Transportation Research Part C 93 (2018) 544–565

Contents lists available at ScienceDirect

Transportation Research Part C

journal homepage: www.elsevier.com/locate/trc

Position synchronization for track geometry inspection data via

T
big-data fusion and incremental learning☆
⁎
Yuan Wanga,b,c, Ping Wanga,b, , Xin Wanga,b, Xiang Liuc
a
School of Civil Engineering, Southwest Jiaotong University, Chengdu, China
b
Key Laboratory of High-speed Railway Engineering, Ministry of Education, Chengdu, China
c
Department of Civil and Environmental Engineering, Rutgers, The State University of New Jersey, NJ, USA

A R T IC LE I N F O ABS TRA CT

Keywords: Track geometry inspection data is important for managing railway infrastructure integrity and
Track geometry inspection operational safety. In order to use track geometry inspection data, having accurate and reliable
Railway position information is a prerequisite. Due to various issues identified in this research, the po-
Big data sitions of different track geometry inspections need to be aligned and synchronized to the same
Information fusion
location before being used for track degradation modeling and maintenance planning. This is
Position synchronization
Incremental learning
referred to as “position synchronization”, a long-standing important research problem in the area
of track data analytics. With the aim of advancing the state of the art in research on this subject,
we propose a novel approach to more accurately and expediently synchronize track geometry
inspection positions via big-data fusion and incremental learning algorithms. Distinguishing it
from other relevant studies in the literature, our proposed approach can simultaneously address
data exceptions, channel offsets and local position offsets between any two inspections. To solve
the Position Synchronization Model (PS-Model), an Incremental Learning Algorithm (IL-
Algorithm) is developed to handle the “lack of memory” challenge for the fast computation of
massive data. A case study is developed based on a dataset with data size of 18 GB, including 58
inspections between February 2014 and July 2016 over 323 km (200 miles) of tracks belonging
to China High Speed Railways. The results show that our proposed model performs robustly
against data exceptions via the use of multi-channel information fusion. Also, the position syn-
chronization error using our proposed approach is within 0.15 meters (0.5 feet). Our proposed
data-driven, incremental learning algorithm can quickly solve the complex, data-extensive, po-
sition synchronization problem, using an average of 0.1 s for processing one additional kilometer
of track. In general, the data analysis methodology and algorithm presented in this paper are also
suitable to address other relevant position synchronization problems in transportation en-
gineering, especially when the dataset contains multiple channels of sensors and abnormal data
outliers.

1. Introduction

Track geometry defects are considered one of the most important factors in operational stability and safety (Esveld, 2001; Higgins
and Liu, 2017; Liu et al., 2013; Quiroga and Schnieder, 2012). Track geometry data from track inspection cars is useful for railway
maintenance. There are multiple inspection channels corresponding to different types of track geometry, and each channel relates to a

☆
This article belongs to the Virtual Special Issue on “Big Data Railway”.
⁎
Corresponding author at: School of Civil Engineering, Southwest Jiaotong University, Chengdu, China.
E-mail address: [email protected] (P. Wang).

https://doi.org/10.1016/j.trc.2018.06.018
Received 27 September 2017; Received in revised form 22 June 2018; Accepted 27 June 2018
Available online 07 July 2018
0968-090X/ © 2018 Published by Elsevier Ltd.
Y. Wang et al. Transportation Research Part C 93 (2018) 544–565

Table 1
Selected track inspection parameters and methods.
Channel Type Sketch Map Measurement Method Sensors

1 Gauge Fig. 1 (1) Laser ranging Laser and displacement transducer

2 Longitudinal profile (two sides) Fig. 1 (4) Inertial method Accelerometer and displacement transducer
3 Alignment (two sides) Fig. 1 (3)
4 Crosslevel Fig. 1 (2) Automatic acceleration compensation Accelerometer
5 Warp (twist) – Difference of crosslevel with a distance of 3 meters Calculated from crosslevel

specific type of sensor. Taking the GJ-4 track inspection car of the Chinese Ministry of Railways as an example, some track inspection
parameters are listed in Table 1 (Ren et al., 2010). The illustrative sketches of different types of track geometry are shown in Fig. 1.
Each channel corresponds to a specific type of track geometry.
There have been many studies based on track inspection data, including data measurement (Haigermoser et al., 2015; Weston
et al., 2007; Bocciolone et al., 2007; Tsunashima, 2008), track condition evaluation (Tsunashima, 2008; Alfelor et al., 2001; Sadeghi,
2010; Sadeghi and Askarinejad, 2011) and track degradation prediction (Kawaguchi et al., 2005; Bartram et al., 2008; Liu et al.,
2010; Xu et al., 2011, 2012; Xu, 2012; Selig et al., 2008). Nearly all the methods and models require high quality inspection data. The
use of raw track geometry inspection data from the track geometry car is not always valid due to various data issues, such as
measurement errors, abnormal data outliers and positional errors. Among these errors, milepost positional error is one common issue,
requiring extensive effort to match and align the positions of the same inspected location from multiple inspections (Xu, 2012; Selig
et al., 2008; Qu, 2012; Xu et al., 2013). This effort is not trivial because of the need for estimating and predicting location-specific
track geometry deterioration in railroad track maintenance planning. This paper aims to address position synchronization problem
from different inspection runs, with a high precision and computational efficiency. The research outcomes can be used for all types of
railway systems, particularly high-speed railways, whose track asset management demands a high accuracy in positional information.
In practice, an initial milepost can be manually selected. The subsequent mileage information is obtained according to the
rotation angles (by counting the grating encoder impulse number) and the wheel radius (Allotta et al., 2002), as illustrated in Fig. 2a.
However, there are inevitable positional errors caused by radial errors of the wheels, faulty encoder output (Qu, 2012), degraded
adhesive conditions (Soleimani and Moavenian, 2017, Liu and Bruni, 2015) or track geometry irregularities (Fig. 2d). Due to these
factors, the positional error accumulates. To address these issues, the Global Positioning System (GPS) (Specht et al., 2017;
Tsunashima, 2008; Allotta et al., 2002), Differential GPS (DGPS) (Allotta et al., 2002; Hanreich et al., 2002) and radio-frequency
identification (RFID) (Yang, 2009) are introduced as an absolute reference to control the accumulation of positional errors.
Even though many advanced techniques and devices are used, the positional errors cannot be eliminated and can sometimes reach
100 meters (328 feet). Furthermore, other environmental conditions could lead to abnormal data points. For example, a film of water
from rain on the rail-head can cause laser sensor malfunction (Fig. 2c). This kind of abnormal data outlier may influence the
performance of the position synchronization method. In Section 2, we review the related work in the literature that addresses this
research problem, the respective merit and limitations of each method and the intended contributions of our proposed new approach
to the body of knowledge.

2. Related prior work

Positional errors can be classified into three categories, which are (1) absolute position errors (APE); (2) relative position errors
(RPE); and (3) channel-inside position offset (CPO). Since our study focuses on position synchronization of data from different runs
with multiple measurement channels, our review focuses on RPE and CPO. It should be noted that position synchronization is only
focus on RPE and CPO. The track inspection dataset used in this paper has undergone a preliminary processing based on the Key
Equipment Identification (KEI) model proposed in Xu et al. (2013).

Alignment mid-line of rail

top-line of rail

Gauge Zero crosslevel

Crosslevel

Longitudinal profile

Fig. 1. Schematic diagrams of track gauge, crosslevel, alignment and.

545
Y. Wang et al. Transportation Research Part C 93 (2018) 544–565

Track geometry
Weather irregularity
Inspection car
Rains

Track

Laser sensors

wheel Rail
move

film

Slipping

Fig. 2. Positioning principle of the inspection car and the causes of the positional errors. (a) shows the mileage measuring principle; (b) shows the
degraded adhesive conditions that leads to relative slippage between rail and wheel; (c) shows unforeseen influences such as weather; (d) track
geometry irregularity.

2.1. Absolute Position Error (APE)

The positional difference of the inspection data compared to its actual position is called the Absolute Position Error (APE). The
absolute positional accuracy is important when a worker attempts to locate a track defect that is observed in the inspection. The APE
magnitude is determined by absolute reference along the track. Since there are inevitable errors in the selected reference, the APE
cannot be eliminated. In addition to using better inspection technologies, researchers have developed mathematical models to deal
with APE. For example, optimization models were established to minimize the sum of the squares of the difference between two sets
of selected track geometry data from a certain section (Sui, 2009). Models are established to correct the local offset of curve sections
based on the least square method and correlation coefficient (Li and Xu, 2010; Pedanekar, 2006). More recently, a key equipment
identification model was proposed to correct the APE of track inspection data by combining the real mileage information of track
equipment, which can reduce the APE to under 5 meters (Xu, 2012; Xu et al., 2013). There are more details regarding APE in Vu et al.
(2012).

