sensors

Article
Bridge Structure Deformation Prediction Based on
GNSS Data Using Kalman-ARIMA-GARCH Model
Jingzhou Xin 1 ID
, Jianting Zhou 1, * ID
, Simon X. Yang 2, * ID
, Xiaoqing Li 1 and Yu Wang 3
1 School of Civil Engineering, Chongqing Jiaotong University, Chongqing 400074, China;
xinjz@mails.cqjtu.edu.cn (J.X.); 622160086010@mails.cqjtu.edu.cn (X.L.)
2 School of Engineering, University of Guelph, Guelph, ON N1G 2W1, Canada
3 School of Engineering, Cardiff University, Cardiff CF24 3AA, UK; WangY219@cardiff.ac.uk
* Correspondence: jtzhou@cqjtu.edu.cn (J.Z.); syang@uoguelph.ca (S.X.Y.); Tel.: +86-23-6265-2702 (J.Z.)

Received: 3 November 2017; Accepted: 15 January 2018; Published: 19 January 2018

Abstract: Bridges are an essential part of the ground transportation system. Health monitoring is
fundamentally important for the safety and service life of bridges. A large amount of structural
information is obtained from various sensors using sensing technology, and the data processing has
become a challenging issue. To improve the prediction accuracy of bridge structure deformation
based on data mining and to accurately evaluate the time-varying characteristics of bridge structure
performance evolution, this paper proposes a new method for bridge structure deformation
prediction, which integrates the Kalman filter, autoregressive integrated moving average model
(ARIMA), and generalized autoregressive conditional heteroskedasticity (GARCH). Firstly, the raw
deformation data is directly pre-processed using the Kalman filter to reduce the noise. After that,
the linear recursive ARIMA model is established to analyze and predict the structure deformation.
Finally, the nonlinear recursive GARCH model is introduced to further improve the accuracy of the
prediction. Simulation results based on measured sensor data from the Global Navigation Satellite
System (GNSS) deformation monitoring system demonstrated that: (1) the Kalman filter is capable
of denoising the bridge deformation monitoring data; (2) the prediction accuracy of the proposed
Kalman-ARIMA-GARCH model is satisfactory, where the mean absolute error increases only from
3.402 mm to 5.847 mm with the increment of the prediction step; and (3) in comparision to the
Kalman-ARIMA model, the Kalman-ARIMA-GARCH model results in superior prediction accuracy
as it includes partial nonlinear characteristics (heteroscedasticity); the mean absolute error of five-step
prediction using the proposed model is improved by 10.12%. This paper provides a new way for
structural behavior prediction based on data processing, which can lay a foundation for the early
warning of bridge health monitoring system based on sensor data using sensing technology.

Keywords: bridge engineering; deformation prediction; structural health monitoring; bridge sensor data

1. Introduction
As the key node of interoperability of traffic systems, the bridge is a vital guarantee for the
development of the economy and social safety. Bridge construction plays a crucial role in resolving
overcapacity of traditional industries, promoting the development of new strategic industries and
the third industry, boosting the integration and upgrading of industries, and stimulating economic
growth [1]. With the continuous development of sensing technology and information science, health
monitoring systems are widely used in the safety guarantee of the bridge structure. As a paramount
monitoring index that reflects the overall stiffness of the structure, deformation is the macro response
of the bridge micro complex mechanical mechanism, which contains the internal mechanical evolution
information of the structure and affect the safety of bridge structure significantly. Thus, predicting the

Sensors 2018, 18, 298; doi:10.3390/s18010298 www.mdpi.com/journal/sensors

Sensors 2018, 18, 298 2 of 17

structural deformation is of great scientific significance and engineering application value in order to
give full play to the early warning benefit of the health monitoring system, as well as the structural
safety of the bridges [2].
For the prediction of structural deformation, many studies have been conducted, which can
be broadly classified into two categories. One is the prediction method based on the mechanical
mechanism, which emphasizes the deterministic function relationship between cause and effect, and is
a predictive model with a priori characteristic. For instance, Liu et al. [3] established a shrinkage and
creep model of high crack resistance and compact concrete. The method for forecasting a long time
structural behavior of Hong Kong-Zhuhai-Macao Bridge was obtained by adding and accumulating the
increment of strain in every minute period for 30 years. Sajedi et al. [4] proposed an analytical procedure
that can predict the flexural behavior of intact and corroded reinforced concrete beams, with or without
lap splices considering the bond stress-slip behavior at the steel–concrete interface. Han et al. [5]
built a new computational framework for Bayesian inference regarding the long-term deflection of
concrete structures, importance sampling technology, and response surface approach, and a stochastic
process was introduced to improve computational efficiency and describe the random properties of
creep. As the solution of nonlinear differential equations of complex structures is arduous to realize,
the above-mentioned prediction methods often rely on numerical simulation models. However,
the prediction results of this model need to be further verified due to the complex and changeable
bridge service environment, the immature constitutive theory of concrete, and the high randomness and
uncertainty of material parameters that appear with an increase of service life. Besides, the performance
evolution mechanism of different structural forms is different, which leads to the universality of this
method. The second type is the method based on data mining of monitoring information. The method
can make full use of the macro response information of the actual structure, and avoids the complex
internal mechanism of structural deformation. It is an effective approach to dynamic control of
structural information; such as linear models, for example, the autoregressive moving average model
(ARMA) [6,7] and autoregressive integral moving average model (ARIMA) [8,9]; and nonlinear models,
such as the artificial neural network method (ANN) [10–12], extreme learning machine (ELM) [13,14],
support vector machines (SVM) [15], and ant colony optimization algorithm (ACO) [16]. The research
shows that the monitoring data of large structures has nonlinear and non-stationary characteristics
because of the ambiguous service environment [10]. The linear models are only capable of stationary
linear or simple non-stationary linear time series prediction, which are hard to express in terms of
deformation time series with high nonlinear and non-stationary characteristics, and linear versions
can only get limited prediction accuracy [17]. ANN has some limitations, such as easy oscillations and
slow convergence speed [18]. It is arduous to determine the key parameters and avoid the subjectivity
caused by artificial selection of ELM, SVM and ACO [19–21].
In recent years, the structural time series model (STM), fusing state space model and Kalman filter,
gets a growing concern because of its excellent prediction capability. The structural time series model
was first proposed by Harvey, a British econometrics economist, in 1983 [22], and the principle of
STM was formally presented and elaborated in his monograph in 1989 [23]. The structural time series
model decomposes the original time series into a variety of random components, such as trend, cycle,
seasonal, irregular and so on. The structural time series model has a natural form of state space. It does
not only express the unobservable components by using state vectors, but also estimates, smoothes
and predicts every component of the state vector based on the Kalman filter, which is widely applied
in economic and non-economic fields [24–26]. Aamir et al. proposed an ARIMA-Kalman model for
forecasting Pakistan’s monthly crude oil price [27]; Sebasti’an et al. presents for the first time a state
space representation for generalized autoregressive conditional heteroscedasticity model (GARCH)
family of time series models and proposes a novel estimation procedure based on the extended Kalman
filter [28]. In the realization aspect, Ian et al. provide a summary of a selection of the high-quality
published computational time series research using R [29]. The above research promotes the theoretical
development and practical application of the STM model. In the field of bridge health monitoring,
Sensors 2018, 18, 298 3 of 17

