A publication of

CHEMICAL ENGINEERING TRANSACTIONS

The Italian Association
of Chemical Engineering
Online at: www.aidic.it/cet

VOL. 33, 2013

Guest Editors: Enrico Zio, Piero Baraldi
Copyright 2013, AIDIC Servizi S.r.l.,
ISBN 978-88-95608-24-2; ISSN 1974-9791

DOI: 10.3303/CET1333118

Condition Monitoring of Shinkansen Tracks

Based on Inverse Analysis
Takahito Kobayashi*a, Yasukuni Naganumab and Hitoshi Tsunashimac
a

Graduate School of Nihon University, 1-2-1 Izumi-cho, Narashino-shi Chiba, Japan

Nippon Kikai Hosen KK., 2-1-95 Kounan, Minato-ku Tokyo, Japan
c
Nihon University, 1-2-1 Izumi-cho, Narashino-shi Chiba, Japan
[email protected],
b

This paper demonstrates the possibility to estimate the track irregularities of Shinkansen tracks using carbody motions only. In an inverse problem to estimate track irregularity from car-body motions, a Kalman
Filter (KF) was applied to solve the inverse problem. Estimation results showed that track irregularity
estimation in vertical direction is possible with acceptable accuracy for real use.

1. Introduction
Track irregularity deteriorates ride comfort and running safety, track condition monitoring for maintenance
is one of the most important tasks for railway companies. In general, track irregularity is measured several
times a month by a specially designed track geometry car. In 2009, the track condition monitoring system
called RAIDARSS started operation in the Tokaido Shinkansen line. Inertial measurement devices are
mounted on six N700 Shinkansen train set and they measure the track several times a day. However,
RAIDARSS needs high frequent maintenance because accelerometers are mounted on axle-boxes of
vehicles. If the track irregularity in level can be estimated from car-body acceleration of in-service vehicle,
it will enable high frequent track condition monitoring by a portable device (Tsunashima et. al., 2012).
This study proposes the track irregularity estimation techniques from car-body motion only. The inverse
analysis technique is applied to estimate the track irregularity. This technique is frequently utilized to
determine an unknown input signal (track irregularity) from a known output signal (car-body acceleration).

2. Construction and validity of vehicle model

2.1 Vehicle model
Figure 1 shows a railway vehicle model in vertical direction. Where zc is a car-body displacement, zt1 and zt2
are front and rear bogie displacement, c is a car-body pitch angle, t1 and t2 are front and rear bogie pitch
angle. Inputs, r1a,r1b,r2a,r2b, denote the track irregularities. The equation of motion for 6 DOF railway vehicle
traveling on a straight track can be written as
Mz(t ) + Cz(t ) + Kz(t ) = Dr(t ) + Er (t ) ,

(1)

where Z (t ) = [zc lc c zt1 lt 2t1 zt 2 lt 2t 2 ] , r (t ) = [r1a r1b r2 a r2b ] . Coefficient matrices of the equations (1) are shown
in the appendix.
The vehicle model was constructed using the discrete state-space model. The state equation and the
measurement equation are expressed by
T

x n = Fx n1 + Bun + Gw n ,

(2)

y n = Hx n + v n ,

(3)

where xn, un, yn are the state vector, the input vector and the output vector, respectively. wn and vn are
process noise and measurement noise.

Please cite this article as: Kobayashi T., Naganuma Y., Tsunashima H., 2013, Condition monitoring of shinkansen tracks based on inverse
analysis, Chemical Engineering Transactions, 33, 703-708 DOI: 10.3303/CET1333118
703

ks
zt2

cs

t2 mb I b

2" c

r2b

cs

ks
mb I b

cp

kp

10

zc

t1

zt1

cp

kp

r2 a

Gain [ms -2/m]

mc I c

r1b

r1a
2" b

Estimated

10

10

1
-2

-1

10

(a) Gain

Phase [rad]