2.2. Relative position error (RPE)

The mileage difference in inspection data between different runs is known as the Relative Position Error (RPE). The RPE exists
because of uncertain rail-wheel contact profiles for different inspection cars running along the same track section. The magnitude of RPE
is determined by comparing the data from multiple inspections. In the scope of this paper, the process to correct RPE is treated as a
problem of position synchronization. The literature concerning RPE is summarized in Table 2. Addressing RPE is one focus of this paper.
Some common issues with the methods mentioned in Table 2 can be summarized as follows:

• A quantitative assessment model of the RPE is lacking. A general approach to address the position error is to observe the coin-
cidence of waveforms from graphs (Xu et al., 2013; Sui, 2009; Li and Xu, 2010). Xu (2012) applies two indirect assessments,
including the correlation coefficient and summation of gauge change between the measured waveforms of two inspections, to
address the performance of mileage corrections. Xu et al. (2015) and Xu et al. (2016) present an indirect measurement by using
standard deviation of the inspection data of different runs. A smaller standard deviation indicates better performance of the
model.
• The reference for position synchronization is determined both subjectively and empirically. Xu et al. (2015), Sui (2009) and Xu
et al. (2016) use the latest set of the previous inspection data as a reference to synchronize the current inspection data. Li and Xu
(2010) and Pedanekar (2006) use a reference data library or some static files, which are generated by railway operators and will
be updated when maintenance is carried out.
• All the aforementioned methods are based on a default assumption that the inspection data has no data exception issues, or the
dataset is thought to be exception-free after preprocessing. However, that is not always the case. Some models may become
invalid or even lead to erroneous positioning results, since the position error due to data exceptions in the reference data will
spread to current processed data.
• Only a single measurement channel is used for position synchronization in the above literatures. Abnormal data points may exist
in one channel, but the probability of data exceptions in all channels at the same position is low. The performance can be
enhanced by fusing data from multiple channels in one unified model (Section 5.3).

546
 Y. Wang et al.

Table 2
Previous studies related to RPE.
Reference Application scenarios Quantitative assessment Main techniques Reference for position Consideration of data Channels in use
model of RPE synchronization exceptions

Xu (2012) Railway Indirect assessment Dynamic time warping The latest inspection run Preprocessing Track gauge
Xu et al. (2015) Railway Indirect assessment Dynamic time warping The latest inspection run Preprocessing Track gauge
Sui (2009) Railway Not mentioned Optimization The latest inspection run Not mentioned Profile, alignment or gauge (one

547
channel)
Li and Xu (2010) Railway Not mentioned Least square method Reference data library Not mentioned Profile, alignment or gauge (one
channel)
Pedanekar (2006) Railway Not mentioned Correlation coefficient Static files Not mentioned –
Xu et al. (2016) Subway Indirect assessment Dynamic Programming and Cross- The latest inspection run Preprocessing Profile, alignment, gauge, twist, and
correlation cross-level (separately)
Transportation Research Part C 93 (2018) 544–565
Y. Wang et al. Transportation Research Part C 93 (2018) 544–565

Inspection Car A

Different distribution of sensors

Inspection Car B

Fig. 3. Sketch map of the CPO and track geometries. C#1 and C#2 represent two different types of track geometry; d1 and d2 are the relative distances
between the measurement locations of C#1 and C#2, for inspection Car A and Car B, respectively.

2.3. Channel-inside Position Offset (CPO)

The Channel-inside Position Offset (CPO) is the position difference of inspection channels between different inspection runs. It is
derived from differences in the distribution of sensors between different inspection cars. Generally, the sensors are not mounted at the
same location of the inspection car. As illustrated in Fig. 3, two different types of inspection car, Car A and Car B, measure the same
types of track geometry, C#1 and C#2. The relative distances between the measurement locations of sensors, d1 and d2, can be
different. The CPO is the difference between d1 and d2. The existence of CPO between inspection Car A and Car B is caused because
the two inspection cars are designed differently.
To our knowledge, there are no published studies concerning CPO. The concept of CPO in the aforementioned Xu (2012), Xu et al.
(2015), Sui (2009), Li and Xu (2010), Pedanekar (2006) and Xu et al. (2016) does not exist since they only consider the data from a
single measurement channel. On the contrary, in this paper we attempt to deal with data exception issues by fusing data from
multiple sensors into one unified model. Because it is very rare that all sensors malfunction at the same time (unless there are
inspection-car-specific failures), multi-sensor fusion can make our proposed algorithm more robust against sensor data outliers in
aligning track inspection data positions.

3. Contributions and organization of this paper

Considering all the limitations of the prior research presented in Sections 2.2 and 2.3, this paper develops a novel approach to
addressing the problem of position synchronization via big data fusion and incremental learning algorithms.
Firstly, the local milepost offset and waveform similarity between every two inspection runs are estimated based on definitions
given in Section 5.2. Secondly, to deal with data exceptions, the inspection data of multiple measurement channels are fused by MCF-
model in Section 5.3. The channel fusion process improves the precision of position synchronization and enhances the robustness
against abnormal data. The side-effect is that the required memory and computation time increase a lot comparing to previous
methods (Xu et al., 2015, 2016; Sui, 2009; Li and Xu, 2010; Pedanekar, 2006), where only the data of one single channel is used.
As a countermeasure, in Section 6, we propose the concept of knowledge library, which represents the minimal information refined
from the overall information with a high-dimensional structure. The refining process compresses the overall information into several
low-dimensional vectors and matrices by statistical approaches. Whenever a new dataset is obtained from the track inspection car,
the RPE and the CPO of the newly measured data are estimated as referring to historical knowledge, and then position synchroni-
zation is carried out to reduce the RPE and CPO. In return, the knowledge library will be updated. The implement of the updating
process of knowledge library is called IL-algorithm.
For a better understanding of this paper’s work, its organization is illustrated in Fig. 4. The contents with a red border are the main
contributions, which are also summarized as the following six points.

• The inspection data from multiple measurement channels are fused and synchronized through a multi-channel fusion model
established in Section 5.3. A quantitative assessment of RPE is achieved through an optimization model proposed in Section 5.4.
• This paper presents a novel approach to dealing with data exceptions through the establishment of three matching criteria with
the thresholds determined through statistical methods presented in Section 5.5.
• The position synchronization operation is achieved through a two-phase interpolation approach presented in Section 5.6.
• An incremental learning algorithm is developed to execute position error estimating, multi-channel fusing and position syn-
chronizing processes in Section 6. The required computational time is minimized via these advanced data processing and analytic
algorithms.
• A case study is developed based on real-life data from part of the China High Speed Railway to demonstrate the practical values of
this research for industrial practice (Section 7).
• The tradeoff between the computation efficiency and the accuracy of position synchronization is discussed in Section 8.

548
Y. Wang et al. Transportation Research Part C 93 (2018) 544–565

Section 1 Section 5 Section 6 Section 8

Introduction Methodology Algorithms Discussion

OI-Algorithm
Definitions of Section 6.1
LMO, OIM, RMP and UMP
Section 2 Section 5.2 Tradeoff between
IL-Algorithm
execution time and
Section 6.2
Multiple Channel Fusion precision
Related Prior Work (MCF-Model)
Section 5.3 Section 7
Case study
Section 3 Reliable Milepost Estimation
(RME-Model) Multi-Channel fusion
Section 5.4 Section 7.1 Section 9
Contributions and
Organization Robustness against abnormal data
Bad Matching Criterion
Section 7.2
(BMC-Model)
Section 5.5 Precision and number of iterations
Section 4 Conclusion
Section 7.3
Position Synchronization
Abbreviations and (PS-Model) Computational time
Nomenclature Section 5.6 Section 7.4

Fig. 4. Organization of this paper.

4. Abbreviations and nomenclature

See Table 3.

5. Methodology

5.1. Technical framework

The purpose of this section is to introduce the overall model framework, which contains five parts, one part is model preparation
with some definitions and four sub-models, as illustrated in Fig. 5. The four models, Multiple Channel Fusion Model (MCF-Model),
Reliable Milepost Estimate Model (RME-Model), Bad Matching Criterion Model (BMC-Model) and Position Synchronization Model
(PS-Model), are established in order of the dependency relationship. The purposes of each process are presented below.