Omenzetter et al. formulate a vector seasonal autoregressive integrated moving average model for
the recorded strain signals [30]. The coefficients of the ARIMA model are allowed to vary with time
and Sensors
are identified
2018, 18, 298
using an adaptive Kalman filter, and the time series model here is mainly used
3 of 18
for structural damage identification as well as other research [31–33]. The method of combining time
series withmodel
series Kalman herefilter is notused
is mainly yet applied for deformation
for structural prediction.
damage identification as well as other research [31–
33].
DueThe method
to the of combining
difficulty time
of structure series with prediction
deformation Kalman filter is not
as well as yet
the applied for deformation
insufficiency of the existing
prediction.
method, a new federated deformation prediction method is established by integrating Kalman filter,
ARIMA and Due GARCH
to the difficulty of structure
model based on thedeformation prediction
time series essence as wellmonitoring
of health as the insufficiency of theSTM,
data. Unlike
existing method, a new federated deformation prediction method is established by
the time series model in this paper is a traditional Box-Jenkins time series model, and the main function integrating
Kalman filter, ARIMA and GARCH model based on the time series essence of health monitoring data.
of Kalman filter here is to reduce the noise of the initial data. The performance of this proposed method
Unlike STM, the time series model in this paper is a traditional Box-Jenkins time series model, and
is demonstrated on a long span urban rail-transit cable-stayed bridge. Compared with the prediction
the main function of Kalman filter here is to reduce the noise of the initial data. The performance of
accuracy of the ARIMA
this proposed method model, the proposedon
is demonstrated method
a longcan improve
span urban the accuracycable-stayed
rail-transit of the monitoring
bridge.data
of a Compared
complex bridge and achieve superior prediction results.
with the prediction accuracy of the ARIMA model, the proposed method can improve the
accuracy of the monitoring data of a complex bridge and achieve superior prediction results.
2. Multiple Sensors Used for Bridge Structure Deformation Prediction
2.The
Multiple
paperSensors Used for
is organized as Bridge Structure
follows. Deformation
In Section Prediction
2.1., a health monitor system, including various
sensors of a real
The bridge,
paper is stated.asInfollows.
is organized SectionIn2.2., a Global
Section 2.1., Navigation Satellite
a health monitor System
system, (GNSS)various
including kinematic
deformation
sensors ofmonitoring system
a real bridge, is further
is stated. introduced,
In Section and in
2.2., a Global Section 2.3.,
Navigation the principle
Satellite of GNSS
System (GNSS)
kinematic deformation monitoring system is further introduced, and in Section 2.3., the principle of
is provided.
GNSS is provided.
2.1. Bridge Sensors
2.1. Bridge Sensors
The Caijia Jialing River Bridge, located in Chongqing, China, is the second longest rail transit
The Caijia
cable-stayed bridgeJialing
in theRiver Bridge,
world. The located in Chongqing,
total length China,
of the bridge is is the m,
1250 second
and longest
the main railstructure
transit is
cable-stayed bridge in the world. The total length of the bridge is 1250 m, and the main structure is a
a dual-pylon dual-cable plane concrete cable-stayed bridge with rhomboic towers (Figure 1). The layout
dual-pylon dual-cable plane concrete cable-stayed bridge with rhomboic towers (Figure 1). The
of spans is 60 m + 135 m + 250 m + 135 m + 60 m = 640 m. The bridge had established a comparatively
layout of spans is 60 m + 135 m + 250 m + 135 m + 60 m = 640 m. The bridge had established a
comprehensive bridge health monitoring system to ensure the safety of serviceability, durability,
comparatively comprehensive bridge health monitoring system to ensure the safety of serviceability,
and durability,
sustainability of the bridge.ofThe
and sustainability the overall layout
bridge. The of sensors
overall is sensors
layout of shown is inshown
Figurein2.Figure 2.

Figure 1. The Caijia Jialing River Bridge.

Figure 1. The Caijia Jialing River Bridge.
Combined with the construction progress of the bridge, the health monitoring sensors were
Combined with
appropriately the at
installed construction progress
the corresponding of theasbridge,
positions follows.the health monitoring sensors were
appropriately installed
(1) Main girder. at the corresponding
Along the longitudinal positions asthe
direction of follows.
bridge, the vibrating wire strain sensors
were attached
(1) Main to the
girder. lower
Along theedge of the top direction
longitudinal slab and the upper
of the edge the
bridge, of the bottom wire
vibrating slab, strain
which sensors
are
shown in Figure 3. It is worth mentioning that the sensor could also measure the temperature
were attached to the lower edge of the top slab and the upper edge of the bottom slab, which are of the
structure
shown and 3.
in Figure modify the initial
It is worth temperature.
mentioning that Also, static could
the sensor level gauges based onthe
also measure thetemperature
pipe principleof the
were installed to monitor the static long-term deformation of the girder. Meanwhile, the GNSS was
structure and modify the initial temperature. Also, static level gauges based on the pipe principle
installed to monitor the spatial mid-span deformation. Also, the displacement sensors and the
were installed to monitor the static long-term deformation of the girder. Meanwhile, the GNSS
acceleration sensors were installed to measure the width of the expansion joint and to test the
dynamic characteristics, respectively.
Sensors 2018, 18, 298 4 of 17

was installed to monitor the spatial mid-span deformation. Also, the displacement sensors and the
acceleration sensors were installed to measure the width of the expansion joint and to test the dynamic
Sensors 2018, 18, 298 4 of 18
characteristics, respectively.
Sensors 2018, 18, 298 4 of 18

D（1）
FD（（11）
）
GF（（11）
）
HG（（11）
） J（2） D（1）
I（
H （11）
） JJ（
（22）
） DF（
（11）
）
I（1） JJ（
（22）
） F（1）
JJ（
（22）
）
J（2）
Jinshan Temple Caojiawan
Jinshan Temple Caojiawan

A（6） A（6） A（6） A（6）A（6） A（6） A（6） A（6） A（6） A（6） A（6）
A（
C（ 6）
2） BA
（（ 6） A（6） BA（
1） （16）
）BA（
（16）
） A（
B 6）
（1 ） A
B（（61）
） A（6） A（6） AB（（61）） AC
（（6）
2）
C（2） CB
（（ 1）
2） B（1）CB（（21）
） B（
C （12）
） B
C（（12）
） BC（（1）2） C（2）
C（2） C（2） C（
D （12）
） C（2） C（2）
E（2） D（1） E（2）
E（2） E（2）

Figure
Figure2. The
2.2.The overall
Theoverall layout
overalllayout
layoutofof the
ofthe sensor
thesensor
sensor measuring
measuring point: A, thestress
the
A,the
point: A, stresssensor
stress sensor(including
(includingtemperature
temperature
measurement as well); B, the acceleration sensor; C, the static level gauge; D, Global Navigation
measurement as well); B, the acceleration sensor; C, the static level gauge; D, Global Navigation Satellite Satellite
System (GNSS);
(GNSS); E, Linear Variable Displacement
Displacement Transducer; F, the tiltmeter; G, the
System (GNSS); E, Linear Variable Displacement Transducer; F, the tiltmeter; G, the temperature and temperature and
humidity
humidity sensor;
sensor; H,
H, the
the pluviometer;
pluviometer; I,
I, the
the dogvane
dogvane and
and anemoscope
anemoscope and
and J,
J, the
humidity sensor; H, the pluviometer; I, the dogvane and anemoscope and J, the anchor cable meter anchor cable meter
through the
throughthe cable.
thecable. Also,
cable.Also,
Also,thethe numbersin
thenumbers
numbers inparentheses
in parenthesesindicate
parentheses indicatethe
indicate the
the amount
amount
amount ofof
of each
each
each sensor.
sensor.
sensor.

Stain Sensor
Stain Sensor

11 2 33

44 5 66

Figure3.
Figure
Figure 3.3.The
Thelayout
The layoutdiagram
layout diagram of
diagram of strain
strain sensors
sensors on
on the
the surface
surfaceof
ofconcrete.
concrete.

(2)Stay-cable.
(2) Stay-cable.Intelligent
Intelligent anchor
anchor cable
cable meters
meters with with temperature
temperature measurement
measurementincluded includedwere
were
(2) Stay-cable.
embeddedtotomonitor Intelligent
monitorthe thecable anchor
cableforce cable
forcebefore
before the meters
the cable with
cable tension temperature measurement included were
embedded tension inin the
theconstruction
constructionstage stageand
andanchorage
anchorage
embedded to and
monitor theThecable force before the cable tension in the construction stage and anchorage
ofofthe
thecable
cable tower.
and tower. intelligent
The intelligent six-string
six-string meters
meters were
were applied
applied for formeasuring
measuringthe thecable
cableforce
force
ofunder
the cable and
theeccentrictower.
eccentricload The
loadto intelligent
toensure
ensurethe six-string
the accuracy.
accuracy. meters were applied for measuring the cable force
under the
under (3) theMain
eccentric
tower. loadTheto ensure the
tiltmeters andaccuracy.
GNSS were were installed
installed on
(3) Main tower. The tiltmeters and GNSS on the
thetop
topofofthe
thetower,
tower,and andaamonitoring
monitoring
(3) Main tower. The tiltmeters and GNSS were installed onincluding
the top ofwindthe tower, and a monitoring
station was established for monitoring meteorological factors including wind speed,wind
station was established for monitoring meteorological factors speed, winddirection
direction
station was established
and rainfall, and humidity. for monitoring meteorological factors including wind speed, wind direction
and rainfall, and humidity.
and rainfall, and humidity.
2.2. Global Navigation Satellite System Kinematic Deformation Monitoring System
2.2.
2.2. Global
Global Navigation
Navigation Satellite
Satellite System
System Kinematic
Kinematic Deformation
Deformation Monitoring System
Monitoring System
To measure the deformation of the pivotal components of the bridge, the GNSS kinematic
To measure the
To measure the deformation
deformation of of the pivotal
pivotal components
components of of the bridge,
bridge, the the GNSS
GNSS kinematic
kinematic
deformation monitoring system wasthe applied, which consists of the a data acquisition system, data
deformation
deformation monitoring system was applied, which consists of a data acquisition system, data
transmission monitoring
system, lightning system was applied,
protection which and
system, control consists
analysis ofsystem,
a dataand acquisition system,
power system. A
transmission
data system, lightning protection system, control and analysis system, and power system. A
totaltransmission
of four stations system,
were set lightning protection
up, including system, station
one reference controlestablished
and analysis in thesystem, and power
non-deformation
total
system.of four stationsfour
A total werestations
set up, including set one reference station established in theestablished
non-deformation
area, and three ofmonitoring were
stations located up, including
at the top of theonetwo reference station
main towers in the
and the mid-span,
area, and
non-deformationthree monitoring
area, and three
respectively. Considering
stations located
monitoring
the impact of train
at the
stations top
operation
of
located the
on at
two
the the
GPS,
main
top mid-spantowers and
of the twomeasurement the mid-span,
main towers and the
point
respectively.
mid-span, Considering
respectively. the
Considering impact theof train
impact operation
of train on GPS,
operation
was placed on the outside of the lateral deck. Dual frequency double star monitoring receiver onthe mid-span
GPS, the measurement
mid-span point
measurement
was
point placed
was placed
GMX902GG on(Leica
the
on outside
the outside
Geosystems of theof
AG, lateral
the deck.
St.lateral
Gallen,deck. Dual
Dualfrequency
Switzerland) frequency double star
doubleAS10
and antenna star monitoring
monitoring receiver
receiver
(Leica Geosystems
GMX902GG
GMX902GG
AG) are applied (Leica Geosystems
(LeicainGeosystems AG, St. Gallen,
AG, St. Gallen,
the four measuring Switzerland)
Switzerland)
stations. The data from and
and the antenna
antenna AS10
stations AS10
are(Leica(Leica Geosystems
Geosystems
transmitted AG)
by fiber
AG) aretransmission
optics applied in the four measuring
independently. Thestations.
control and Theanalysis
data from the stations
system realizes arethe transmitted
automation ofbydata
fiber
optics transmission
observation, independently.
data decoding The control
and processing, andand analysis system
equipment management realizesbythe GNSSautomation of data
Spider (Leica
observation, data decoding and processing, and equipment management
Geosystems AG). The working network of GNSS deformation monitoring system is shown in Figure by GNSS Spider (Leica
4.
Geosystems AG). The working network of GNSS deformation monitoring system is shown in Figure 4.
Sensors 2018, 18, 298 5 of 17

