q
T
k
1
Q
k
exceeds
a predetermined threshold, h, that is chosen to achieve a desired performance. The robust
observer fault detection system is shown in Figure 5.
Figure 5. Robust observer fault detection system.
D
o
w
n
l
o
a
d
e
d
b
y
[
M
c
M
a
s
t
e
r
U
n
i
v
e
r
s
i
t
y
]
a
t
0
4
:
1
8
1
3
M
a
r
c
h
2
0
1
3
Vehicle System Dynamics 709
3.2. Fault detection based on the Kalman lter and GLRT
Assume that the track irregularities are white noises, then a Kalman lter is designed for the
suspension system (25) and (26) as follows:
x
k+1
= (G K
C)x
k
+ K y
k
,
r
k
= y
k
Cx
k
. (34)
The no fault (H
0
) and fault (H
1
) hypotheses test are described in terms of the innovation as
follows:
H
0
:
k
= r
k
, (35)
H
1
:
k
= r
k
+ g
k
(
), (36)
where r
k
is the residual in the absence of the fault case. g
k
(
. g
k
is generated by the failure signature dynamical equation
k+1
= (G K
C)
k
+ (H
f
K
D
f
)s
k
, (37)
g
k
=
C
k
+
D
f
s
k
. (38)
The primary principle behind is that for each time instant k, check if there is a failure in the
past time with the generalised likelihood ratio
k
(
, ) =
p(
kL
,
kL+1
, . . . ,
k
|H
1
,
, )
p(
kL
,
kL+1
, . . . ,
k
|H
0
)
=
j=k
j=kL
p(
j
|H
1
,
, )
p(
j
|H
0
)
(39)
for all k [k L, k], where L is the sliding window length.
Taking the log of the above ratio, it follows that
k
(
, ) = (
)
1
2
2
S
k
(
), (40)
where
k
(
) =
k
j=
g
T
j
(
)R
1
j
j
, (41)
R
j
= CP
j
C
+ DD
, (42)
S
k
(
) =
k
j=
g
T
j
(
)R
1
j
g
j
(
), (43)
where P
j
is the system noise covariance.
The generalised log likelihood ratio is given by
l
k
= max
(kL,k)
max
R
k
(
, ). (44)
Further explanation and detailed algorithm of GLRT can be found in [1923]. The Kalman
lter and GLRT-based fault detection system is shown in Figure 6.
D
o
w
n
l
o
a
d
e
d
b
y
[
M
c
M
a
s
t
e
r
U
n
i
v
e
r
s
i
t
y
]
a
t
0
4
:
1
8
1
3
M
a
r
c
h
2
0
1
3
710 X. Wei et al.
Figure 6. Kalman lter and the GLRT-based fault detection system.
3.3. A brief description of PCA-based fault detection method
The standard PCA-based fault detection consists of three steps and is formulated as follows:
Data collection and pre-processing: Consider a data matrix X R
Nm
consisting of N
samples and m sensors collected from process. Matrix X is then scaled to zero mean, and
often to unit variance. Let the scaled data be
Y =
y
T
1
.
.
.
y
T
N
R
Nm
(45)
with y
i
R
m
, i = 1, . . . , N, denoting the ith scaled vector.
Decomposition of covariance matrix: The covariance matrix is formed as
0
=
1
N 1
Y
T
Y. (46)
By means of singular value decomposition (SVD), the covariance matrix is decomposed as
follows:
1
N 1
Y
T
Y = PP
T
, =
_
pc
0
0
res
_
, (47)
where
pc
= diag(
2
1
, . . . ,
2
l
),
res
= diag(
2
l+1
, . . . ,
2
m
), with
2
i
, i = 1, . . . , m, is the ith
singular value of the covariance matrix, PP
T
= I
mm
and
_
P
T
pc
P
T
res
_
[P
pc
P
res
] =
_
I
ll
0
0 I
(ml)(ml)
_
.
On-line fault detection: When a new scaled measurement y R
m
is available, the squared
prediction error (SPE) and Hostelings T
2
indices can be computed as
T
2
= y
T
P
pc
1
pc
P
T
pc
y, (48)
SPE = y
T
P
res
P
T
res
y. (49)
The fault detection logic is SPE J
th,SPE
and T
2
J
th,T
2 fault-free. Otherwise faulty.