• The first process is aimed at estimating the local offsets and waveform similarities between every two runs of inspection data.
(Section 5.2)
• The MCF-Model is established to estimate and fuse the CPO among different inspection channels. (Section 5.3)
• The RME-Model is established to estimate the RPE by fusing the output results of the MCF-Model. (Section 5.4)
• The BMC-Model is established to deal with data exceptions by filtering the CPO and RPE. Three criteria are defined and the
thresholds are given according to probability distributions of local offsets and waveform similarities. (Section 5.5)
• The PS-Model is established to conduct position synchronization operation according to a two-phase interpolation. (Section 5.6)
5.2. Fundamentals of the models

This section presents some basic definitions concerning the local offsets between every two runs, including Local Milepost Offset
(LMO), Overall Information Matrix (OIM), Reliable Matching Point (RMP) and Unreliable Matching Point (UMP).

5.2.1. Local Milepost Offset

Local Milepost Offset (LMO) is defined to describe the milepost difference between two data samples, data #1 and data #2 in
Fig. 6. If there is no milepost offset between the two sets of data, this situation is referred to as an ideal mapping, illustrated in Fig. 6a.
Real mapping is different from ideal mapping because of RPE. For example, the data sample R1 in data #1 and R2 in data #2 share the
same milepost range, as in curve A-B in Fig. 6a. When R1 is longer than R2, it indicates data #1 is compressed in comparison to data
#2 within the A-B range. Therefore, the LMO is defined as the difference between real mapping and ideal mapping, see Fig. 6b.
The following presents a mathematical definition of LMO. For discrete signal X = {x i |i = 1, 2, …, N } and Y = {yi |i = 1, 2, …, N } , in
considering a local waveform matching scale s, the local samples of X and Y at location k are defined as

{ s
⎧ X (s, k ) = x i |k− 2 + 1 ⩽ i < k +
s
2 }
⎨ Y (s, k ) = {y |k−
⎩
i
s
2
+1⩽i<k+
s
2 } (1)
For i〈1 or i〉N , x i , yi = 0 the local offset δ (s, k ) and waveform similarity ρ(s, k ) of Y to X are defined as follows

549
Y. Wang et al. Transportation Research Part C 93 (2018) 544–565

Table 3
Abbreviations and Nomenclature used in this Paper.
Abbreviations Explanation

APE Absolute Position Error

RPE Relative Position Error
CPO Channel-inside Position Offset
RMP Reliable Matching Point
UMP Unreliable Matching Point
RME Reliable Milepost Estimate
LMO Local Milepost Offset
OIM Overall Information Matrix
MCF-Model Multiple Channel Fusion Model
RME-Model Reliable Milepost Estimate Model
BMC-Model Bad Matching Criterion Model
PS-Model Position Synchronization Model
OI-Algorithm Overall-Iterative Algorithm
IL-Algorithm Incremental-Learning Algorithm

Notation Explanation

s The scale parameter for waveform matching

ds The step for waveform matching, 0.25 m per point; point by point if ds = 1.
δ Local mileage offset
ρ Local similarity, ranging from [−1, 1].
k Refers to a mileage position
c , c∗ A data channel, c∗ indicates the best channel to be selected.
δ (s, k ) (X , Y ) The mileage offset of X to Y at location k under scale s.
ρ(s, k ) (X , Y ) The similarity of X to Y at location k under scale s.

δij(k, c) The mileage offset of the jth run to the ith run at location k based on data from channel c.

ρij(,kc) The similarity of the jth run to the ith run at location k based on data from channel c.

Δ(ck ) (X ) , Δc (X ) The local offset matrix and overall offset matrix at location k based on data from channel c. Where X is a matrix containing data of multiple
runs.
ϒ (ck ) (X ) , ϒc (X ) The local similarity matrix and overall similarity matrix at location k based on data from channel c. Where X is a matrix containing data of
multiple runs.
U = {(δi(,kj,)c, ρi(,kj,)c )|i, j ∈ [1, n];k ∈ [1, N / ds];c ∈ [1, t ]} , the Overall Information Matrix. Also written as {(Δc (X ), ϒc (X ))|c = 1, 2, …, t } .
Ci, j, c The overall channel offset of the jth run to the ith run for channel c.
Di, c The estimation of overall channel offset for the ith run.
di(k ) , di(k ) ∗ The RME of the ith run at the location k. * indicates the best selected channel.
(k ) ∗
di̇ The first-order derivative of di(k ) ∗.
δ0 ,ρ0 ,δ1 The thresholds ofδij(k, c) , ρij(,kc) and di̇
(k ) ∗
for the BMC.
rUMP card ({bo (δi(,kj ), ρi(,kj ) ) = 0 | i, j ∈ [1, n];k ∈ [1, N / ds];c ∈ [1, t ]})
= , the proportion of the UMPs.
card (U )

ui(k, n) The average of the LMO of the ith run at location k, considering a total of n runs of inspection data.

ni(k, n) The count of the RMPs of the ith run at location k, considering a total of n runs of inspection data.

K (k, n) = {(ui(k, n), ni(k, n) )|i, j ∈ [1, n];k ∈ [1, N / ds]} , local information matrix.
fr The distribution of variable r, where r can be δ , ρ or d .̇
M A parameter of IL-Algorithm, which represents the maximal number of inspection runs that are to be cached in memory.
Pr (p) The precision of position synchronization at a given confidence level p.

Operator Explanation

norm(x ) = ‖x‖2 = ∑ |xi |2 , vector norm.

cov(x , y ) = ∑ (xi−x )(yi −y ) , covariance.
Percentile(x , p) The percentile value of x at a percentage of p .
corr(x , y ) =
cov (x , y )
, normalized correlation coefficient.
norm (x − x )·norm (y − y )
xcorr2(A, B ) 2D cross-correlation of matrix A and B.
card({·}) Number of elements in set {·} .
interp(x , y, x ′) Resampling the sequence ( x , y ) by a new x ′ through interpolation.
bo(δ , ρ) 1, if ρ ⩾ ρ0 and |δ| ⩽ δ0
=⎧ , exception judgment operation.
⎨
⎩ 0, otherwise

550
Y. Wang et al. Transportation Research Part C 93 (2018) 544–565

INPUT

Descriptions
1
Definitions of Estimate the local offset and similarities
LMO, OIM, RMP and UMP between any two inspection runs at
Section 5.2 different locations
2
Multiple Channel Fusion Quantitative assessment of
Estimate CPO between different channels,
(MCF-Model) RPE and CPO
Section 5.3 then select the best channel

3
Reliable Milepost Estimation
Estimate of the reliable milepost offsets
(RME-Model)
Section 5.4
for each run

4
Bad Matching Criterion Deal with data exception. Determine the
(BMC-Model) thresholds according to probability Data exception consideration
Section 5.5 distributions of local offset and similarity
5
Position Synchronization Achieve position synchronization Conduct the operation to reduce
(PS-Model) RPE and eliminate CPO
according to a two-phase interpolation.
Section 5.6

OUTPUT

Fig. 5. Overall model framework.

compression section

extreme offset point

Fig. 6. LMO between data #1 and data #2.

(s, k ) (X , Y ) ≜ = argmax (s, k + θ ) , Y (s, k ) )|

⎧δ −Δ ⩽ θ ⩽ Δ|cov (X
⎨ ρ(s, k ) (X , Y ) ≜ = corr (X (s, k + δ ) , Y (s, k ) )
(s , k )
⎩ (2)

where Δ is a limitation of possible offset range.

It can be interpreted that the local offset δ (s, k ) is defined as a local shift θ that causes X (s, k + θ) to experience the strongest covariance
with Y (s, k ) within the range of [−Δ, Δ]. The similarity ρ(s, k ) is defined as the normalized correlation coefficient of X (s, k + δ (s, k )) and Y (s, k ) .
Introducing a matching step ds (ds ⩾ 1) to sample LMO discretely, the location parameter k is specified as a natural number within
the range of [1, N / ds]. It should be noted that δ (s, k ) (X , Y ) may not equal δ (s, k ) (Y , X ) when X and Y are unevenly stretched or
compressed around the location k.