are applied in the four measuring stations. The data from the stations are transmitted by fiber
optics transmission independently. The control and analysis system realizes the automation of
data observation, data decoding and processing, and equipment management by GNSS Spider
(Leica Geosystems AG). The working network of GNSS deformation monitoring system is shown
in Figure 4.
Sensors 2018, 18, 298 5 of 18

Figure
Figure 4.
4. GNSS
GNSS deformation
deformation monitoring
monitoring system.
system.

2.3. Principle of Global Navigation Satellite System Kinematic Deformation Monitoring

2.3. Principle of Global Navigation Satellite System Kinematic Deformation Monitoring
Global Navigation Satellite System refers to all satellite navigation systems, including American
Global Navigation Satellite System refers to all satellite navigation systems, including American
GPS, Russian Glonass, European Galileo, and China’s Beidou navigation systems and related
GPS, Russian Glonass, European Galileo, and China’s Beidou navigation systems and related
enhancement systems. There are many shared characteristics in the independent satellite system, for
enhancement systems. There are many shared characteristics in the independent satellite system,
instance, each system consists of a ground control component, space satellites and a user component.
for instance, each system consists of a ground control component, space satellites and a user component.
The satellite signal contains three levels, namely, carrier, pseudo code and navigation message data
The satellite signal contains three levels, namely, carrier, pseudo code and navigation message data
code [34]. The main functions of ground control component are tracking GNSS satellites, determining
code [34]. The main functions of ground control component are tracking GNSS satellites, determining
the orbits of satellites and satellite clock parameters, forecasting and compiling navigational
the orbits of satellites and satellite clock parameters, forecasting and compiling navigational messages.
messages. The main function of the space satellite is to continually broadcast the ranging signals and
The main function of the space satellite is to continually broadcast the ranging signals and navigation
navigation messages for navigation and positioning so that the earth can receive a sufficient number
messages for navigation and positioning so that the earth can receive a sufficient number of satellite
of satellite signals at any time. The navigation message includes satellite clock and satellite orbit, etc.
signals at any time. The navigation message includes satellite clock and satellite orbit, etc.
One of the key problems of GNSS positioning is how to determine the distance from the satellite
One of the key problems of GNSS positioning is how to determine the distance from the
to the ground sensor. The methods including the code pseudo measurement and the carrier phase
satellite to the ground sensor. The methods including the code pseudo measurement and the carrier
measurement are typically adopted, and carrier phase measurement is the one to achieve higher
phase measurement are typically adopted, and carrier phase measurement is the one to achieve
precision positioning. The GNSS signal carrier is a cosine wave with a short wavelength. The
higher precision positioning. The GNSS signal carrier is a cosine wave with a short wavelength.
wavelength of two frequencies is 19.0 cm or 24.4 cm [34]. The measurement error at the level of mm
The wavelength of two frequencies is 19.0 cm or 24.4 cm [34]. The measurement error at the level of
is completely acceptable in the monitoring of large bridge structure [35]. The observation equation of
mm is completely acceptable in the monitoring of large bridge structure [35]. The observation equation
carrier phase measurement can be constructed by
of carrier phase measurement can be constructed by
q  = ( X i  X ) 2  (Y i  Y ) 2  ( Z i  Z ) 2 -cV  cV  N   ( V )  ( V ) , (1)
i iR ii s i ion i trop i
2 2 2
ϕi λ = ( X i − X ) + (Y i − Y ) + ( Zi − Z ) − cViR + cVii s − Ni λ − (Vion )i − (Vtrop )i , (1)
where i  1, 2, 3 , and i represents the satellite number;  is the carrier wavelength; i
where i = 1, 2, 3 . . ., and i represents the satellite number; λ is ithei carrier wavelength; ϕ represents
represents phase difference which less than an entire i icycle;
i (X , Y , Zi ) is the location of ithe satellite
phase difference which less than an entire cycle; ( X , Y , Z ) is the location of the satellite in space based
in
onspace based on
the satellite the satellite
ephemeris; ( X,ephemeris; ( X , Y , Z) ofisthe
Y, Z ) is the location theGNSS
location of thecGNSS
sensor; sensor;ofclight;
is the speed is theVspeed
i R and
of light; ViR and Vii s are the receiver clock error and satellite clock error, respectively; Vion and
Vtrop are ionospheric delay and tropospheric, respectively; and Ni is the ambiguity of whole cycles.

3. Prediction Model
In this section, a brief review of the Kalman filter, time series model, and GARCH will be
Sensors 2018, 18, 298 6 of 17

Vii s are2018,
Sensors
the18,
receiver
298
clock error and satellite clock error, respectively; Vion and Vtrop are ionospheric
6 of 18
delay and tropospheric, respectively; and Ni is the ambiguity of whole cycles.
Due to the complexity of the bridge structure monitoring environment, there may be a
3. Prediction Model
significant amount of random noise in the observation data—the deformation curve shows the
In this section,
characteristic a brief
of a small review
value with aoflarge
the fluctuation.
Kalman filter, timestudy,
In this seriesthe
model, and
optimal GARCH Kalman
estimation will be
filter algorithm
presented, in the least mean square sense is used to de-noise the random noise of the data. In
respectively.
most cases, the discrete Kalman filter is the major application types [36]. The mathematical model is
3.1. Kalman Filter
as follows:
Due to the complexity of the bridge structure monitoring environment, there may be a significant
xk = Fk|k-1xk-1  wk , xk , xk-1  R , (2)
amount of random noise in the observation data—the deformation curve shows the characteristic of
a small value with a large fluctuation. In this study, the optimal estimation Kalman filter algorithm in
yk = Hthe
the least mean square sense is used to de-noise k krandom
R , of the data. In most cases, the discrete
x  vk , yk noise (3)
Kalman filter is the major application types [36]. The mathematical model is as follows:
where R and R represent  - and  -dimensional real variable domains, respectively; xk and
Fk|k−k1 x1
xk -1 are the state vectors at stepsxkk=and + wk , xk , xk−1 y∈k Risη ,the observed measurement value
k −1, respectively; (2)

at step k ; Fk|k-1 is the transition matrix; yk =and

H , ythe∈measurement
Hk xk +k vis Rv , matrix. Both transition and (3)
k k
measurement matrices can be variable matrices or constant matrices. Vectors w k and v represent
where Rη and Rv represent η- and v-dimensional real variable domains, respectively; xk k and xk−1
the the
are process and measurement
state vectors at steps k and noises,
k − 1,respectively.
respectively; They yk is theareobserved
assumedmeasurementto be additive, valuewhite, and
at step k;
p(w)
Fk|k−1 is the transition matrix; and Hk is the measurement matrix. Both transition and measurement.
independent of each other, and have the probability distribution ~ N (0, Q) and p(v) ~ N (0, R)