J
th,T
2 and J
th,SPE
are the thresholds for SPE and T
2
test, respectively.
The PCA methods can be extended to take the serial correlations into account by aug-
menting each observation vector with the previous l observations and stacking the data
D
o
w
n
l
o
a
d
e
d
b
y
[
M
c
M
a
s
t
e
r
U
n
i
v
e
r
s
i
t
y
]
a
t
0
4
:
1
8
1
3
M
a
r
c
h
2
0
1
3
Vehicle System Dynamics 711
matrix in the following manner:
Y(l) =
y
T
k
y
T
k1
y
T
kl
y
T
k1
y
T
k2
y
T
kl1
.
.
.
.
.
.
y
T
k+ln
y
T
k+ln1
y
T
kn
, (50)
where y
T
k
is the m-dimensional observation vector in the training set at time instant k. This
approach of applying PCA to Equation (50) is referred to as dynamic PCA (DPCA).
The vehicle suspension system is a dynamical system and the measured outputs are cor-
related with the past measurements. Hence, DPCA rather than PCA is applied for its fault
detection.
3.4. CVA-based fault detection method for dynamic processes
The stacked past and future vectors, p and f, are represented as follows:
p
kl
k1
= [y
T
k1
y
T
kl
]
T
, (51)
f
k+l+1
k
= [y
T
k
y
T
k+l1
]
T
, (52)
where y denotes output vectors, subscript k is the present time index for y, and l is the number
of the lag or the lead. In CVA, the stacked future and past vectors are normalised by using
d
f ,k
=
1/2
ff
f
k+l+1
k
, d
p,k1
=
1/2
pp
p
kl
k1
, (53)
where
ff
= E(f
k+l+1
k
f
k+l+1
k
T
) and
pp
= E(p
kl
k1
p
kl
k1
T
) (E() denotes the expectation oper-
ator). The normalised vectors d
f ,k
and d
p,k1
are dened as the scaled stacked future and past
vectors at time k, respectively. The conditional expectation of the scaled future vector takes
the following form:
E(d
f ,k
|d
p,k1
) =
1/2
ff
fp
1/2
pp
d
p,k1
, (54)
where
E(|) denotes the expectation of under condition . The physical meaning of
fp
, dened as E(f
k+l+1
k
p
kl
k1
T
), is the well-known Hankel matrix. Thus,
1/2
ff
fp
1/2
pp
indicates the scaled Hankel matrix. The scaled Hankel matrix can be factorised by using SVD,
d
f ,k
=
1/2
ff
fp
1/2
pp
d
p,k1
= USV
T
d
p,k1
U
q
S
q
V
T
q
d
p,k1
, (55)
where
d
f ,k
denotes
E(d
f ,k
|d
p,k1
) and q represents the state order. U
q
andV
q
consist of the rst
q column vectors of U and V, respectively, and the diagonal matrix S
q
is the q q principal
submatrix of S. Then, the past and future canonical variates at time k are given by
z
k
= U
T
q
d
f ,k
= U
T
q
1/2
ff
f
k+l+1
k
= L
q
f
k+l+1
k
,
m
k
= V
T
q
d
p,k1
= V
T
q
1/2
pp
p
kl
k1
= J
q
p
kl
k1
, (56)
respectively.
D
o
w
n
l
o
a
d
e
d
b
y
[
M
c
M
a
s
t
e
r
U
n
i
v
e
r
s
i
t
y
]
a
t
0
4
:
1
8
1
3
M
a
r
c
h
2
0
1
3
712 X. Wei et al.
Similar to the T
2
and SPEindices, the following indices are used for the CVAfault detection
test in this paper
T
2
s
= p
kl
k1
T
J
q
J
T
q
p
kl
k1
, (57)
SPE
s
=
T
k
k
, (58)
where
k
= (I J
q
J
q
)p
kl
k1
.
4. Simulations
In this section, the simulation environment and the fault generation principle are introduced
at rst. After that the fault detection result for all the four detection methods is presented in
detail. The advantages and disadvantages are compared and pointed out.