Estimated

Coherence

10

Frequency [1/m]

Figure 1: Railway vehicle model

0.5

Vehicle Model

-2

10

Frequency [1/m]

0
-2
-4

-1

Vehicle Model

-2

-1

10

10

10

Frequency [1/m]

(b) Phase
Figure 3: Frequency response functions

Figure 2: Coherence

2.2 Frequency response functions and system identification

To confirm a validity of the simple model and parameter identification, frequency analysis was carried out. In
the input and output relations of vertical direction, input is track irregularity, and output is car-body
acceleration.
Figure 2 shows the coherence of track irregularity and car-body acceleration. This relation shows strong
correlation, thus it can be assumed the dynamics as linear system. Figure 3 shows the frequency response
functions of gain characteristic and phase characteristic. Dotted line given in Figure 3 indicates the frequency
response function calculated with the vehicle model, and solid line is estimated from measured track
irregularity and car-body acceleration. There is good correspondence between the estimated and the
calculated values. It demonstrates that the model and parameters are appropriate.

3. Track irregularity estimation technique

Two approaches for inverse analysis are applied in this study. First approach focuses on the impulse
response of the vehicle, and second one uses a new state equation of a simplified vehicle model. Details of
each method will be discussed in next section. Figure 4 shows the flow of inverse analysis using a KF.
Track geometry car

Vehicle model
With impulse response

With simplified vehicle model

Kalman Filter

10m-chord versine

Measured

Estimated

Estimated

Comparison and Evaluation

Figure 4: KF based signal-processing methods

In this study, to confirm the validity of theoretical inverse analysis, we carried out the following steps.
(Step 1) Generate a random number, and create the track geometry data (more than 6m wavelength) that is
similar to the frequency characteristics of the actual track.
(Step 2) Calculate the car-body acceleration and car-body pitch rate using the track geometry data.
(Step 3) Add a Gaussian noise to vehicle motion, and it is used as the measurement data.
(Step 4) Estimate the track geometry from measurement data using a KF.
(Step 5) Compare the track geometry.

704

A KF approach is a well-known estimation technique proposed in various fields. The track irregularities are
estimated from the measured car-body acceleration containing noise. The KF based estimation algorithms can
be written as
Time update equations
x n|n - 1 = Fx n - 1|n - 1 + Gun ,

(4)

Pn|n1 = FPn1|n1F T + GQG T ,

(5)

Measurement update equations

1
K n = x n|n 1 H T (HPn|n 1 H T + R ) ,

(6)

xn|n = xn|n1 + K n ( yn Hxn|n1 ) ,

(7)

Pn|n = Pn|n 1 K n HPn|n 1 ,

(8)

where xn|nT = [ xn xn-1 xn-L+1 ] is the track irregularities, and ynT = [ yn yn-1 yn-L+1 ] is the car-body
acceleration. Q is the covariance matrix of process noise. R is the covariance matrix of measurement noise.
In a conventional state equation, the external input un is treated as a known deterministic input. However,
in the inverse analysis, it is an unknown state to be estimated. Therefore, track geometries are defined as
the random walk model with the external input and the process noise. The state transition matrix F of
the equation (2) can be shown as
0    0
1    

F = 0    

    
0  0 1 0 .

4. Estimation of track irregularities

In this study, we are focusing on the 10m-wavelength track irregularity that affect the running safety.
Transformation of the 10m-chord versine method by actual track geometry can be approximated by
x (l ) = y (l ) ( y (l + 5) + y (l 5)) / 2 , where y(l) represents the track geometry, and x(l) represents the 10m-chord
versine.
4.1 Estimation using impulse response
In this section, a estimation technique using the impulse response of the vehicle (Figure 5) is evaluated.
The impulse response can be obtained by giving an impulse input to equation (2) and (3). When applying
an impulse response to inverse analysis, the observation matrix H of equation (3) can be given as
H = [ h ( L ) h ( L 1)   h ( 0) ] . The symbol h denotes the impulse response, and the symbol L denotes
the total number of impulse response (Kobayashi et. al., 2012).
Figure 6 shows the estimation results. The track geometry and 10m-chord versine can be estimated with
good accuracy with the actual data. The maximum estimation error is about 1.5mm. The result revealed
that the method is effective for track condition monitoring with acceptable accuracy.