Let x k and x k represent the a priori and a posteriori state estimates
matrices can be variable matrices or constant matrices. Vectors wk and vk represent the process and
at step k , respectively, where
measurement noises, respectively. They are assumed to be additive, white, and independent of each
the error covariance matrices are calculated by
other, and have the probability distribution p(w)∼ N (0, Q) and p(v)∼ N (0, R). Let x̂k− and x̂k represent

the a priori and a posteriori state estimates Pk atEstep   T where the error covariance matrices
[( xk k,xrespectively,
k )( xk  x k ) ] , (4)
are calculated by
T
Pk− = E[( xk − x̂k− )( xk − x̂ k− )T ], (4)
Pk  E[( xk  x k )( xk  x k ) ] . (5)
T
Pk = E[( xk − x̂k )( xk − x̂k ) ]. (5)
The Kalman filter uses a predictor-corrector algorithm to estimate xk as shown in Figure 5 [36].
The Kalman filter uses a predictor-corrector algorithm to estimate xk as shown in Figure 5 [36].
Firstly, aa tentative
tentativeestimate x kis is
estimatex̂ − calculated based on the value of −x  k 1 , then the measurement value
Firstly, k calculated based on the value of x̂ k −1 , then the measurement value yk
− 
isy used
k
to further
is used refine
to further the the
refine valuevalueof x̂of
k in
x korder to obtain
in order to obtain x k , which
x̂k , which is theisestimate of xkof x
the estimate k

Predictor
xk =F x k 1
k|pre
k-1

Pk =Fk|k -1 Pk -1 Fk|k -1T  Q

Corrector
Pk H k T
K
( H k Pk H k T  R)
 
x k = x k  K( yk  H k x k )
Pk =（1-KHk）Pk

Figure 5. Predictor-corrector algorithm of Kalman filter.

Figure 5. Predictor-corrector algorithm of Kalman filter.

3.2. Test for Stationary of Time Series

In this paper, the runs test using to check the stationarity of time series is conducted, then the
corresponding model is established according to the test results. The specific steps of the run test are
shown below:
Step 1: Calculate the mean value X of sample series X (t ) .
Sensors 2018, 18, 298 7 of 17

3.2. Test for Stationary of Time Series

In this paper, the runs test using to check the stationarity of time series is conducted, then the
corresponding model is established according to the test results. The specific steps of the run test are
shown below:
Step 1: Calculate the mean value X of sample series X (t).
Step 2: Code values above X as positive and values below X as negative. Thus, a sequence
of symbols corresponding to the original series can be obtained. Each consecutive sequence of the
symbolic sequence is defined as a run, and the total number of runs is r.
Step 3: For random sequences, assume the length of the sequence is N, and N1 , N2 reflect the
number of positive and negative occurrences in the sequence, respectively.
Step 4: The stationarity of a random sequence X (t) can be tested by the following equations:

r − E (r )
Z= , (6)
σ (r )

E(r ) = 2N1 N2 /N + 1, (7)

2N1 N2 (2N1 N2 − N ) 1/2

 
σ (r ) = . (8)
N 2 ( N − 1)
As the statistic Z obeys the normal distribution approximately, at the 5% significance level, a test
statistic with an absolute value greater than 1.96 indicates non-stationary.

3.3. Time Series Model

A time series model can be classified into two categories: a stationary model, such as autoregressive
model (AR), moving average model (MA), ARMA; or a non-stationary model, such as ARIMA.
For non-stationary deformation time series, ARIMA( p, d, q) model should be adopted;
ARIMA( p, d, q) can be constructed below,

φ( B)(1 − B)d X (t) = θ ( B) a(t), (9)

B = X (t − 1)/X (t), (10)

φ( B) = 1 − φ1 B − φ2 B2 · · · − φ p−1 B p−1 − φ p B p , (11)

θ ( B ) = 1 − θ 1 B − θ 2 B 2 · · · − θ q −1 B q −1 − θ q B q , (12)

where X (t) represent the measured deformation for the moment t (t = 1, 2, 3, · · · ), while a(t) is the
residual errors at the same moment, which should satisfy the Gaussian white noise process with
a mean value of 0; θ j ( j = 1, 2, 3, · · · q) is the parameters to be estimated for the model; B is the
difference operator; and p, d and q represent autoregressive order, difference order, and moving
average order of the model, respectively. In particular, when d = 0, the ARIMA( p, d, q) model is equal
to the ARIMA( p, q) model.

3.4. Explanation and Test Method of Heteroscedasticity

One of the critical hypotheses of the classical linear regression model is that the random
disturbances have equal variance, which guarantees unbiased, effective and consistent regression
coefficients. In other words, they assume that the data is homoscedastic. When the hypothesis is
not well-established, the validity and consistency of the regression estimation coefficient cannot
be guaranteed, which leads to the estimation bias of the regression coefficients and lower fitting
precision of the model. To cope with the above problems, Engle proposed the autoregressive
conditional heteroskedasticity model (ARCH) [37]. The basic idea is that the residuals obey the
Gaussian distribution whose mean is zero and variance is a time-varying variance (conditional
Sensors 2018, 18, 298 8 of 17

heteroscedasticity), and the variance can by expressed by a linear combination of the squared residuals
of the past limited term. The essence of the model is fitting the current heteroscedastic function values
by using the finite moving average term of the squared residual sequence. However, in many cases,
many heteroscedastic functions of residuals have a long-term correlation, if fitting heteroscedasticity
functions using the ARCH model will produce high moving average order, the difficulty of parameters
estimation increased. For that reason, Bollerslev proposed GARCH [38], which defines the current
heteroscedasticity function as a weighted combination of past heteroscedasticity functions and the
squared residuals of the past. The number of parameters in the model and the difficulty of parameter
estimation are reduced. Meanwhile, the processing ability of heteroscedasticity is further improved.
At present, the intuitionistic methods of testing heteroscedasticity include the residual plot test, residual
square test, and analytic method.

3.5. Generalized Autoregressive Conditional Heteroskedasticity Model

The mathematical form of the GARCH model can be expressed by

X (t) = φ1 X (t − 1) + · · · + φ p X (t − p) + a(t) + θ1 a(t − 1) + · · · + θq a(t − q), (13)

p
a(t) = ht et , (14)
m n
ht = β 0 + ∑ βu a2t−u + ∑ αv ht−v , (15)
u =1 v =1

where the mean Equation (13) can be obtained by the ARIMA model; { at } is the residual sequence;
and {et } represents an independent and identically distributed random sequence with mean 0 and
variance 1. It usually has three kinds of distributions, that is, standard normal distribution, t distribution
m n
and generalized error distribution; β u > 0(u = 0, 1, · · · m), αv > 0(v = 1, · · · n), ∑ β u + ∑ αv < 1.
u =1 v =1
The GARCH model uses the m-order moving average (ARCH) term of the squared residual sequence
 2
at and the n-order autoregressive (GARCH) term of the sequence {ht } to fit the current
heteroscedasticity value, and n and m reflect the order of the GARCH term as well as the ARCH
term, respectively. Obviously, the ARCH(m) model is a special case of the GARCH(m, n) model.
In particular, when ht is constant and {et } obeys the Gaussian distribution, the GARCH model is
equivalent to the ARMA model.

4. Results and Discussions

In this section, the establishment process of the proposed model is described in detail, and the
performance of the proposed Kalman-ARIMA-GARCH model is demonstrated by comparison with
the Kalman-ARIMA model. All the simulations are carried out in MATLAB 2011a.

4.1. Data Preprocessing

A collection of deformation data of mid-span of Caijia Jialing River Bridge measured by GNSS
(including 9000 samples) is adopted to examine the effectiveness and reliability of the proposed
method. The sampling period is 10 s. Due to the influence of random errors from light, temperature,
and the sensors, the variation characteristics of little deformation with significant fluctuation often
appear in the deformation monitoring. In this paper, the Kalman filter is used to de-noise the sample
data. According to [39], there is a positive correlation between the measurement noise variance and
the estimation performance of the filter, and a negative correlation between the system noise variance
and the filter performance, in this study, R = 0.3, Q = 1.0. The time series processing by the Kalman
filter is called as { X1t }, which is illustrated in Figure 6.
As shown in Figure 6, the Kalman filter can effectively eliminate random noise, and the fidelity of
the data after filtering is satisfactory.
Sensors 2018, 18, 298 9 of 17
Sensors
Sensors 2018,
2018, 18,
18, 298
298 99 of
of 18
18

40
40
The
The original
original data
data
20 Deflection
Deflection data
data after
after Kalman
Kalman filtering
filtering
20

00
Deformation (mm)

-20
-20

-40
-40

-60
-60

-80
-80

-100
-100
00 1000
1000 2000
2000 3000
3000 4000
4000 5000
5000 6000
6000 7000
7000 8000
8000 9000
9000
Sample
Sample number
number

Figure
Figure 6.
6. The
The deflection
deflection sample
sample series
series after
after the
the Kalman
Kalman filter.
filter.