4.1. Simulation environment and the suspension fault generation
To analyse and study the fault detection performance of the four methods stated in the last
section, a SIMPACK and MATLAB co-simulation environment is built. The parameters are
provided by the vehicle manufacturer to obtain the simulation data for the purpose of algorithm
validation. The multi-body simulation model is built by the construction of coordinates system,
bodies, joints, constraints, force elements, track excitations and so on. The SIMPACK vehicle
model is used to generate the acceleration signals and different faults in both primary and
secondary suspensions. The SIMPACK vehicle suspension simulation model equipped with
acceleration sensors is shown on the right part of Figure 7.
In the SIMPACK simulation environment, it is not easy to generate a component fault in
the course of a simulation. One way is to build a new model, whose components are already
faulty. However, there are several different faults with different magnitudes for the underlying
suspension system. Many SIMPACK models need to be built and it is a very time-consuming
work. In order to simulate the fault components, the SIMPACK and MATLAB co-simulation
environment is built. On the MATLAB side, the data generated by the SIMPACK are acquired
and some pre-processing work is also carried out. Another important task for MATLAB is to
control SIMPACK and simulate the suspension faults with different fault magnitudes at any
time when the vehicle is running. In the following, the fault simulation method is presented.
In the light of the working principle of the damper component, the force generated by the
damper is equal to the damper coefcient times pistons velocity, which is used to prevent the
Figure 7. Co-simulation between SIMPACK and MATLAB/SIMULINK.
D
o
w
n
l
o
a
d
e
d
b
y
[
M
c
M
a
s
t
e
r
U
n
i
v
e
r
s
i
t
y
]
a
t
0
4
:
1
8
1
3
M
a
r
c
h
2
0
1
3
Vehicle System Dynamics 713
movement of the piston. When the damper is faulty, for instance, the coefcient has a 25%
reduction, a smaller force is generated by the damper and the reduced force is proportional
to the product of the pistons velocity and the reduced damper coefcient. A virtual force is
generated in the light of the fault scenarios and acts on the position where the damper is xed.
As shown in Figure 7, for instance, sensors are equipped at the position of a secondary damper
to measure its moving velocity. Assume that the damper coefcient is reduced to half of its
normal value at the 15th second (controlled by MATLAB), then an external force (the fault
signal in Figure 7) is exerted on the piston to reduce the resistance generated by the damper.
The direction of the virtual force is opposite to the force generated by the damper and the
value of force is equal to the fault magnitude times the pistons velocity. In a similar way,
spring faults and other components faults can also be simulated.
4.2. Performance assessment by simulation experiments
The fault detection results for the railway vehicle suspension system are shown in this section.
The track irregularity used in the simulations is the fth-grade track irregularity spectrum of
the US railway lines. The proposed sensor conguration and the fault detection result do not
depend on the special track irregularity. All vertical suspension component faults are fully
studied, but only some typical faults detection results are demonstrated here. The typical
faults considered are two primary suspension spring coefcient reduction scenarios and two
secondary suspension damper coefcient reduction scenarios listed in Table 2. The considered
spring or damper coefcient reduction of 25% and 75% represents a fault at its early stage
and a severe fault, respectively.
Remark 4.1 For urban railway suspension systems, the performance of the dampers and
springs degrades signicantly after one year or two years. This can be indicated by the damper
and spring coefcients. When the damper has leakage problem, the damper coefcient also
changes. These slow change faults can be described by the small fault, and the condition
monitoring device can provide useful information for component maintenance. For abrupt
and server fault case, fault detection device should send an alarm to the driver after the fault
occurs and an emergency braking is needed to guarantee the safety of the train. Fault detection
can be used to detect small faults to provide maintenance decision on the one hand. On the
other hand, it can also be applied to guarantee the safety of the train when severe faults come
forth.
The threshold for the fault detection alarm depends on the system disturbances (track irreg-
ularities), the residual generation algorithm and the residual evaluation method as well. In the
simulations, the threshold is trivially taken as a constant which is larger than the maximal value
of the residual evaluation function output when no faults occur in the system in the course
of a long enough simulation. For practical application, the threshold determination should
Table 2. Fault scenarios used in the simulation.