Vertical
acceleration[m/s ]

80
60
40
20
0
-20
-40
0

50

100
150
Distance [m]
Figure 5: Impulse response

200

705

Track
geometry [mm]

60

True

30

Estimated

0
-30
-60
0

1000

2000

3000

4000

5000

Distance [m]

10m-chord
versine [mm]

(a) Track geometry

6

True

Estimated

0
-3
-6
0

200

400

600

800

1000

Distance [m]

Estimation
error [mm]

(b)10m-chord versine
2
1
0
-1
-2
0

200

400

600

800

1000

Distance [m]

(c)Estimation error
Figure 6: Simulation results
mc I c z
c

ks
mb

cs

2" c

kp

cp

cs

ks
mb

zt2

zt1

kp

cp

r1

r2
r2b

r2 a

r1b

r1a
2" b

Figure 7: Simplified vehicle model

4.2 Estimation using a simplified vehicle model
The forward analysis of vehicle model gives the 8 inputs (track geometry of 4 axes and its differential
value). But it is impossible to calculate back all these correctly with the inverse analysis. To develop a
numerical model for inverse analysis, the following two assumptions were made for simplification
(Naganuma et. al., 2009).
(a)

(b)

In practice, track irregularity is managed using 10m-chord versine whose wavelength bandwidth of 5
m to 100 m for running stability. Precisely, the vertical car-body acceleration does not contain
frequency components correspond to the wavelength of twice of the wheel base of the bogie.
Therefore, it is good enough for track management to know the average of irregularities at the front
and rear wheel of the bogie.
A vehicle model is not necessary to express the pitching motion of bogies because the purpose is to
estimate the track irregularity from the car-body acceleration.

Simplified vehicle model for inverse analysis is shown in Figure 7. In the previous section, the track
geometry is expressed by the random walk model (un+1=un+en), and it is shown that the KF based
algorithm can be applied to the inverse analysis (Naganuma et. al., 2012). In this approach, the external
force vector (first-order differentiation of the track geometry) is included in the state vector. The statespace model can be written as
x n +1 F
u = O
n +1
y n = [H

B xn I
+
un O

x
O ] n + v n
,
un

O wn
I e n ,

(9)
(10)

where is the unit vector. The track geometry can be estimated by this state-space model (9) and (10),
and it is estimated to continuously and stably as one-component of the state vector. This model observes
the pitch rate and the acceleration of car-body. Figure 8 shows the estimation results using the simplified
model. The maximum estimation error is about 1.0mm. The result revealed that the KF is effective for track
condition monitoring with acceptable accuracy as with the impulse model.

706

Track
geometry [mm]

60

True

30

Estimated

0
-30
-60

1000

2000

3000

4000

5000

Distance [m]

10m-chord
versine [mm]

(a) Track geometry

6

True

Estimated

0
-3
-6
0

200

400

600

800

1000

800

1000

Distance [m]
Estimation
error [mm]

(b) 10m-chord versine

2
1
0
-1
-2

200

400

600

Distance [m]

RMSE [mm]

(c) Estimation error

Figure 8: Simulation results
1.5
With impulse model

With simplified vehicle model

1
0.5
0

200

400

600

800

1000

Distance [m]

Figure 9: RMSE of the estimation results

5. Evaluation of the estimation results

5.1 RMSE (Root Mean Square Error)
In this section, RMSE of 20m section based on the result of 10 m-chord versine is calculated. The RMSE
is an index evaluate the difference between the true value and the estimated value. It can be written as
N

(X x )