5.2.2. Overall information matrix

When considering a dataset of n (n ⩾ 2) runs {X1 , X2 , …, Xn } , with each containing t channels, all intermediate results, including
local offset and waveform similarity, can be gathered into a high-dimension matrix
U = {(δi(,kj,)c , ρi(, kj,)c )|i, j ∈ [1, n];k ∈ [1, N / ds];c ∈ [1, t ]} , which is named the Overall Information Matrix (OIM). The detailed structure of
U can be found in Appendix A. The formation process of U is highly parallel since the elements in U are independent.

551
Y. Wang et al. Transportation Research Part C 93 (2018) 544–565

5.2.3. Reliable and unreliable matching point

In practice, for some reasons, such as abnormal data, maintenance carried out by heavy machinery or track condition degradation,
a bad correlation will exist at some milepost points, namely ρ(s, k ) (X , Y ) ≪ 1. Therefore, the similarity threshold ρ0 is introduced to
judge whether a matching exception exists at location k. Another situation is also taken as a matching exception when the estimated
local offset exceeds a given threshold δ0 . The determination of similarity threshold is based on the probability distribution, which is to
be addressed in Section 5.5.4. As a result, the exception judgment operation bo(δ , ρ) is defined as follows:

1, if ρ ⩾ ρ0 and |δ| ⩽ δ0
bo(δ , ρ) ≜ = ⎧
⎨
⎩ 0, otherwise (3)
bo(δi(,kj ),
If ρi(, kj ) )
= 1, it indicates that the milepost point at location k is reliable according to the inspection data from the ith and jth
runs; we call this kind of matching point a Reliable Matching Point (RMP). Otherwise it is an Unreliable Matching Point (UMP), and
should be ignored. The ratio between the number of UMPs and the number of all elements in U is denoted as rUMP . The ratio rUMP
represents the proportion of bad matching points. Moreover, the larger the ratio rUMP is, the poorer the repeatability of track geometry
is, so that the ratio rUMP is an important index to describe the reliability of position synchronization.

card ({bo (δi(,kj ), ρi(, kj ) ) = 0|i, j ∈ [1, n];k ∈ [1, N / ds];c ∈ [1, t ]})
rUMP ≜ =
card (U ) (4)

5.3. Multiple Channel Fusion Model (MCF-Model)

The purpose of this section is to estimate the CPO of each inspection run and establish the model for fusing multiple inspection
channels. All inspection channels provide information for position synchronization after the corresponding CPO is corrected. The
fusion of multiple channels is the essential part of the models in this paper to deal with data exception.
The estimation of CPO according to the overall information matrix U is presented in Fig. 7. The cube on the left side of Fig. 7
represents the overall offset matrix Δc1 (X ) , please refer to ①. Subscript c1 indicates the channel of gauge. The right cube represents
matrix Δct (X ) , see ②. Subscript ct indicates any other channel but that of gauge. The matrixes Δc1, k (X ) and Δct , k (X ) are slices at
location k of Δc1 (X ) and Δct (X ) , respectively, please refer to ③ and ④. The two vectors δi, j, c1 and δi, j, ct are extracted from Δc1 (X ) and
Δct (X ) , respectively, please refer to ⑤ and ⑥. In an ideal situation, δi, j, c1 and δi, j, ct should be coincident, while in actuality there may
exist a constant deviation. The difference between δi, j, c1 and δi, j, ct is the CPO between the ith and jth run for channel ct, that is denoted
as Ci, j, ct .Ci, j, ct , the constant offset of channel ct between the ith and jth run, can be obtained by solving Eq. (5).
argminCi, j, ct = ‖δi, j, c1−δi, j, ct −Ci, j, c ‖2 ; c = 2, 3, …, t (5)

The best estimation of the channel offset of the ith run, denoted as Di, c , can be obtained by solving Eq. (6).
argmin Di, c = ‖Ci, j, ct −Di, c ‖2 ; i = 1, 2, …, n; c = 2, 3, …, t (6)
Finally, the final local mileage offset δij(k ) ∗ between the ith and jth runs at location k can be obtained by subtracting Di, c∗ from δij(k, c)∗,
and the final waveform similarity ρij(k ) ∗ equals to the one estimated within the channel c∗, as presented in Eq. (7).
(k ) ∗
⎧ δij = δij(k, c)∗−Di, c∗
⎨ ρ (k ) ∗ = ρ (k )∗
⎩ ij ij, c (7)

Fig. 7. Estimation of channel offset.

552
Y. Wang et al. Transportation Research Part C 93 (2018) 544–565

5.4. Reliable Milepost Estimation Model (RME-Model)

The purpose of this section is to establish a model to estimate the reliable milepost offsets for each run of inspection data. All data
from other inspection runs are taken as a reference to determine the reliable milepost offset for each run. The reliable milepost
estimation (RME) of the ith run at location k is denoted as di(k ) . The RME-Model relies on two principles:

• The RME minimizes the sum of squared differences between the ith and other inspection runs;
• The sum of RME of all runs at each location k equals zero.
Therefore, the RME-Model can be described as an optimization model given in Eq. (8).
n n
⎧ min=
⎪ ∑ ∑ (δij(k) ∗−di(k) )2
⎪ i=1 j=1
⎨ n
⎪ ∑ di(k ) = 0
⎪ (8)
⎩ i = 1

Eq. (8) contains an optimization objective and a constrain, which are corresponding to the two principles given above. Eq. (8) is a
constrained least squares problem, which can be solved using the Augmented Lagrangian method, please refer to Appendix B.

5.5. Bad Matching Criterion Model (BMC-Model)

Bad matching points, such as the UMPs, lead to invalid estimations of local offsets and have a great influence on the position
synchronization of inspection data. Especially in the event of a large erroneous estimation of local offset which is not rectified, the
waveform will distort, the effects of which are all but irreversible. The purpose of this section is to minimize the probability for UMPs
to be treated as RMPs through the establishment of the BMC-Model. In this paper, the BMC-Model includes three parts: (1) amplitude
criterion, (2) similarity criterion, and (3) 1st derivative criterion. Multiple criteria can achieve better performance against the fault
matching points.

5.5.1. Amplitude criterion

The amplitude criterion is used for the local offsets, and it works when the magnitude of a local offset is beyond the given
threshold δ0 , a limitation not likely to be exceeded. It is especially useful when the inspection data waveform is strongly periodic. The
amplitude criterion is expressed by Eq. (9).

|δi(,kj )| ⩽ δ0 (9)

5.5.2. Similarity criterion

The similarity criterion is used for the local waveform similarity, and it works when the similarity of two waveforms is sig-
nificantly less than 1. A low similarity indicates unreliable waveform matching. The similarity criterion is expressed by Eq. (10).

|ρi(, kj ) | ⩾ ρ0 (10)

5.5.3. 1st derivative criterion

Unlike the above two criteria, the 1st derivative criterion is used for the RME-Model. It works when the change rate of the RME
along the railway is faster than a given threshold δ1, a limitation not likely to be exceeded. The 1st derivative criterion is expressed by
Eq. (11).
(k ) ∗
|di̇ | ⩽ δ1 (11)
(k ) ∗ (k ) ∗
where di̇ is the first order derivative of di(k ) ∗
along the railway, namely the change rate of the RME. di̇
is defined by Eq. (12). It
should be noted that, when δ1 < 1, the 1st derivative criterion is able to guarantee the monotonicity of the newly generated position
coordinates in Section 5.6.
(k ) ∗ d
di̇ = (di(k ) ∗)
dk (12)

5.5.4. Determination of δ0 , ρ0 and δ1

The amplitude threshold δ0 and similarity threshold ρ0 are determined based on the joint probability distribution of δ and ρ . As
the δ0 becomes larger, it is more likely for a UMP to be mistakenly taken as an RMP. Generally, for a lager local milepost offset, the
corresponding similarity will also be larger. In practice, the aim to reduce the probability of false matching can always be achieved by
increasing the values of the thresholds δ0 and ρ0 .
The joint probability distribution of the local milepost offset and similarity is illustrated in Fig. 8. The log-normal distribution of ρ

553
Y. Wang et al. Transportation Research Part C 93 (2018) 544–565

Fig. 8. The joint probability distribution of local milepost offset and waveform similarity. (b) Is an enlarged view of the range in the highlighted
square of (a). The scattered points in the highlighted circles indicate the UMPs.

is given in (1−ρ) ∼ lnN (−2.9, 0.82) . In this paper, the thresholds δ0 and ρ0 are given according to the 99% confidence level, or stricter
(k ) ∗
95% confidence. As for the threshold δ1, the distribution of di̇ can also be obtained from the overall information matrix U.
Similarly, the value of δ1 can be obtained according to the 99% confidence level, or stricter 95% confidence.
It should be noted that the confidence level can be taken as a percentage threshold, which is given empirically according to
subjective experience. Higher confidence level leads to more misjudgments of Reliable Matching Points (RMPs) and fewer mis-
judgments of Unreliable Matching Points (UMPs), and vice versa. The percentage should not be too large or too small. In this paper,
two percentages are suggested, 95% and 99%.