Scenario Fault Fault pattern Faulty time (s)
Fault 0 No fault
Fault 1 C
2
25% reduction 15
Fault 2 C
2
75% reduction 15
Fault 3 K
1
25% reduction 15
Fault 4 K
1
75% reduction 15
D
o
w
n
l
o
a
d
e
d
b
y
[
M
c
M
a
s
t
e
r
U
n
i
v
e
r
s
i
t
y
]
a
t
0
4
:
1
8
1
3
M
a
r
c
h
2
0
1
3
714 X. Wei et al.
make a tradeoff between sensitivity to faults and robustness to uncertainties such as the track
irregularities and model uncertainties. Different thresholds can be chosen for different track
segments to improve the fault detection performance.
In all the gures, the red solid line indicates the threshold. The blue dashed-dotted line
represents the result of no fault case. The suspension component losing 25% of its value is
indicated by the green dashed line. For the last scenario, 75% reduction of the components
coefcient is represented by the black dotted line. In all the simulations, the vehicle velocity
is set to 80 km/h and the sampling time for fault detection is 0.1 s. All the faults of different
scenarios are introduced into the suspension system at 15 s after the SIMPACK-MATLAB
co-simulation starts. The detailed simulation results are presented below.
Robust observer-basedmethod:Arobust observer is designedbasedonthe former work[17].
The parameters used for the observer synthesis are the same as those used for building the
SIMPACKvehicle dynamical model. The residual generatedbythe robust observer is evaluated
by the MCUMSUM algorithm. The results are shown in Figures 8 and 9. It can be seen clearly
from Figure 8 that the algorithm detects the 75% reduction fault of the secondary suspension
damper coefcient after 7 s when it occurs. The S
k
value is larger than the predened threshold
in most of the time after the fault occurs. This fault is clearly detected by the robust observer
method. On the other hand, the 25% reduction fault of the damper coefcient is not detected
clearly by the robust observer and at only several time instants, the S
k
value is above the
predened threshold. It is hardly concluded whether a fault occurs or not. For the early stage
secondary suspension damper fault, the robust observer-based approach is not very effective.
The fault detection result for the primary spring fault scenario is shown in Figure 9. Similar to
the damper fault case, the severe fault (75%coefcient reduction) is clearly detected around 7 s
after it emerges. Nevertheless, the small fault (25% coefcient reduction) is not successfully
detected. There are only several time instants where the S
k
is larger than the threshold. The
alarm is not persistent and it is hardly determined that a fault has come forth already.
0 10 20 30 40 50 60 70 80 90
0
10
20
30
40
50
60
70
S
k
Time(s)
Simulation results for secondary suspension damper faults
threshold
no fault
C
2
25% reduction
C
2
75% reduction
Figure 8. Robust observer and MCUMSUM algorithm for C
2
fault.
D
o
w
n
l
o
a
d
e
d
b
y
[
M
c
M
a
s
t
e
r
U
n
i
v
e
r
s
i
t
y
]
a
t
0
4
:
1
8
1
3
M
a
r
c
h
2
0
1
3
Vehicle System Dynamics 715
0 10 20 30 40 50 60 70 80 90
0
10
20
30
40
50
60
70
80
90
S
k
Time(s)
Simulation results for primary suspension spring faults
threshold
no fault
K
1
25% reduction
K
1
75% reduction
Figure 9. Robust observer and MCUMSUM algorithm for K
1
fault.
Remark 4.2 The model used for the fault detection does not include the coupling between
roll and lateral motion of the car body and bogie. This causes some model uncertainties. A
more precise model is helpful to improve the performance for detecting small fault.
GLRT based on the Kalman lter method: Figures 10 and 11 demonstrate the fault detection
results by using GLRTbased on the Kalman lter method for the secondary suspension damper
fault and the primary spring fault, respectively. For the damper fault scenario, both the early
stage fault (25%coefcient reduction) and the severe fault case (75%coefcient reduction) are
detected successfully. For the 75% coefcient reduction fault case, the rst alarm is triggered
out at around 10 s after the fault occurs. It takes more than 17 s to trigger the rst alarm for the
early stage fault (25%coefcient reduction). For the primary suspension spring fault, it can be
seen clearly that the severe fault is detected successfully 15 s after it comes forth. Nevertheless,
for the early stage fault, the alarm is triggered out only at very limited instants, even though
the rst alarm is triggered out just 4 s after the fault emerges.