According
According to
According to [7,8,10],
to [7,8,10], aaa single
[7,8,10], single
single prediction
prediction step
prediction step is
step is set
is to
set to
set one
to one minute.
minute. If
one minute. the
If the
If prediction
prediction is
the prediction is directly
is directly
directly
carried out
carried out
carried by
out by using
by using the
using the time
the time series {X
series {{X
time series } , the
X11tt }} ,, the model
the model needs
model needs six
needs six steps
six steps ahead
steps ahead for
ahead for each
for each forward
each forward prediction
forward prediction
prediction
1t
of
of 11 min.
min. It
It is
is obvious
obvious that
that too too many
many prediction
prediction steps
steps will
will affect
affect the
the prediction
prediction accuracy
accuracy of of time
time series
series
models.
models. In Inthis
In this study,{{X
thisstudy,
study, {XX }}
11tt1t }is
is processed
is processed
processed through
through
through the average
the
the average
average value
value
value method
method
method every minute.
every
every In
minute.
minute. In this
In
this way,
this
way,
way,
the the advanced
the advanced
advanced min1prediction
11 min min prediction
prediction modelmodel
model needsneeds
needs only only
only one
one oneCompared
step.
step. step. Compared
Compared with
with thewith
the the approach
approach
approach of
of the of the
the simple
simple
simple adjustment
adjustment
adjustment of of the of
the data the data
data sampling sampling
sampling frequency, frequency,
frequency, the the
the maximum maximum
maximum deformation deformation
deformation value value in value
in the
the 60 in the
60 ss is 60 s is
is retained
retained
retained
while
while the while
the GPS the GPS
GPS performanceperformance
performance is is guaranteed.
is guaranteed.
guaranteed. The The sequence
sequence {{X
The sequence { X } (including
X22tt}} 2t(including
(including 1500 1500 samples)
1500 samples) after
samples) afterafter
averaging is shown in Figure 7. Here, the time series is divided into two parts: the 1st–1000th sampling
averaging
averaging is is shown
shown in in Figure
Figure 7. 7. Here,
Here, the
the time
time series
series is
is divided
divided into
into twotwo parts:
parts: the
the 1st–1000th
1st–1000th sampling
sampling
points for training and the subsequent 1001th–15000th data for testing.
points
points for
for training
training andand the
the subsequent
subsequent 1001th–15000th
1001th–15000th data data for
for testing.
testing.

10
10

00

-10
-10

-20
-20
Deformation (mm)

-30
-30

-40
-40

-50
-50

-60
-60

-70
-70

-80
-80
00 500
500 1000
1000 1500
1500
Sample
Sample number
number

Figure 7. Deformation
Figure7.
Figure 7. Deformation time
timeseries
series{X }.
2t}.
{X2t
2t

4.2.
4.2. Model
Model Establishment
Establishment
In
In this
this section,
section, the
the detailed
detailed modeling
modeling process
process of
of ARIMA-GARCH
ARIMA-GARCH is
is presented.
presented.

4.2.1.
4.2.1. Test
Test for
for Stationary
Stationary of
of Time
Time Series
Series
 Sensors 2018, 18, 298 10 of 18
Sensors 2018, 18, 298 10 of 17
Sensors 2018,
The18, 298
runs test is adopted to test the stationarity of time series {X2t } . As the result shows10there
of 18 is

4.2.1. Test for Stationary

nonstationarity in {X2of t
} ,Time
ARIMA(Series
p , d , q ) is used to fit the sequence {X2t } . The 1st–1000th sampling
The runs test is adopted to test the stationarity of time series {X2t } . As the result shows there is
points
The are
runssubjected
test is to first order
adopted to differencing.
test the stationarityData ofprocessing
time seriesby{ first
X 2t } . difference
As is recorded
the 1st–1000th
result shows as {X
there is3t }
nonstationarity in {X2t } , ARIMA( p , d , q ) is used to fit the sequence {X2t } . The sampling
nonstationarity
in Figure 8. The { X2t }test
in runs result of d,{Xq)3t }is used
, ARIMA(p, showstostationarity, thus,{in
fit the sequence X2tthe}. The
ARIMA 1st–1000th
( p , d , q )sampling
model, the
points are subjected to first order differencing. Data processing by first difference is recorded as {X3t }
points are subjected to first order differencing. Data processing by first difference is recorded as
order of the difference is equal to one.
in
{ XFigure 8. The 8.
3t } in Figure runs
Thetest
runs result
test of {X3of
result t
} {shows stationarity,
X3t } shows thus, thus,
stationarity, in theinARIMAthe ARIMA( p , d ,(p,
q ) d,
model, the
q) model,
the order
order 30of difference
of the the difference is equal
is equal to one.
to one.

30
20

20
10
Deformation (mm)

10
Deformation (mm)

0
-10

-10
-20

-20
-30
0 500 1000 1500
Sample number
-30
0 500
FigureSample
8. Time series {X3t}. 1000
number
1500

4.2.2. Model Identification Figure

Figure 8.
8. Time
Time series
series {X
{X3t3t}.}.

4.2.2. Figure 9 shows analysis results of {X3t } by autocorrelation (ACF) and partial ACF (PACF). As
4.2.2. Model
Model Identification
Identification
seen from Figure 9, both ACF and PACF exhibit tails of infinite shock. The ARMA model is
Figure
Figure 99 shows
shows analysis resultsofof{ X{X3}t }byby
analysisresults autocorrelation
autocorrelation (ACF)
(ACF) andand partial ACF
partial (PACF).seen
As
established to fit {X3t } , which is equivalent
3t to establishing an ARIMA( p,1, q ) ACF
model(PACF).
fittingAs
sequence
seen
from from
FigureFigure
9, both9,ACF both ACF
and PACF and PACFtails
exhibit exhibit tails shock.
of infinite of infinite shock. model
The ARMA The ARMA model to
is established is
 X 2t  .
established to fit is{X
fit { X3t }, which } , whichto
equivalent is establishing
equivalent toan establishing
ARIMA(p, an ARIMA(
1, q) p ,1, q
model fitting)sequence
model fitting
{ X2tsequence
}.
3t

X  .
2t
1

10.5
ACF

0.5 0
ACF

0
-0.5
0 5 10 15 20
-0.5 1 Lag
0 5 10 15 20
Lag
10.5
PACF

0.5 0
PACF

0
-0.5
0 5 10 15 20
-0.5 Lag
0 5 10 15 20
Figure
Figure 9. The
9. The autocorrelation
autocorrelation (ACF) Lag
(ACF)
andand partial
partial ACFACF (PACF)
(PACF) forfor
thethe
{X3t{X } series.
} 3tseries.

4.2.3. ModelFigure
Order9.Determination
The autocorrelation (ACF) and partial ACF (PACF) for the {X3t} series.

4.2.3. Model Order Determination

Sensors 2018, 18, 298 11 of 17
Sensors 2018, 18, 298 11 of 18

The Akaike information criterion (AIC) is used to determine the value of the autoregressive
4.2.3. Model Order Determination
order p and moving average order q in ARIMA( p,1, q ). The mathematical form of the AIC
The is
criterion Akaike
given information
by criterion (AIC) is used to determine the value of the autoregressive
order p and moving average order q in ARIMA(p, 1, q). The mathematical form of the AIC criterion is
given by AIC =  2ln(c1）+ 2(c2 ), (16)
AIC = −2 ln(c1 ) + 2(c2 ), (16)
where c1 is the maximum likelihood, and c2 is the number of independent parameters.
where c1 is the maximum likelihood, and c2 is the number of independent parameters.
By comparison, the model fitting is most reasonable when p  3 , q  3 , thus, the most suitable
By comparison, the model fitting is most reasonable when p = 3, q = 3, thus, the most suitable
non-stationary ARIMA( p , d , q ) model is the ARIMA(3,1,3) model.
non-stationary ARIMA(p, d, q) model is the ARIMA(3,1,3) model.

4.2.4. Parameter Estimation

Maximum likelihood estimation is used to estimate the parameters of the ARIMA(3,1,3) model,
ARIMA(3,1,3) model
the equation of ARIMA(3,1,3) model is
is obtained
obtained by
by
X(t )  0.3620 X(t  1)+0.4762 X(t  2)  0.3935X(t  3)  0.5077 X(t  4)+a(t )  0.2128 a(t  1)
X (t) = −0.3620X (t − 1) + 0.4762X (t − 2) + 0.3935X (t − 3) − 0.5077X (t − 4) (17)
 0.4317 a(t  2)  0.6995a(t  3) (17)
+ a(t) + 0.2128a(t − 1) − 0.4317a(t − 2) − 0.6995a(t − 3)

4.2.5. Test of
of Heteroscedasticity
Heteroscedasticity
In this
this study,
study, the
theresidual
residual plot
plot test
test and
andthe
theLagrange
Lagrange multiplier
multiplier verify
verify (ARCH-LM)
(ARCH-LM) are used to
examine the heteroscedasticity. As shown in Figure 10, the aggregation
aggregation of the
the residuals
residuals indicates
indicates that
the residuals of the ARIMA(3,1,3) model may have heteroscedasticity. If there is heteroscedasticity, the
GARCH model would be established, if not, the prediction would be made by the mean Equation (17).