5.6. Position Synchronization Model (PS-Model)

The purpose of this section is to establish a model to conduct position synchronization according to the estimated RPE and CPO.
Assuming di(k ) ∗ is the best estimated mileage offset of the ith run at location k, the position synchronization can be achieved by
moving the data point from location k to a new location of k + di(k ) ∗. The channel offset Di, c for different runs needs to be included to
ensure channel synchronization. The process can be described as a two-phase interpolation approach that can be written as Eq. (13).

Xi∗, c = interp (interp (dm, dm + di(k ) ∗ + Di, c , d v ), Xi, c , d v ); i = 1, 2, …, n; c = 1, 2, …, t (13)

where Xi∗, c is the synchronized data, and d v = (1, 2, …, N )T . dm = {(i−1)·ds + 1|i = 1, 2, …, N / ds} , is the matching position sequence
with a step of ds .
The first interpolation (the inside interpolation in Eq. (13) is aimed at generating a new position coordinate according to the
mapping of (dori, dnew ) . Since the new position coordinate does not share the same sampling rate with that of the original, the second
interpolation (the outside interpolation in Eq. (13)) is conducted to achieve data resampling by d v . Piecewise linear interpolation is
adopted in the first interpolation, while both piecewise linear interpolation and cubic spline interpolation can be used in the second
interpolation.
The PS-Model process is illustrated in Fig. 9. The solid line in Fig. 9(a) represents the estimated RME of one run of inspection data.
The estimated di(k ) ∗ are taken as a discrete sampling of RME. The piecewise linear interpolation for the first interpolation of Eq. (13) is
equivalent to a linear approximation of the RME. The second interpolation of Eq. (13) can reduce the RME, as shown in Fig. 9(b). The
residual RME is the interpolation remainder of the first interpolation of Eq. (13).
For x ∈ [x 0 , x1], the linear interpolation remainder is expressed as:

f ′ (ξ )
R (x ) = (x −x 0)(x −x1)
2 (14)

It can be seen from Eq. (14) that, theoretically, the residual RME will converge towards 0 as |x1−x 0 | approaches 0. In this paper, the
value of |x1−x 0 | equals the matching step ds . However, the smaller |x1−x 0 | is, the more calculations there will be. The optimization of
computation efficiency and precision will be discussed further in Section 8.

6. Computational algorithms

The purpose of this section is to develop algorithms to achieve the estimating, fusing and synchronizing processes. A direct
solution is proposed in this paper through an Overall-Iteration Algorithm (OI-Algorithm). As an improvement, the Incremental-
Learning Algorithm (IL-Algorithm) is developed to handle the “lack of memory” and massive computation challenges.

554
Y. Wang et al. Transportation Research Part C 93 (2018) 544–565

Fig. 9. Illustration of the PS-Model process.

6.1. OI-Algorithm

Considering that deviations may exist in the RME values due to distortions of the original waveform in the inspection data, instead
of simply decreasing the matching step ds , further iterations of the PS-Model on the newly processed results can more efficiently
enhance the position synchronization performance. According to Eq. (14), the farther the position is to the interpolation points, the
larger the residual RME becomes. A new iteration for position synchronization can be better when it avoids the previous RMPs. It is
achieved in this paper by changing the beginning position of the matching sequence dm according to the iteration number, as in Eq.
(15).

s N
dm∗ = ⎧(i−1)·ds + (j−1)· +1 i = 1, 2, …, ; j = 1, 2, …, No ⎫
⎨
⎩ No ds ⎬
⎭ (15)

where dm∗represents the matching sequence with multiple iterations; No is the overall number of iterations; j is the iteration number.
As a result, the OI-Algorithm, which takes the simultaneous input of data from all inspection runs, contains several iterations and
in each iteration the four models are solved in order. A new iteration is carried out on the results of the previous iteration (or the
original data for the first iteration). The flow chart of the OI-Algorithm is presented in Appendix C.
The implement of OI-Algorithm is exactly corresponding to the data processing flow provided in Fig. 5, the overall model fra-
mework. From the calculation steps illustrated in Fig. 17, it can be found that the input dataset will undergo the five sub-procedures
given in Sections 5.2-5.6, one after the other until the final result is obtained.
Though the OI-Algorithm is a complete implementation of the four models presented in Section 5, there are two main challenges
for practical application of the algorithm.

• Challenge 1: lack of memory. Take a 300 km track section for example, the required buffer memory reaches 12 GB to cache the
data of 50 inspection runs. More memory is needed to carry out the OI-Algorithm. The size of the overall information matrix U is
50 × 50 × 10,000 × t for a matching step of 30 m, where t is the number of inspection channels in use. U is larger when pro-
cessing data from more inspection runs.
• Challenge 2: massive computations. Each time a new inspection run is conducted, the OI-Algorithm needs to be re-executed
with massive repeated calculations. A direct countermeasure can be helpful to reduce repeated computation, such as storing the
matrix U. Nevertheless, with an increasing number of inspection runs, it is uneconomical to store such a quantity of intermediary
data.

An incremental algorithm is of great importance to achieve the models instead of the OI-Algorithm. The aim is not simply to
transform the OI-Algorithm into an incremental style, but to minimize the required memory, and optimize computations through
incremental probability estimation and the establishment of a “Knowledge Library”.

6.2. IL-Algorithm to improve the OI-Algorithm

6.2.1. Knowledge library

In fact, it is not necessary to store all elements in the overall information matrix U. The concept of a “knowledge library”
represents the minimal information refined from the high-dimensional matrix U. The refining process compresses U into several low-
dimensional matrices or vectors through statistical approaches. Each time a new set of data is obtained from the track inspection car,

555
Y. Wang et al. Transportation Research Part C 93 (2018) 544–565

the

Fig. 10. The refining processing for local knowledge.

the RPE and CPO of the newly measured data is estimated by reference to the historical knowledge. In return, the knowledge will be
updated.
The knowledge library contains two parts: local knowledge and global knowledge. Local knowledge is the refined information
from the overall information matrix U for different inspection runs at different locations; this process is shown in Fig. 10.
Δ(ck ) (X ) represents the slice of Δc (X ) at location k. It should be noted that the UMPs should not be included, so the values δij(k ) are
excluded when bo(δij(k ), ρij(k ) ) = 0 . The average and the number of elements in the ith row of Δ(ck ) (X ) , denoted as ui(k ) and ni(k ) re-
spectively, are taken as local knowledge, as shown in Eq. (16).
n
⎧u (k ) = 1
bo (δij(k ) ∗, ρij(k ) ∗)·δij(k ) ∗
⎪ i ni(k )
∑
⎪ j=1
n
⎨ (k )
⎪ ni = ∑ bo (δij(k ) ∗, ρij(k ) ∗)
⎪ j=1 (16)
⎩
The value of bo(δij(k ) ∗, ρij(k ) ∗)
equals 1 for a RMP, and 0 for a UMP, so ni(k )
is calculated by summing up bo(δij(k ) ∗,
directly for ρij(k ) ∗)
j = 1, 2, …, n .
Global knowledge retains the overall features of the matrix U, including the statistical distributions of δ , ρ and d ,̇ that are denoted
as fδ [p], fρ [p] and fd ̇ [p], respectively. The range Lr (r = δ , ρ , d )̇ can be divided into 2N ′ parts. The number of parameters in each part p
are counted according to Eq. (17).

p p+1
Nr , p = card ⎛ ⎧r⎜ Lr ,1 ⩽ r < Lr ,2 ⎫ ⎞; p = −N ′, 1−N ′, …, N ′−1
⎟

⎝⎨⎩ N′ N′ ⎬
⎭⎠ (17)
where the boundaries of δ , ρ , and d ̇ are given as:

⎧ Lδ,(1,2) = [−Δ, Δ]
⎪
Lρ,(1,2) = [−1, 1]
⎨
⎪ Ld,(1,2) = [−Δ1, Δ1]
⎩ ̇ (18)
The total number of elements in U isNr = card({r|Lr ,1 ⩽ r < Lr ,2}) . Global knowledge fr [p] can be calculated according to Eq. (19).
Nr , p
fr [p] = ;
Nr (19)

6.2.2. Incremental implementation of the RME di(k ) ∗

According to Appendix B, the result di(k ) ∗ consists of two parts, di′ (k ) and d (k ) ∗, the incremental implementation of di(k ) ∗ can
therefore be derived from two parts, as expressed in Eq. (20).
n+1
1
di(k ) ∗ = di′ (k )−d (k ) ∗ = ui(k, n + 1)− ∑ ui(k, n + 1)
n+1 i=1 (20)

556
Y. Wang et al. Transportation Research Part C 93 (2018) 544–565

Fig. 11. Local knowledge updating process.