Remark 4.3 In the Kalman lter and GLRT-based fault detection method, the disturbance is
assumed to be white noise. However, in the suspension system, the disturbance is the track
irregularity and it is not totally white. The fault detection performance could be improved if
more precise system model and disturbance model are used.
DPCA method: In this data-driven fault detection approach, a data set generated by the
SIMPACK without faults is used for building the statistical model. The time lag steps l in
Equation (50) is selected as 20. One hundred and twenty principle components are nally
selected to be retained in the model. Both T
2
test and SPE test are applied in the simulations.
It is found that the fault detection results based on the SPE test outperform the T
2
test. Hence,
only the SPE test results for the two typical fault scenarios are presented in the following.
The fault detection results for the secondary suspension damper faults and the primary
suspension spring faults are shown in Figures 12 and 13, respectively. It can be seen clearly
D
o
w
n
l
o
a
d
e
d
b
y
[
M
c
M
a
s
t
e
r
U
n
i
v
e
r
s
i
t
y
]
a
t
0
4
:
1
8
1
3
M
a
r
c
h
2
0
1
3
716 X. Wei et al.
0 10 20 30 40 50
0
0.5
1
1.5
2
2.5
x 10
4
Simulation results for secondary suspension damper faults
Time(s)
l
k
threshold
no fault
C2 25% reduction
C2 75% reduction
Figure 10. Kalman lter and GLRT algorithm for C
2
fault.
0 10 20 30 40 50
0
0.5
1
1.5
2
2.5
3
3.5
x 10
4
Simulation results for primary suspension spring faults
Time(s)
l
k
threshold
no fault
K1 25% reduction
K1 75% reduction
Figure 11. Kalman lter and the GLRT algorithm for the K
1
fault.
that for the severe fault cases (C
2
75%reduction and K
1
75%reduction), the SPEtest values are
much higher than the predened thresholds. For the early stage faults (C
2
25% reduction and
K
1
25% reduction), they are also successfully detected. Nevertheless, the SPE test values are
very close to the predened threshold. The response of the DPCA-based method to the faults
is much faster than the robust observer- and Kalman lter-based approaches stated before.
D
o
w
n
l
o
a
d
e
d
b
y
[
M
c
M
a
s
t
e
r
U
n
i
v
e
r
s
i
t
y
]
a
t
0
4
:
1
8
1
3
M
a
r
c
h
2
0
1
3
Vehicle System Dynamics 717
10 15 20 25 30 35 40 45
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
S
P
E
Time(s)
Simulation results for secondary suspension damper faults
threshold
no fault
C
2
25% reduction
C
2
75% reduction
Figure 12. DPCA and SPE for the C
2
fault.
10 15 20 25 30 35 40 45
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
S
P
E
Time(s)
Simulation results for primary suspension spring faults
threshold
no fault
K
1
25% reduction
K
1
75% reduction
Figure 13. DPCA and SPE for the K
1
fault.
CVA method: For the CVA approach, the dimension of the measured data is reduced from
12 to 6 by PCA techniques rst. This avoids the problem that
1/2
ff
and
1/2
pp
in Equation
(53) are not real. The time lag and lead steps l in Equation (51) are selected as 20. In the
CVA model, 18 states are selected. Both T
2
s
test and SPE
s
test are applied in the simulations.
D
o
w
n
l
o
a
d
e
d
b
y
[
M
c
M
a
s
t
e
r
U
n
i
v
e
r
s
i
t
y
]
a
t
0
4
:
1
8
1
3
M
a
r
c
h
2
0
1
3
718 X. Wei et al.
10 15 20 25 30 35 40
60
80
100
120
140
160
180
200
220
240
T
2s
Time(s)
Simulation results for secondary suspension damper faults
threshold
no fault
C
2
25% reduction
C
2
75% reduction
Figure 14. CVA and T
2
s
for C
2
fault.
10 15 20 25 30 35 40
0
1000
2000
3000
4000
5000
6000
7000
8000
T
2s
Time(s)
Simulation results for primary suspension spring faults
threshold
no fault
K
1
25% reduction
K
1
75% reduction
Figure 15. CVA and T
2
s
for the K
1
fault.