25

20

15

10
Residuals (mm)

-5

-10

-15

-20
0 100 200 300 400 500 600 700 800 900 1000
Sample number

Figure 10. The residual error of autoregressive integrated moving average

average (ARIMA).
(ARIMA).

4.2.6. Establishment of the GARCH Model

resultsshow
The results showthatthat
thethe residuals
residuals of the
of the ARIMA(3,1,3)
ARIMA(3,1,3) model model exhibit
exhibit heteroscedasticity,
heteroscedasticity, in
in which
which
case, thecase,
GARCH the model
GARCH model
should should be As
be established. established.
the GARCH(1,1)As the GARCH(1,1)
model model describes
describes heteroscedasticity
heteroscedasticity
concisely and has aconcisely
satisfactoryandfitting
has aperformance,
satisfactory itfitting performance,
is considered as theitbenchmark
is considered
modelas [40].
the
benchmarkthis
Therefore, model
study[40]. Therefore,
established this
the study established
GARCH(1,1) model, the andGARCH(1,1)
it is found thatmodel, and it is found
the GARCH(1,1) that
model
the GARCH(1,1)
based model has
on t distribution basedtheon t distribution
best has
fitting effect the bestthe
through fitting effect through
comparison of the the comparison
optimal of
principle.
the optimal
Use principle.
the maximum Use the estimation
likelihood maximum to likelihood
determineestimation to determine
the parameters of thethe parameters
ARCH term and of the
ARCH term
GARCH term,and
thethe GARCH
GARCH term,isthe
model GARCH
obtained bymodel is obtained by

ht  3.325  104  0.7761at21 +0.2109ht 1 , (18)

Sensors 2018, 18, 298 12 of 17

ht = 3.325 × 10−4 + 0.7761a2t−1 + 0.2109ht−1 , (18)

where the coefficient of the ARCH term and the GARCH term are greater than zero and satisfy
m n
the condition ∑ β u + ∑ αv < 1. In this way, a prediction model of ARIMA(3,1,3)-GARCH(1,1) is
u =1 v =1
established, where Equation (17) is the mean equation, and Equation (18) is the variance equation.

4.3. Recursive Time Series Model

When utilizing a recursive time series model for advanced multi-step prediction, the parameters
of the model are updated by the new predictive values after obtaining the prediction of the next
moment by iteration. Then a new model including prediction information is obtained, and then used
to predict the next step by the new model. The detailed procedure is described as follows:
Step 1: Modeling { X2t (1), X2t (2), · · · , X2t (1000)} by following Sections 4.2.1–4.2.6, and obtain
the model equation. Realize one-step prediction and get the predictive value X2t (1).

Step 2: Re-estimate the parameters of the model equation by X2t (2), · · · , X2t (1000), X2t (1)
(follow Sections 4.2.4–4.2.6), then obtain the new model equation, including the predictive information
and make a further prediction, after that, complete five-step prediction in the same way.
Step 3: Make a five-step prediction for { X2t (2), X2t (3), · · · , X2t (1001)} by repeating the above two
modeling steps until the five-step of { X2t (496), X2t (497), · · · , X2t (1495)} is achieved in the same way.

4.4. Evaluation Criteria

To quantitatively evaluate the accuracy and stability of proposed model, the mean absolute error
(MAE), mean relative percentage error (MRPE), root mean square error (RMSE) and root mean square
relative error (RMSRE) are utilized in this study, the calculation formulas are shown as follows:

1 nte
nte i∑
MAE = | x i − x i |, (19)
=1

1 nte  xi − xi 
 

nte i∑  x  × 100%,
MRPE = (20)
=1 i
s
1 nte
nte i∑
RMSE = ( x i − x i )2 , (21)
=1
v
u 1 nte  xi − xi 2
u

nte i∑
RMSRE = t , (22)
=1
xi

where xi and xi represent the measured data and the predicted data at the time t, respectively; nte is
the number of the data for performance evaluation.

4.5. Forecasting Results

For time series { X2t }, the ARIMA model (without considering of heteroscedasticity) and the
ARIMA-GARCH model (consider heteroscedasticity) are used to predict the deformation, and the
prediction results are shown in Figures 11 and 12. The error analysis results are shown in Tables 1–3.
Sensors 2018, 18, 298 13 of 17
Sensors 2018, 18, 298 13 of 18

-20

-30

-40
Deformation (mm)

-50

-60

The original deformation data

-70 One-step prediction by Kalman-ARIMA-GARCH
One-step prediction by Kalman-ARIMA

-802018, 18, 298

Sensors 13500
of 18
0 50 100 150 200 250 300 350 400 450
Sample number
-20 (a)
-20

-30
-30

-40
(mm)

-40
(mm)
Deformation

-50
Deformation

-50

-60
-60
The original deformation data
-70 One-step prediction by Kalman-ARIMA-GARCH
-70 One-step
The originalprediction by Kalman-ARIMA
deformation data
Three-step prediction by Kalman-ARIMA-GARCH
-80 Three-step prediction by400
Kalman-ARIMA
0 50 100 150 200 250 300 350 450 500
-80 Sample number
0 50 100 150 200 250 300 350 400 450 500
(a) number
Sample
(b)
(b)
Figure 11. Results of the predictions for the original deformation series {X2t} by the ARIMA and the
Figure 11. Results of the predictions for the original deformation series {X2t } by the ARIMA
Figure 11. Results
autoregressive of the predictions
conditional for the original
heteroscedasticity model deformation
(GARCH): (a)series {X2t} by
One-step the ARIMA
prediction; (b)and the
Three-
and the autoregressive conditional heteroscedasticity model (GARCH): (a) One-step prediction;
autoregressive
step prediction.conditional heteroscedasticity model (GARCH): (a) One-step prediction; (b) Three-
(b) Three-step prediction.
step prediction.
-20
-20

-30
-30

-40
(mm)

-40
(mm)
Deformation

-50
Deformation

-50

-60
-60

-70 The original deformation date

-70 Five-step prediction
The original by Kalman-ARIMA-GARCH
deformation date
Five-step
Five-step prediction
prediction by
by Kalman-ARIMA
Kalman-ARIMA-GARCH
-80 Five-step prediction by Kalman-ARIMA
0 50 100 150 200 250 300 350 400 450 500
-80 Sample number
0 50 100 150 200 250 300 350 400 450 500
Sample number
(a)

Figure 12. Cont.

Sensors 2018, 18, 298 14 of 17
Sensors 2018, 18, 298 14 of 18

-10

-20

-30
Deformation (mm)

-40

-50

The original maximum deformation date

Five-step maximum deformation prediction by
-60
Kalman-ARIMA-GARCH
Five-step maximum deformation prediction by
Kalman-ARIMA
-70
0 50 100 150 200 250 300 350 400 450 500
Sample number
(b)
Figure 12. Results of the predictions for the original deformation series {X2t} by the ARIMA and the
Figure 12. Results of the predictions for the original deformation series {X2t } by the ARIMA and the
GARCH: (a) Five-step prediction; (b) Five-step maximum deformation prediction.
GARCH: (a) Five-step prediction; (b) Five-step maximum deformation prediction.
Table 1. Analysis of the predictions given in Figure 11a.
Table 1. Analysis of the predictions given in Figure 11a.
Comparis
Indices Kalman-ARIMA Kalman-ARIMA-GARCH Improvements
on Results
MAE of average deformation prediction (mm) 4.407 3.402 Comparison
1.005 22.80%
Indices Kalman-ARIMA Kalman-ARIMA-GARCH Improvements
MRPE of average deformation prediction (%) 10.16 7.96 Results
2.20 21.65%
RMSE of average deformation prediction (mm) 5.317 4.341 0.976 18.36%
MAE of average deformation prediction (mm) 4.407 3.402 1.005 22.80%
RMSRE of average deformation prediction (%) 13.20% 10.59% 0.0261 19.77%
MRPE of average deformation prediction (%) 10.16 7.96 2.20 21.65%
MAE: deformation
RMSE of average mean absolute error;(mm)
prediction MRPE: mean relative percentage error;
5.317 4.341 RMSE: root mean0.976square error;
18.36%
RMSRE ofRMSRE:
average deformation
root mean prediction (%)
square relative 13.20%
error; 10.59%
ARIMA: autoregressive integrated moving0.0261
average model;19.77%
MAE:GARCH:
mean absolute error; autoregressive
generalized MRPE: mean relative percentage
conditional error; RMSE: root mean square error; RMSRE: root
heteroskedasticity.
mean square relative error; ARIMA: autoregressive integrated moving average model; GARCH: generalized
autoregressive conditional heteroskedasticity.
Table 2. Analysis of the predictions given in Figure 11b.
Kalman-ARIMA- Comparison
Table
Indices 2. Analysis ofKalman-ARIMA
the predictions given in Figure 11b.
GARCH Results
Improvements