It should be noted that the di(k ) ∗ in Eq. (20) is the total estimation of reliable mileage offset based on the historical inspection data.
When the data of the (n + 1)th inspection run is included, the adjustment value is given in Eq. (21).
n+1
1 ⎛
di(k, n + 1) ∗ = di(k ) ∗−di(k, n) ∗ = (ui(k, n + 1)−ui(k, n) )− ∑ ui(k,n+1)−ui(k,n)⎞⎟
n + 1 ⎜ i=1
⎝ ⎠ (21)

where un(k+, n1 ) is defined as zero because there is no previous information for the (n + 1)th run.

6.2.3. Local knowledge updating

The updating process of local knowledge contains four parts, as presented in Fig. 11.

Process A. Obtain the local milepost offset {(δi(,kn)+∗1, ρi(, kn)+∗1)|i = 1, 2, …, N } by performing the waveform matching for data of the
(n + 1)th run to each of the previous n runs, respectively, using the model presented in Section 5.2.1.
Process B. Obtain the local milepost offset {(δn(k+)1,∗ i, ρn(k+)1,∗ i )|i = 1, 2, …, N } by performing the waveform matching for each of the
previous n to (n + 1)th runs, respectively, using the model presented in Section 5.2.1.

It should be noted that the aforementioned data from the previous n runs are the latest version after mileage synchronization.

Process C. The matrix K (k, n) is updated into K (k, n + 1) with {(δi(,kn)+∗1, ρi(, kn)+∗1)|i = 1, 2, …, N } . The formula is as follows:
(k , n + 1) 1
⎧ u = (ni(k, n) ·ui(k, n) + bo (δi(,kn)+ 1, ρi(, kn)+ 1)·δi(,kn)+ 1)
⎪ i ni(k , n + 1)

⎨ ni(k, n + 1) = ni(k, n) + bo (δi(,kn)+ 1, ρi(, kn)+ 1)

⎪ (22)
⎩
Eq. (22) is an incremental form of Eq. (16). The mean value ui(k, n + 1) is updated based on the previous value ui(k, n) and the newly
input δi(,kn)+ 1.

Process D. The (n + 1)th row of K (k, n + 1) is generated with {(δn(k+)1,∗ i, ρn(k+)1,∗ i )|i = 1, 2, …, N } . The formula is as follows:
n
⎧u (k, n + 1) = 1
bo (δn(k+)1,∗ j, ρn(k+)1,∗ j )·δij(k )
⎪ n+1 nn(k+, n1 + 1)
∑
⎪ j=1
n
⎨ (k, n + 1)
⎪ nn + 1 = ∑ bo (δn(k+)1,∗ j, ρn(k+)1,∗ j )
⎪ j=1
⎩ (23)

Eq. (23) is similar to Eq. (16), because there is no previous knowledge for the (n + 1)th inspection dataset.

6.2.4. Global knowledge updating

The global knowledge updating process can be achieved as Eq. (24).

557
Y. Wang et al. Transportation Research Part C 93 (2018) 544–565

1
f r(n + 1) [p] = ·(f (n) [p]·Nr(n) + Nr(,np+ 1) );r = δ , ρ , d ̇
Nr(n + 1) r (24)

where

p p+1
⎝
{
Nr(,np+ 1) = card ⎛ rij | Lr ,1 ⩽ rij <
N′ N′
Lr ,2;iorj = n + 1 ⎞
⎠
} (25)

Nr(n + 1) = Nr(n) + card({rij |Lr ,1 ⩽ rij < Lr ,2;iorj = n + 1}) (26)

6.2.5. IL-Algorithm
The flow chart of the IL-Algorithm is presented in Fig. 18 in Appendix D. The initialization of the knowledge library can be
achieved by applying the OI-Algorithm on M inspection runs. After the data of the (n + 1)th inspection run is processed (n ⩾ M ), the
data of the (n + 1−M )th inspection run is saved to the disk and released from memory, so there are only M + 1 inspection runs of
data in memory. Each time data from a new inspection run is imported, the four models are to be solved using the newly processed
historical data, rather than the original measured data. It should be noted that in this paper the data arrival order is arranged
according to the inspection date. The early inspection data is processed before the late one.
There are two main improvements in the IL-Algorithm compared to the OI-Algorithm:

• The IL-Algorithm only needs to cache data from M inspection runs, rather than all historical inspection data as in the OI-
Algorithm. The parameter M is determined according to the required precision and hardware environment, such as memory
capacity.
• Instead of caching the entire matrix U, only the knowledge library needs to be cached in memory. The knowledge library is
updated incrementally with few redundant calculations during the execution process of the algorithm.

7. China high speed railway case study

The purpose of this section is to demonstrate the performance and contribution of the proposed track inspection data position
synchronization methodology. It should be noted that there are a lot of results to be presented corresponding to the models given in
Section 5. However, this section only presents the most important and interesting points due to limited space. This section focuses on
four aspects, that are (1) multi-channel information fusion, (2) robustness against data exception, (3) precision and number of
iterations, and (4) computational time. The precision of position synchronization is defined as the percentile value of the estimated
LMO (δi(,kj ) ) at a given confidence level p, as given in Section 5.5.4.
A case study was carried out based on a dataset with 58 inspection runs (17.7 GB) between February/2014 and July/2016 on the
China High Speed Railway. Each inspection run contained 13 measurement channels, including milepost, gauge, crosslevel, long-
itudinal profile (left and right side), alignment (left and right side), superelevation, curvature, carbody acceleration (vertical and
lateral direction), speed and ALD (please refer to Xu et al., 2013 for a description of ALD). The length of the railway line was 323 km.
The sampling distance of the inspection car was 0.25 m. Some default parameters of the IL-Algorithm are: matching scale s = 50 m;
matching step ds = 20 m; matching range Δ = 40 m and percentage threshold for BMC-Model is 99%. The dataset has undergone
preliminary processing using the Key Equipment Identification (KEI) model proposed in Xu et al. (2013).

7.1. Multi-channel information fusion

Fig. 12 illustrates the estimated CPO of the original and processed inspection data. It can be found that the CPO of the original
channel offset is within the range of [−1.25, 1.25] meters by reference to the gauge channel, as illustrated in Fig. 12(a). There are
differences between different runs and channels. As for the processed data, the CPO drops to less than 0.025 m, which can never-
theless be disregarded, as found in Fig. 12(b).

7.2. Robustness against abnormal data

This section presents a comparison between the waveform before and after position synchronization. The inspection data from
two channels, Channel #1 for track gauge and Channel #2 for crosslevel, within the range of K641.6-K641.8 are taken as an example,
please refer to Fig. 13.
It can be seen that Channel #1 (gauge) has some abnormal data points, which randomly occurred in different inspection runs, as
highlighted in Fig. 14(a), (b). And there are data exceptions in the measured gauge of the inspection run colored red, see Fig. 13(a),
(b). By contrast, Channel #2 (crosslevel) does not have any abnormal data points within the same range of positions. As a result, the
UMPs in Channel #1 are substituted by the RMPs in Channel #2 through multi-channel fusion. That explains the robustness of the
methods in this paper against abnormal data for the synchronized gauge.

558
Y. Wang et al. Transportation Research Part C 93 (2018) 544–565

Fig. 12. The estimation of CPO; (a) shows the CPO of original data; (b) illustrates the CPO of the processed data.

Fig. 13. Comparison of the waveform before and after position synchronization (Channel #1 gauge). (a and b) illustrate the waveforms of the
original and processed gauge, respectively; (c and d) illustrate the waveforms of the original and processed crosslevel, respectively.