Contrary to the DPCA methods, the fault detection results based on the T
2
s
test outperforms
the SPE
s
test. Hence, only the T
2
s
test results for the two typical fault scenarios are presented
in the following.
The fault detection results are shown in Figures 14 and 15. It can be seen that the CVA
method detects all the faults successfully. Especially, for the primary suspension spring fault
D
o
w
n
l
o
a
d
e
d
b
y
[
M
c
M
a
s
t
e
r
U
n
i
v
e
r
s
i
t
y
]
a
t
0
4
:
1
8
1
3
M
a
r
c
h
2
0
1
3
Vehicle System Dynamics 719
Table 3. Comparison of the fault detection methods.
Method Sensitivity Response speed Complexity Limitation
Observer Worst Slow Complex Precise model required
GLRT Bad Slowest Complex Precise model required
PCA Good Fast Simple Fault isolation is difcult
CVA Best Fast Simple Fault isolation is difcult
cases, CVA even outperforms the DPCA method. Nevertheless, for the C
2
fault cases, DPCA
is better than the CVA approach.
4.3. Comparison of the four fault detection methods
In this part, the four fault detection methods are compared in terms of the sensitivity to
faults, response speed to faults, complexity of the methods and application limitations. The
comparison results are given Table 3.
It is surprising that the data-driven approaches have better performance in all aspects.
Actually, this result is much easier to be accepted when we notice the point that the model-
based method needs a precise model to achieve a good performance. Since the rail vehicle
suspension system is very complicated and there is nonlinearity existing in the system, the
model is certainly not a precise one, while the data-driven approaches do not need the model.
However, the model-based approach also has advantages such as dealing with fault isolation
and identication problems. The model-based fault detection approach is still being developed
and better detection performance is expected from the improvement of modelling techniques
and fault detection algorithms.
5. Conclusions
In this paper, the fault detection problem of the urban railway vehicle suspension system
is studied. A novel sensor conguration is proposed where the underlying vehicle system
is equipped with only acceleration sensors in the four corners of the carbody, the leading
and trailing bogies, respectively. A mathematical model is built for the considered vehicle
suspension system. After that a comparative study on fault detection methods of urban rail
vehicle suspension systems is considered. Both model-based and data-driven approaches are
studied for the suspension fault detection problem. The robust observer, the Kalman lter
combined with the GLRT method, the DPCA and the dynamical CVA approaches are applied
to the fault detection problem, respectively. The simulation results show that the data-driven
methods outperform the model-based methods.
In this paper, the vertical suspension system is considered for our comparative study. How-
ever, the sensor conguration, modelling and the fault detection methods can be trivially
extended to the lateral suspension system. In addition, the proposed methods can also being
applied to the gyro sensor conguration framework when robust and cheap gyros are available.
In this paper, only the fault detection issue for the suspension system is concerned. It
is necessary to further investigate the suspension system model, robustness and sensitivity
analysis of the fault detection methods. In addition, how to isolate these faults and estimate
the magnitude of each fault are not studied yet. This will be carried out in the future work.
D
o
w
n
l
o
a
d
e
d
b
y
[
M
c
M
a
s
t
e
r
U
n
i
v
e
r
s
i
t
y
]
a
t
0
4
:
1
8
1
3
M
a
r
c
h
2
0
1
3
720 X. Wei et al.
Acknowledgements
This work is partly supported by Chinese 863 program(Contract No. 2011AA110503-6), State Key Laboratory of Rail
Trafc Control and Safety (Contract No.RCS2010ZT003) and Ph.D. Programs Foundation of Ministry of Education
of China (Grant number: 20110009120037).
References
[1] S. Bruni, R. Goodall, T.X. Mei, and H. Tsunashima, Control and monitoring for railway vehicle dynamics, Veh.
Syst. Dyn. 45 (2007), pp. 765771.
[2] R. Goodall and T. Mei, Advanced control and monitoring for railway vehicle suspensions, International Sym-
posium on Speed-up and Service Technology for Railway and Maglev Systems(STECH06), Chengdu, China,
2006, pp. 1016.