MAE of average deformation prediction (mm) 5.887 5.098 0.789 13.40%

MRPE of average deformation prediction (%) 13.99 12.02 1.97
Comparison 14.08%
Indices
RMSE of average Kalman-ARIMA
deformation prediction (mm) 7.890 Kalman-ARIMA-GARCH
6.448 1.442Results Improvements
18.28%
RMSRE of average deformation prediction (%) 17.69 15.78 1.91 10.80%
MAE of average deformation prediction (mm) 5.887 5.098 0.789 13.40%
MRPE of average deformation prediction (%) 13.99 12.02 1.97 14.08%
Table
RMSE of average deformation prediction 3. Analysis7.890
(mm) of the predictions given in Figure 12.
6.448 1.442 18.28%
RMSRE of average deformation prediction (%) 17.69 15.78 1.91 10.80%
Kalman- Kalman-ARIMA- Comparison
Indices Improvements
ARIMA GARCH Results
Five step average deformation prediction
Table 3. Analysis of the predictions given in Figure 12.
MAE of average deformation prediction (mm) 6.505 5.847 0.658 10.12%
MRPE of average deformation prediction (%) 14.95 13.91 1.04 6.96%
RMSE of average deformation prediction (mm) 7.885 7.293 0.592 Comparison7.51%
Indices Kalman-ARIMA Kalman-ARIMA-GARCH Improvements
RMSRE of average deformation prediction (%) 20.11 18.19 1.92 Results 9.55%
Five step maximum deformation prediction
Five step average deformation prediction
MAE of maximum deformation prediction (mm) 6.151 5.666 0.485 7.88%
MAE of average deformation prediction (mm)
MRPE maximum deformation prediction (%) 6.50516.13 14.885.847 1.25 0.658 7.75% 10.12%
MRPE of RMSE
average deformation prediction (%)
of maximum deformation prediction (mm) 14.957.811 7.15613.91 0.655 1.04 8.39% 6.96%
RMSE of average
RMSRE deformation prediction (mm)
of maximum deformation prediction (%) 7.88523.07 21.077.293 2.00 0.592 8.67% 7.51%
RMSRE of average deformation prediction (%) 20.11 18.19 1.92 9.55%
Five step maximum deformation prediction
As shown in Table 1, the prediction accuracy of the Kalman-ARIMA-GARCH model is satisfactory
MAE of maximum deformation prediction (mm) 6.151 5.666 0.485 7.88%
with only 3.402 mm of MAE and 10.59% of RMSER. With the increase of prediction step, the degradation
MRPE maximum deformation prediction (%) 16.13 14.88 1.25 7.75%
ofofthe
RMSE accuracy
maximum appears,prediction
deformation as MAE(mm) of three-step prediction is 5.098 mm,
7.811 7.156 and RMSRE is 15.78%.
0.655 The MAE of
8.39%
RMSRE
fiveoflead
maximum deformation
forecast is 5.847 prediction
mm, and(%) the RMSRE 23.07
is 18.19%. Compared 21.07
with the Kalman-ARIMA2.00 8.67%
model, the
Kalman-ARIMA-GARCH model has obviously better forecasting performance.
As shown in Table 1, the prediction accuracy of the Kalman-ARIMA-GARCH model is satisfactory
with only 3.402 mm of MAE and 10.59% of RMSER. With the increase of prediction step, the degradation
of the accuracy appears, as MAE of three-step prediction is 5.098 mm, and RMSRE is 15.78%. The MAE
Sensors 2018, 18, 298 15 of 17

of five lead forecast is 5.847 mm, and the RMSRE is 18.19%. Compared with the Kalman-ARIMA
model, the Kalman-ARIMA-GARCH model has obviously better forecasting performance.
As can be seen from Figures 11 and 12, the prediction results have lag characteristics. The property
has little influence on the prediction results by sustained growth (or decrease) in deformation, however,
the prediction results will bring obvious error when the deformation mutation occurs. The main
reason for the hysteresis is that the prediction is realized by the historical deformation sequence, so it
is closely related to the historical information. Also, nonlinear behavior of a variable is simplified
to a linear model by the time series method, which causes a certain gap between simplify and the
actual situation; a recursive multi-step prediction model based on the historical deformation sequence
produce accumulative errors. As shown in Figures 11 and 12, as the GARCH model is a nonlinear
model, it can explain part of the nonlinear characteristics (heteroscedasticity). As shown by Tables 1–3,
the prediction effect of ARIMA-GARCH is better than that of the linear ARIMA model.
Based on the success of the five-step prediction by using GARCH model, the maximum
deformation value of five minute ahead can be predicted. The calculation method is as follows:
(1) according to the recursive time series modeling method, obtain five minute ahead average prediction
value; (2) calculate the correlation coefficient H of { X1t } between mean value of 60 s with the maximum
value according to the least square method, in this study, H = 0.97; and (3) revise the predicted values
obtained by step (1) according to H, and obtain the maximum predictive value in the five minutes.
The comparison between the predicted maximum value and the measured maximum one are shown
in Figure 12b and Table 3. The MAE of the proposed method is 5.666 mm, and the RMSRE is 21.07%.
Compared with Kalman-ARIMA model, the prediction result of Kalman-ARIMA-GARCH model is
more reliable.
The prediction model adopted in this paper is composed of a linear ARIMA model (explaining
subject part) and nonlinear GARCH model (explaining residuals part), which can only explain partial
nonlinear characteristics. To further explain this nonlinearity, the nonlinear model can be introduced
to explain the main part, while the residual term explained by the GARCH model. The specific form
can be expressed as
X (t) = f ( X (t − 1), · · · X (t − p)) + a(t) + θ ( B) × at , (23)

where X (t), a(t), p and θ ( B) are consistent as defined in Section 3.3; a(t) can be explained by the
GARCH model, f ( g) can be interpreted by the kernel density estimation of the Gaussian kernel
function using a window function.

5. Conclusions
In this paper, the residuals of the ARIMA model of the deformation sequence is verified, and
it is found that the deformation time series has an obvious heteroscedasticity effect. To address
this difficulty, the GARCH model is introduced to predict the deformation and compared with the
traditional linear ARIMA model. Some concluding remarks can be summarized as follows.
(1) Tables 1–3 indicate that for non-stationary data with embedded heteroscedasticity,
the prediction accuracy of GARCH model is superior to the ARIMA model, the prediction error
analysis results of the two models show that the GARCH model has certain advantages under the
background that the error appeared in the heteroscedasticity condition. The GARCH model captures
this nonlinear information to a certain extent.
(2) The overall actual predictive performance of the GARCH model is not inferior of ARIMA
model in the mean sense, and the GARCH model has some edges over ARIMA in a few indicators.
In particular, the GARCH model makes up for the imperfections that the ARIMA model is not satisfied
with the assumption of constant variance. Thus, GARCH has advantages from the standpoint of
theoretical rigor.
(3) The GARCH model provides a new feasible method for bridge deformation prediction.
The theories are well designed and established, and the practical prediction result is also satisfactory.
Sensors 2018, 18, 298 16 of 17

The results can lay a foundation for the early warning of bridge health monitoring system based on
sensor technology.
The GARCH model alone is not fully applicable and reasonable to explain the whole nonlinear
characteristics of deformation time series. More relevant models should be analyzed for the
further study.