7.3. Precision and number of iterations

The purpose of this section is to analyze the influence of iterations on the precision of position synchronization. Fig. 14 illustrates
the joint distribution of local milepost offset and waveform similarity after different iterations. It can be observed that the standard
deviation of local milepost offset drops quickly to 0.73 (0.19 m) after the first iteration, then 0.26 (0.25 m) and 0.22 (0.25 m) after the
second and third iteration. The distribution of waveform similarity, however, changes slightly with different iterations. The quan-
titative relationship among the iterations, distributions and thresholds are summarized in Table 4. It can be observed that after the
first iteration, the threshold of δ is reduced from 6.58 m to 0.42 m with 99% confidence, or 0.3 m for 95% confidence. And after the

559
Y. Wang et al. Transportation Research Part C 93 (2018) 544–565

Fig. 14. The joint distribution of value pairs (δ , ρ) . (a–d) show the distributions of the original data, and the data after one, two and three iterations
respectively.

Table 4
Iterations, distribution and the precision Pr (p) .
Iterations Distribution of the estimated local offset (0.25 m) Precision Pr (p) (m)

p = 99% p = 95%

Original N (0.020, 11.32) 6.58 4.65

After 1 iteration N (−0.001, 0.732) 0.42 0.30
After 2 iterations N (0.003, 0.262) 0.15 0.10
After 3 iterations N (0.002, 0.222) 0.13 0.09

second iteration, the threshold of δ drops to 0.15 m and 0.1 m for 99% and 95% confidences, respectively. Here, the value of 0.15 m is
less than the sampling distance of the inspection car, which is 0.25 m in China. However, the third iteration does not seem to show
any significant improvement, so it is suggested that two iterations are necessary for position synchronization.

7.4. Computational time

This section analyzes the execution efficiency of OI- and IL-Algorithms. The algorithms are implemented with the MATLAB
R2016b platform, in a CPU/I7-5820 k (6CPUs × 3.3 GHz) and RAM/16 GB hardware environment. The percentage threshold for the
BMC-Model is 95% and the iteration was performed twice. The polynomial fitting of the execution time for both algorithms is
performed, quadratic fitting for OI-Algorithm and linear fitting for IL-Algorithm. The results are presented in Fig. 15. It can be
observed that the execution time for the OI-Algorithm (developed in Section 6.1 is 0.03n2 + 0.05n−0.03 sec/km, where n represents
the number of inspection times. In contrast, the execution time for the IL-Algorithm is 0.1n + 0.43 sec/km.

Fig. 15. The execution time per kilometer for OI-Algorithm (a) and IL-Algorithm (b).

560
Y. Wang et al. Transportation Research Part C 93 (2018) 544–565

Fig. 16. The execution time and precision Pr (p = 99.7%) of position synchronization considering different matching step ds.

8. Discussion

The purpose of this section is to discuss the computation efficiency and precision of position synchronization. According to Eq.
(14), the interpolation remainder is closely related to the distance between two adjacent interpolation points, which is determined by
the matching step (ds), a parameter introduced in Section 5.2.1. The smaller the matching step ds is, the higher the interpolation
precision that can be achieved, but the more execution time the IL-Algorithm will require. As a result, execution time and precision
are two conflicting objectives. There are two essential parameters, the matching step ds and the number of total runs of inspection
data cached in memory (the parameter M in the IL-Algorithm), which control the execution time and precision of position syn-
chronization.
To optimize the execution time and precision, the IL-Algorithm is executed with different values of ds, from 5 m to 200 m. The
results are illustrated in Fig. 16. There is an inversely proportional relationship between the execution time and matching step ds.
When the matching step ds is less than 40 m, the execution time is seen to increase dramatically. The precision Pr (confidential level
p = 99.7%) of position synchronization is illustrated in Fig. 16. Fig. 16 shows that the precision of two iterations is much better than
that of only one iteration, with an improvement of approximately 0.45 m on average. As the matching step becomes larger, the
precision gradually drops. However, when the matching step is less than 40 m, the precision shows no improvement with smaller
values. For a given requirement of precision according to the sampling distance of inspection cars, namely 0.25 m, the minimal
matching step should be 50 m.

9. Conclusion

Position synchronization of track inspection data is crucial for track degradation modeling and maintenance scheduling. This
paper develops a novel data-driven methodological framework for addressing the accurate and computationally efficient position
synchronization of track infrastructure inspection data via big data fusion and an incremental learning algorithm. A number of data
analysis models are established to mine, fuse and synchronize the inspection data of multiple runs with multiple sensors. The model
also accounts for possible measurement data exceptions. An incremental learning algorithm (IL-Algorithm) is developed to facilitate
the fast computation of enormously large inspection data volumes.
The proposed method has been applied to a section of track on the China High Speed Railway network. The results show that (1)
our proposed method is robust against measurement data exceptions via multi-sensor data fusion; (2) the proposed algorithm can
sufficiently synchronize positions within a small number of iterations (in some cases, only two iterations are sufficient); (3) because of
these methodological merits, our approach can reduce the relative position error (RPE) within 0.15 meters at a 99 percent confidence
level. This is a significant improvement in positioning accuracy, considering the fact that the RPE error is much smaller than the
sampling interval (0.25 meters in our dataset); (4) for processing one additional kilometer of track, the proposed algorithm would
take an extra 0.1n + 0.43 s, where n is the number of inspection runs.
Finally, we discuss the relationship between the computational efficiency and the accuracy of position synchronization, and the
proper number of inspection runs that should be cached in memory in the IL-algorithm. In summary, the proposed position syn-
chronization methodology considers both positioning accuracy and computational time, with a promising application to data-rich
railroad inspection information processing problems, in support of a wide array of track safety and maintenance research activities.

Acknowledgements

The present work has been supported by the National Natural Science Foundation of China (51425804, 51,608,459 and
51378439) and the United Fund for Key Projects of China’s High-Speed Railway (U1334203 and U1234201). Yuan Wang is funded by
China Scholarship Council CSC (No. 201707000036). Dr. Xiang Liu is funded by the US Federal Railroad Administration at the time of
writing this paper. However, the authors are solely responsible for all views and analyses in this research.

561
Y. Wang et al. Transportation Research Part C 93 (2018) 544–565

Appendix A. The structure of OIM

δi(,kj,)c and ρi(, kj,)c represent the local offset and similarity of the jth run to the ith run at location k based on the data of the cth channel,
as shown in Eq. (28). Xi, c represents the data of the cth channel of Xi .

δi(,kj,)c = δ (s,(k − 1)·N / ds) (Xi, c , Xj, c ) (28a)

ρi(, kj,)c = ρ(s,(k − 1)·N / ds) (Xi, c , Xj, c ) (28b)

Slicing up the matrix U at a given location k and channel c, we get the local offset matrix Δ(ck ) (X ) and local similarity matrix
ϒ (ck ) (X ):
(k ) (k ) (k ) (k )
⎡ δ1,1, c ⋯ δ1, n, c ⎤ ⎡ ρ1,1, c ⋯ ρ1, n, c ⎤
Δc(k ) (X ) = ⎢ ⎥
⋮ ⋱ ⋮ ; ϒ c (X ) = (k ) ⎢ ⋮ ⋱ ⋮ ⎥
⎢ (k ) ⎥ ⎢ ⎥
(k ) ( k)
⎢ ρn,1, (k ) ⎥
⎢ δ
⎣ n,1, c
⋯ δn, n, c ⎥
⎦ ⎣ c ⋯ ρn, n, c ⎦ (29)
Then the overall offset matrix Δc (X ) and the overall similarity matrix ϒc (X ) is denoted by Eq. (30). The Overall Information
Matrix U can be rewritten as Eq. (31).