RMSE =

(11)

i =1

where X is the true value, xi is the measured value, and N is the number of the data. Figure 9 shows the
RMSE evaluation of the estimation results. The results showed that the method with simplified vehicle has
better performance than the method with impulse response.
5.2 MPC metrics
In this section, we evaluated the estimation results of 10m-chord versine using the Sprague & Geers
Metrics method (Ray et. al., 2008). This method calculate the MPC metrics using the Sprague & Geers
correlation function. The MPC metrics treats the two waves magnitude and phase. The magnitude
component M should be sensitive to difference in magnitude. The phase component P should be sensitive
to difference in phasing. Component C is the combination with the magnitude and phase. These
characteristics of MPC metrics allow the analyst to identify the aspects of the curves that do not agree. For
each component of MPC metrics, zero indicates that the two waves are identical. Table 1 shows the
equations of each component and calculation results. The symbol ei and mi in the equation represents the
estimated values and true values. Table 1 data indicate that there is a high correlation between the true
value and estimated value.
Table 1: MPC metrics
Magnitude
Equations

M =

ei2
1
mi2

Phase
P=

cos 1

ei mi
ei2 mi2

Combination

C = M 2 + P2

With impulse response

0.082

0.072

0.109

With simplified vehicle model

0.075

0.060

0.096

707

6. Conclusion
In this paper, the track irregularities estimation techniques from car-body acceleration and/or pitch rate are
proposed. Main results of the study are:
(a)
(b)

The track geometry expressed by the random walk model, can be applied to the inverse analysis with
KF formulation.
The KF was used to estimate longitudinal level irregularity from car-body motion. The results
confirmed that this filter can be used to calculate back the track irregularity with accuracy sufficient for
track condition monitoring.

The proposed methods can be applied for the on-board condition monitoring device because they have
real-time processing algorithm with the consideration of vehicle speed change.
References
Kobayashi, T., Naganuma, Y. and Tsunashima, H., 2012, Condition Monitoring of Shinkansen Tracks
Based on Vehicle Model, The 6thInternational Symnosium on Speed-up, Safty and Service
Technology for Railway and Maglev Systems, Korea.
Naganuma, Y. and Yoshimura, A., 2009, Reconstruction and estimation of railway track geometry using
regularization methods, Proceedings of IAVSD, Stockholm, Sweden.
Naganuma, Y., Kobayashi, T. and Tsunashima, H., 2012, Study on track geometry estimation from carbody acceleration, The 19th Jointed Railway Technology Symposium,No.12-79.187-190.
Ray, M. H., Anghileri, M. and Mongiardini, M., 2008, Comparison of validation metrics using repeated Fullscale automobile crash tests, WCCM 8, Venice, Italy.
Tsunashima, H., Naganuma. Y., Matsumoto, A., Mizuma, T. and Mori, H., 2012, Condition Monitoring of
Railway Track Using In-Service Vehicle, Reliability and Safety in railway, Intech, 333-356.
Appendix : Coefficient matrices in Eq. (1)
Equations of motion of the vehicle model in vertical direction can be written as follows

M = diag mc I c / lc2 mt1 I t1 / lt21 mt 2 I t 2 / lt22 ,

2cs
0

cs
C =
0
cs

0
0

c p
D =
c p
0

708

cs

2cs
cs

cs

2(c p + cs )

2c p

cs

0
cp

0
0

cp
0
0

0
cp
cp

0
0
0

0
cp

c p ,

cs

0
cs
0
0
0

0
0
2(c p + cs ) 0

0
2c p ,

2k s
0

k s
K =
0
k s

0
0

k p
E =
k p
0

ks

0
0

2k s

ks

ks

2(k p + k s )

0
ks

0
0

2k p
0

0
kp

0
0

kp
0
0

0
kp
kp

0
0
0

0
kp

k p

ks

0
0
ks
0
0

0
0
2(k p + k s ) 0

0
2k p