Acknowledgments: This work was supported by the National Key Research and Development Program of China
(2017YFC0806007, 2016YFC0802202), the National Science Fund for Distinguished Young Scholars (51425801),
the National Natural Science Foundation of China (51708067) and the Scientific and Technological Project of
Chongqing Municipal Education Commission (KJ1600501).
Author Contributions: J.X. analyzed the monitoring data and wrote this paper. J.Z. designed the monitoring
system. S.X.Y. advised on the data analysis and paper organization. Y.W. involved in paper writing. X.L. wrote
part of the program.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Zhang, X.G.; Liu, G.; Ma, J.H.; Wu, H.B.; Fu, B.Y.; Gao, Y. Status and prospect of technical development for
bridges in China. Chin. Sci. Bull. 2015, 61, 415–425. [CrossRef]
2. Zhou, J.T.; Li, X.G.; Xia, R.C.; Yang, J.; Zhang, H. Health monitoring and evaluation of long-span bridges
based on sensing and data analysis: A survey. Sensors 2017, 17, 603. [CrossRef] [PubMed]
3. Liu, M.Y.; Li, Q.; Huang, Y.B.; Lu, Z.F.; Zhang, Q. Ultra-long time performance of steel-concrete composite
continuous beam in Hong Kong-Zhuhai-Macao bridge with creep and shrinkage of concrete slabs. China J.
Highw. Transp. 2016, 29, 60–69.
4. Sajedi, S.; Huang, Q.D. Load-deflection behavior prediction of intact and corroded RC bridge beams with or
without lap splices considering bond stress-slip effect. J. Bridge. Eng. 2016, 22, 04016102. [CrossRef]
5. Han, B.; Xiang, T.Y.; Xie, H.B. A Bayesian inference framework for predicting the long-term deflection of
concrete structures caused by creep and shrinkage. Eng. Struct. 2017, 142, 46–55. [CrossRef]
6. Cao, J.; Ding, W.Y.; Zhao, D.S.; Song, Z.G.; Liu, H.M. Time series forecast of foundation pit deformation
based on LSSVM-ARMA model. Rock Soil Mech. 2014, 35, 579–586. [CrossRef]
7. Chen, Y.; Ye, Y.Q.; Sun, B.N.; Lou, W.J.; Yu, J.H. Application of model prediction technology to bridge health
monitoring. J. Zhejiang Univ. Eng. Sci. 2008, 42, 157–163.
8. Chen, G.L.; Lin, X.G.; Yue, Q.; Liu, H. Study on separation and forecast of long term deflection based on time
series analysis. J. Tongji Univ. Nat. Sci. Ed. 2016, 44, 962–968.
9. Cheng, C.M.; Qin, P. Prediction of seawall settlement based on a combined LS-ARIMA model.
Math. Probl. Eng. 2017, 2017, 7840569. [CrossRef]
10. Yang, J.X.; Zhou, Y.X.; Zhou, J.T.; Chen, Y. Prediction of bridge monitoring information chaotic using time
series theory by multi-step BP and RBF neural networks. Intell. Autom. Soft Comput. 2013, 19, 305–314.
[CrossRef]
11. Lai, J.X.; Qiu, J.L.; Feng, Z.H.; Chen, J.X.; Fan, H.B. Prediction of soil deformation in tunnelling using artificial
neural networks. Comput. Intel. Neurosci. 2016, 2016, 6708183. [CrossRef] [PubMed]
12. Kao, C.Y.; Loh, C.H. Monitoring of long-term static deformation data of Fei-Tsui arch dam using artificial
neural network-based approaches. Struct. Control Health Monit. 2013, 20, 282–303. [CrossRef]
13. Kang, F.; Liu, J.; Li, J.J.; Li, S.J. Concrete dam deformation prediction model for health monitoring based on
extreme learning machine. Struct. Control Health Monit. 2017, 24, e1997. [CrossRef]
14. Lian, C.; Zeng, Z.; Ya, W.; Tang, M. Ensemble of extreme learning machine for landslide displacement
prediction based on time series analysis. Neural Comput. Appl. 2014, 24, 99–107. [CrossRef]
15. Zhou, J.T.; Yang, J.X. Prediction of bridge life based on svm pattern recognition. Intell. Autom. Soft Comput.
2011, 17, 1009–1016. [CrossRef]
16. Huang, T.; Chen, X.F.; Liu, L. Deflection prediction of long span pre-stressed concrete beam bridgebased on
ant colony optimization algorithm. J. Southeast Univ. Nat. Sci. Ed. 2013, 235–240. [CrossRef]
17. Jiang, Y.; Huang, G.Q. Short-term wind speed prediction: Hybrid of ensemble empirical mode decomposition,
feature selection and error correction. Energy Convers. Manag. 2017, 144, 340–350. [CrossRef]
Sensors 2018, 18, 298 17 of 17

18. Jiang, Y.; Huang, G.Q.; Peng, X.Y.; Li, Y.L. Method of short-term wind speed forecasting based on generalized
autoregressive conditional heteroscedasticity model. J. Southwest Jiaotong Univ. 2016, 51, 663–669. [CrossRef]
19. Huang, G.; Huang, G.B.; Song, S.J.; You, K.Y. Trends in extreme learning machines: A review. Neural Netw.
2015, 61, 32–48. [CrossRef] [PubMed]
20. Yu, P.S.; Yang, T.C.; Chen, S.Y.; Kuo, C.M.; Tseng, H.W. Comparison of random forests and support vector
machine for real-time radar-derived rainfall forecasting. J. Hydrol. 2017, 552, 92–104. [CrossRef]
21. Mahi, M.; Ömer, K.B.; Kodaz, H. A new hybrid method based on particle swarm optimization, ant colony
optimization and 3-opt algorithms for traveling salesman problem. Appl. Soft. Comput. 2015, 30, 484–490.
[CrossRef]
22. Harvey, A.C.; Todd, P.H.J. Forecasting economic time series with structural and box-jenkins models: A case
study. J. Bus. Econ. Stat. 1983, 1, 299–307. [CrossRef]
23. Harvey, A.C. Forecasting, Structural time Series Models and the Kalman Filter; Cambridge University Press:
Cambridge, UK, 1989.
24. Harvey, A.C.; Trimbur, T.M.; Dijk, H.K.V. Trends and cycles in economic time series: A bayesian approach.
J. Econom. 2007, 140, 618–649. [CrossRef]
25. Doorley, R.; Pakrashi, V.; Caulfield, B.; Ghosh, B. Short-term forecasting of bicycle traffic using structural time
series models. In Proceedings of the 17th International Conference on Intelligent Transportation Systems,
Qingdao, China, 8–11 October 2014; pp. 1764–1769. [CrossRef]
26. Soldaat, L.; Visser, H.; Roomen, M.V.; Strien, A.V. Smoothing and trend detection in waterbird monitoring
data using structural time-series analysis and the Kalman filter. J. Ornithol. 2007, 148, 351–357. [CrossRef]
27. Aamir, M.; Shabri, A. Modelling and forecasting monthly crude oil price of Pakistan: A comparative study
of ARIMA, GARCH and ARIMA Kalman model. AIP Conf. Proc. 2016, 1750, 060015. [CrossRef]
28. Ossandon, S.; Bahamonde, N. On the Nonlinear Estimation of GARCH Models Using an Extended Kalman
Filter. In Proceedings of the World Congress on Engineering 2011 Vol I WCE 2011, London, UK, 6–8 July 2011.
29. McLeod, A.I.; Yu, H.; Mahdi, E. Time Series Analysis with R, Handbook of Statistics; Elsevier: Oxford, UK, 2012;
Volume 30, pp. 661–712.
30. Omenzetter, P.; Brownjohn, J.M.W. Application of time series analysis for bridge monitoring. Smart. Mater.
Struct. 2006, 15, 129–138. [CrossRef]
31. De Lautour, O.R.; Omenzetter, P. Damage classification and estimation in experimental structures using time
series analysis and pattern recognition. Mech. Syst. Signal Process. 2010, 24, 1556–1569. [CrossRef]
32. Gul, M.; Catbas, F.N. Statistical pattern recognition for structural health monitoring using time series
modeling: Theory and experimental verifications. Mech. Syst. Signal Process. 2009, 23, 2192–2204. [CrossRef]
33. Ernst, F.; Helmut, W. The application of data mining in bridge monitoring projects: Exploiting time series
data of structural health monitoring. In Proceedings of the 22nd International Workshop on Database and
Expert Systems Applications, Toulouse, France, 29 August–2 September 2011; pp. 297–301.
34. Yu, J.Y.; Shao, X.D.; Yan, B.F.; Zhu, P. Research and development on global navigation satellite system
technology for bridge health monitoring. Chin. J. Highw. Transp. 2016, 29, 30–41.
35. Chen, Z.S.; Zhou, J.T.; Tim, T.K.; Gang, H.U.; Yong, L.I.; Xu, W. Alignment control for a long span urban
rail-transit cable-stayed bridge considering dynamic train loads. Sci. China Tech. Sci. 2016, 59, 1759–1770.
[CrossRef]
36. Xie, Y.C.; Zhang, Y.L.; Ye, Z.R. Short-term traffic volume forecasting using Kalman filter with discrete wavelet
decomposition. Comput.-Aided. Civ. Infrastruct. Eng. 2007, 22, 326–334. [CrossRef]
37. Engle, R.F. Autoregressive conditional heteroscedasticity with estimates of the variance of United Kingdom
inflation. Econometrica 1982, 50, 987–1007. [CrossRef]
38. Bollerslev, T. Generalised autoregressive conditional heteroskedasticity. J. Econom. 1986, 31, 307–327. [CrossRef]
39. Guo, Y.S.; Wang, C.; Zhang, Y.Q. Analysis of noise variance’s effect on Kalman filter result. Comput. Eng.
Design 2014, 35, 641–645.
40. Liu, H.P.; Erdem, E.; Shi, J. Comprehensive evaluation of ARMA–GARCH(-M) approaches for modeling the
mean and volatility of wind speed. Appl. Energy 2011, 88, 724–732. [CrossRef]

© 2018 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license (http://creativecommons.org/licenses/by/4.0/).