Δc (X ) = [ Δ(1) (2) (N / ds )
c (X ), Δc (X ), ⋯ Δc (X ) ] (30a)

ϒc (X ) = [ ϒ (1) (2) (N / ds )
c (X ), ϒ c (X ) ⋯ ϒ c (X ) ] (30b)

U = {(Δc (X ), ϒc (X ))|c = 1, 2, …, t } (31)

Appendix B. Solution of RME-Model using the Augmented Lagrangian method

Eq. (8) is a least squares problem with a constraint, which can be solved through the Augmented Lagrangian method. Introduce
variable λ , and Eq. (8) can be transferred into Eq. (32).
n n n
min = ∑∑ (δij(k ) ∗−di(k ) )2 + λ ∑ di(k )
i=1 j=1 i=1 (32)
Through conducting partial derivative operations for each variable and solving the resultant equations, we can obtain the best
estimation of di(k ) , as presented in Eq. (33).
n n n
1 1
di(k ) ∗ = ∑ δij(k ) ∗− ∑∑ δij(k ) ∗ (i = 1, 2, …, n)
n j=1
n2 i=1 j=1 (33)
(k )
The di(k ) ∗ can be divided into two components, di′ and d (k ) ∗.
(k )
di(k ) ∗ = di′ −d (k ) ∗ (34)
where,
n n n
1 1
di′ (k ) = ∑ δij(k ) ∗;d (k ) ∗ = ∑∑ δij(k ) ∗
n j=1
n2 i=1 j=1 (35)
(k )
The first component di′ represents the average of the local offsets between the data from other inspection runs and the ith run at
location k. The second component d (k ) ∗ represents the average of all the local offsets between inspection data from every two runs at
location k, namely the average of Δ(ck∗) (X ) . The magnitude of d (k ) ∗ reflects the degree of waveform irregularity resulting from rar-
efaction and compression, in which the smaller value is better. When Δ(ck∗) (X ) is an antisymmetric matrix, satisfying Eq. (36), we get
d (k ) ∗ = 0 and di(k ) ∗ = di′ (k ) .

δij(k ) ∗ = −δ ji(k ) ∗ (36)

562
Y. Wang et al. Transportation Research Part C 93 (2018) 544–565

Appendix C. The OI-Algorithm Flow Chart

See Fig. 17.

Fig. 17. The OI-Algorithm flow chart.

563
Y. Wang et al. Transportation Research Part C 93 (2018) 544–565

Appendix D. The IL-Algorithm Flow Chart

See Fig. 18.

Fig. 18. The IL-Algorithm Flow Chart.

564
Y. Wang et al. Transportation Research Part C 93 (2018) 544–565

References

Alfelor, R.M., Carr, G.A., Fateh, M., 2001. Track degradation assessment using gauge restraint measurements. Transp. Res. Rec. 1, 68–77.
Allotta, B., Colla, V., Malvezzi, M., 2002. Train position and speed estimation using wheel velocity measurements. Proc. Inst. Mech. Eng. Part F J. Rail Rapid Transit
216 (3), 207–225.
Bartram, D., Burrow, M., Yao, X., 2008. A Computational Intelligence Approach to Railway Track Intervention Planning. Evolutionary Computation in Practice.
Springer, Berlin Heidelberg.
Bocciolone, M., Caprioli, A., Cigada, A., Collina, A., 2007. A measurement system for quick rail inspection and effective track maintenance strategy. Mech. Syst. Sig.
Process. 21 (3), 1242–1254.
Esveld, C., 2001. Modern Railway Track. Delft University of Technology Publishing Service, The Netherlands, pp. 349–406.
Haigermoser, A., Luber, B., Rauh, J., Gräfe, G., 2015. Road and track irregularities: measurement, assessment and simulation. Veh. Syst. Dyn. 53 (7), 878–957.
Hanreich, D.W., Mittermayr, P., Presle, G., 2002. Track geometry measurement database and calculation of equivalent conicities of the OBB network. In: Proceedings
of the AREMA 2002 Annual Conference, Washington, DC.
Higgins, C., Liu, X., 2017. Modeling of track geometry degradation and decisions on safety and maintenance: a literature review and possible future research
directions. Proc. Inst. Mech. Eng. Part F J. Rail Rapid Transit 095440971772187.
Kawaguchi, A., Miwa, M., Terada, K., 2005. Actual data analysis of alignment irregularity growth and its prediction model. Quart. Report of Rtri 46 (4), 262–268.
Li, H., Xu, Y., 2010. A method to correct the mileage error in railway track geometry data and its usage. In: 2010 International Conference on Traffic and
Transportation Studies. pp. 1130–1135.
Liu, R., Xu, P., Wang, F., 2010. Research on a short-range prediction model for track irregularity over small track lengths. J. Transp. Eng. 136 (12), 1085–1091.
Liu, B., Bruni, S., 2015. Analysis of wheel-roller contact and comparison with the wheel-rail case. Urban Rail Transit 1 (4), 215–226.
Liu, X., Saat, M.R., Qin, X., Barkan, C.P., 2013. Analysis of U.S. freight-train derailment severity using zero-truncated negative binomial regression and quantile
regression. Accid. Anal. Prev. 59 (59C), 87–93.
Pedanekar, N.R., 2006. Methods for aligning measured data taken from specific rail track sections of a railroad with the correct geographic location of the sections, US,
US7130753[P].
Qu, J., 2012. Study on Track Irregularity Prediction and Decision-aiding Technology based on TQI of Raised Speed Lines. Ph.D. thesis. Beijing Jiaotong University,
Beijing, China.
Yang, A., 2009. Automatic correction milepost system of geometry inspection car based on RFID. Railway Comput. Appl. 18 (10), 39–41 (In Chinese).
Quiroga, L.M., Schnieder, E., 2012. Monte Carlo simulation of railway track geometry deterioration and restoration. Proc. Inst. Mech. Eng. Part O J. Risk Reliab. 226
(3), 274–282.
Ren, S., Gu, S., Xu, G., Gao, Z., Feng, Q., 2010. Development of GJ-4G track inspection car. Proc. Spie 7544 75440Q–75440Q-6.
Sadeghi, J., 2010. Development of railway track geometry indexes based on statistical distribution of geometry data. J. Transp. Eng. 136 (8), 693–700.
Sadeghi, J., Askarinejad, H., 2011. Development of track condition assessment model based on visual inspection. Struct. Infrastruct. Eng. 7 (12), 895–905.
Selig, E.T., Cardillo, G.M., Stephens, E., Smith, A., 2008. Analyzing and forecasting railway data using linear data analysis. Comput. Railways XI.
Soleimani, H., Moavenian, M., 2017. Tribological aspects of wheel-rail contact: a review of wear mechanisms and effective factors on rolling contact fatigue. Urban
Rail Transit 3 (4), 227–237.
Specht, C., Koc, W., Chrostowski, P., et al., 2017. The analysis of tram tracks geometric layout based on mobile satellite measurements. Urban Rail Transit 3 (4),
214–226.
Sui, G., 2009. Mileage calibration algorithm of track geometry data. J. Transport Inf. Saf.
Tsunashima, H., 2008. Condition monitoring of railway tracks using in-service vehicles: development of probe system for track condition monitoring. J. Mech. Syst.
Transport. Log. 3 (1), 154–165.
Vu, A., Ramanandan, A., Chen, A., Farrell, J.A., Barth, M., 2012. Real-time computer vision/DGPS-aided inertial navigation system for lane-level vehicle navigation.
IEEE Trans. Intell. Transp. Syst. 13 (2), 899–913.
Weston, P., Ling, C., Goodman, C.J., Li, P., Goodall, R.M., 2007. Monitoring vertical track irregularity from in-service railway vehicles. J. Rail Rapid Transit 221,
75–88.
Xu, P., 2012. Mileage Correction Model for Track Geometry Data from Track Geometry Car & Track Irregularity Prediction Model. Ph.D. thesis. Beijing Jiaotong
University, Beijing, China.
Xu, P., Liu, R., Sun, Q., Wang, F., 2012. A novel short-range prediction model for railway track irregularity. Discrete Dyn. Nat. Soc. 2012 (2012), 1951–1965.
Xu, P., Liu, R., Sun, Q., Jiang, L., 2015. Dynamic-time-warping-based measurement data alignment model for condition-based railroad track maintenance. IEEE Trans.
Intell. Transp. Syst. 16 (2), 799–812.
Xu, P., Sun, Q., Liu, R., Souleyrette, R.R., 2016. Optimal match method for milepoint postprocessing of track condition data from subway track geometry cars. J.
Transp. Eng. 142 (8), 04016028.
Xu, P., Sun, Q.X., Liu, R.K., Wang, F.T., 2011. A short-range prediction model for track quality index. Proc. Inst. Mech. Eng. Part F J. Rail Rapid Transit 225 (3),
277–285.
Xu, P., Sun, Q.X., Liu, R.K., Wang, F.T., 2013. Key equipment identification model for correcting milepost errors of track geometry data from track inspection cars.
Transport. Res. Part C Emerging Technol. 35 (9), 85–103.

565