[3] R. Goodall1 and C. Roberts, Concepts and techniques for railway condition monitoring, IET International
Conference on Railway Condition Monitoring, Birmingham, UK, 2006, pp. 9095.
[4] P. Li, R. Goodall, P. Weston, C.S. Ling, C. Goodman, and C. Roberts, Estimation of railway vehicle suspension
parameters for condition monitoring, Control Eng. Pract. 15 (2007), pp. 4355.
[5] Y. Hayashi, H. Tsunashima, andY. Marumo, Fault detectionof railway vehicle suspensionsystems usingmultiple-
model approach, Mech. Syst. Transp. Log. 1 (2008), pp. 8898.
[6] X. Wei, S. Lin, and H. liu, Distributed fault detection observer for rail vehicle suspension systems, Chinese
Control and Decision Conference, Taiyuan, China, 2012, pp. 34083413.
[7] X. Wei, L. Jia, and H. Liu, Fault diagnosis lter design for railway vehicle suspension systems based on LMI
optimization, Int. Interdiscip. J. accepted (2012).
[8] P. Li and R. Goodall, Model based condition monitoring for railway vehicle systems, Control 2004, University
of Bath, 2004, ID-508.
[9] H.T. Yusuke Hayashi and Y. Marumo, Condition monitoring of railway vehicle suspension, International
Conference on Control, Automation and Systems, Gyeonggi-do, Korea, 2010, pp. 584589.
[10] T.X. Mei and X.J. Ding, A model-less technique for the fault detection of rail vehicle suspensions, Veh. Syst.
Dyn. 46 (2008), pp. 277287.
[11] T.X. Mei and X.J. Ding, Condition monitoring of rail vehicle suspensions based on changes in system dynamic
interactions, Veh. Syst. Dyn. 47 (2009), pp. 11671181.
[12] X. Wei, H. liu, and Y. Qin, Fault diagnosis of rail vehicle suspension systems by using GLRT, Chinese Control
and Decision Conference, Mianyang, China, 2011, pp. 19321937.
[13] X. Wei, H. Liu, and Y. Qin, Fault isolation of rail vehicle suspension systems by using similarity measure,
International Conference on Intelligent Railway Tranportation, Beijing, China, 2011, pp. 391396.
[14] S.J. Qin, Data-driven fault detection and diagnosis for complex industrial processes, The 7th IFAC Symposium
on Fault Detection, Supervision and Safety of Technical Processes, Barcelona, Spain, 2009, pp. 11151125.
[15] S. Yin, S. Ding, A. Naik, P. Deng, and A. Haghani, On PCA-based fault diagnosis techniques, 2010 Conference
on control and Fault Tolerant Systems, Nice, France, 2010, pp. 179184.
[16] C. Lee, S.W. Choi, and L. In-Beum, Variable reconstruction and sensor fault identication using canonical
variate analysis, J. Process Control 16 (2006), pp. 747761.
[17] X. Wei, L. Liu, and M. Verhaegen, Fault detection and estimation for LTI systems and its application to a lab
robotic manipulator, Control and Decision Conference, 2009. CCDC09. Chinese, Guilin, 2009, pp. 15951600.
[18] R.B. Crosier, Multivariate generalizations of cumulative sumquality-control schemes, Technometrics 30 (1988),
pp. 291303.
[19] A.S. Willsky and H.L. Jones, A generalized likelihood ratio approach to the detection and estimation of jumps
in linear systems, IEEE Trans. Autom. Control 21 (1976), pp. 108112.
[20] F. Gustafsson, Adaptive Filtering and Change Detection, Wiley, Chichester, 2000.
[21] F. Guatafusson, The marginalized likelihood ratio test for detecting abrupt change, IEEETrans. Autom. Control
41 (1996), pp. 6678.
[22] H.L.V. Trees, Detection, Estimation, and Modulation Theory, Part I, John Wiley and Sons, NewYork, 1968.
[23] R.S. Mangoubi, Robust Estimation and Failure Detection, Springer, London, 1998.
D
o
w
n
l
o
a
d
e
d
b
y
[
M
c
M
a
s
t
e
r
U
n
i
v
e
r
s
i
t
y
]
a
t
0
4
:
1
8
1
3
M
a
r
c
h
2
0
1
3