Research Article: A Data-Driven Approach For Bridge Weigh-in-Motion From Impact Acceleration Responses at Bridge Joints
Research Article: A Data-Driven Approach For Bridge Weigh-in-Motion From Impact Acceleration Responses at Bridge Joints
Research Article: A Data-Driven Approach For Bridge Weigh-in-Motion From Impact Acceleration Responses at Bridge Joints
Research Article
A Data-Driven Approach for Bridge Weigh-in-Motion from
Impact Acceleration Responses at Bridge Joints
Haoqi Wang ,1,2 Tomonori Nagayama,2 Takaya Kawakatsu,3,4 and Atsuhiro Takasu3
1
College of Civil Engineering, Tongji University, Shanghai 200092, China
2
Department of Civil Engineering, Te University of Tokyo, Tokyo 113-8656, Japan
3
National Institute of Informatics, Tokyo 101-8430, Japan
4
Preferred Networks, Tokyo 100-0004, Japan
Received 28 September 2022; Revised 27 February 2023; Accepted 30 March 2023; Published 4 May 2023
Copyright © 2023 Haoqi Wang et al. Tis is an open access article distributed under the Creative Commons Attribution License,
which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Bridge weigh-in-motion (BWIM) serves as a method to obtain the weight of passing vehicles from bridge responses. Most BWIM
systems proposed so far rely on the measurement of bridge global vibration data, usually strain, to determine the vehicle load.
However, because the bridge’s global response is sensitive to all vehicles on the bridge, the global vibration-based BWIM
techniques usually sufer from inaccuracy in the case where multiple vehicles are present on the bridge. In this paper, a data-driven
approach is proposed to extract the passing vehicle’s weight and driving speed from vertical acceleration at the bridge joint. As
a type of local vibration, the impulse acceleration responses at a bridge joint can be recorded only during a short period when
a vehicle is passing over the joint and are thus not sensitive to vehicles at other locations of the bridge. A feld test is conducted at
a bridge to prepare labeled training data for the use of a convolutional neural network. One accelerometer is installed on the bridge
joint to record impulse acceleration, while the vehicle’s weight and driving speed are obtained from a WIM station and a camera
near the bridge, respectively. A network that detects the vehicle’s passage as well as its passing lane is frst proposed, followed by
a 1-D convolutional neural network that uses the raw data of acceleration as the input to predict the vehicle’s gross weight and
driving speed. A comparison is made between the 1-D network and an updated 2-D network that uses the wavelet coefcients as
the input matrix. Te latter one shows better performance, indicating that it is important to choose the proper input data for the
network to be trained. A transfer learning technique is used to test the feasibility of the proposed method. Results show that the
proposed method can be extended with limited data to bridges other than the bridge where the network is trained.
vehicle’s weight is calculated from the bridge responses as an CNN for the vehicle weight and driving speed for passing
inverse dynamic problem. Since this idea was introduced by vehicles, whose performance is compared with the 2-D CNN
Moses in the 1970s [4], research studies have proposed many described in Section 5. To increase the practicality of the
BWIM algorithms. Some of these algorithms are based on an proposed method, a transfer learning technique is used to
extension of Moses’ method [5–7], while others treat the extend the trained model to another bridge joint in Section 6.
problem from a system identifcation point of view [8–13]. Finally, some conclusions are drawn in Section 7.
While easy-to-install BWIM techniques using accelerome-
ters have been proposed, they sufer from two major dis- 2. Field Test and Data Collection
advantages. For example, in multiple vehicle cases, the
bridge acceleration measured by sensors is excited by all 2.1. Experimental Setup. Te feld test was conducted at an
vehicles present on the bridge. In such cases, the identif- expressway bridge in Gifu prefecture, Japan. Te bridge is
cation of each vehicle’s weight can become an ill- a two-span steel girder bridge with a length of 74 m. A
conditioned problem, especially when the distances camera was located near the exit of the bridge to capture the
among vehicles are relatively short compared with the bridge video of each vehicle passing over the joint. To measure the
span, thus reducing the identifcation accuracy of each ve- vehicle-induced responses, two wireless sensors equipped
hicle’s weight [14]. Moreover, it is reported that the lon- with Epson MA351AU three-axis accelerometers were lo-
gitudinal location, as well as the lateral location, of the cated on the entrance joint of the bridge [17, 18], namely,
vehicle at each time instant is also an important factor af- Accelerometer I and Accelerometer II, at each lane of the
fecting the identifcation accuracy [15], while such location bridge. Te sampling frequency was set to be 100 Hz, and an
information is, unfortunately, not necessarily available ac- internal fnite impulse response (FIR) Kaiser flter was
curately in real cases. A method that is not sensitive to the employed. As the proposed data-driven approach necessi-
presence of surrounding vehicles and the vehicle’s exact tates utilizing high-frequency acceleration responses at the
location on the bridge is desired. joint and given that the gentle roll-of flter does not
In this paper, a BWIM method that uses the acceleration eliminate signal components above the cutof frequency,
responses at bridge joints is proposed. When vehicles enter a postprocess band-pass flter was then implemented to
or leave the bridge, impact responses are observed due to the acquire a 10–20 Hz signal, which corresponds to the vehicle
joints at the ends of the bridge. Tis impact acceleration excitation frequency [19]. Te top view of this bridge and the
response is infuenced by many factors, including the ve- experimental setup are shown in Figure 1. In addition, the
hicle’s weight, instant driving speed, and the number of axles weight of all vehicles entering the bridge was measured at
while remaining insensitive to other vehicles at other lo- a weighing station embedded in the pavement several ki-
cations on the bridge. A feld test is conducted at a two-lane lometers away from the bridge.
girder bridge, and a data-driven approach utilizing the Te experimental setup consists of three subsystems,
impact acceleration is proposed. Training data for passing including (a) camera, (b) weighing scale, and (c) acceler-
vehicles’ weight and instant driving speeds at joints were ometers. Te functions of these subsystems are briefy in-
recorded by using a nearby vehicle weighing scale and troduced herein. For each vehicle coming into the bridge, the
a video camera installed near the bridge joints, respectively. camera captured a video of this vehicle, from which the
Accelerometers were used to measure bridge joint responses. driving speed, number of axles, appearance interval, and
Convolutional neural networks (CNNs) are used for the plate number were extracted by implementing a computer-
detection of passing vehicles at joints and then for vehicle vision technique described in [16]. Te vehicle appearances
weight identifcation. Te efect on the performance of the were utilized to compare with the monitoring system at the
network from using diferent types of input dataset is in- weighing scale to extract the measured weight of the cor-
vestigated by comparing a 1-D and a 2-D network structure. responding vehicle [20]. Te accelerometers and the camera
Te network is trained on one lane of the bridge and then are synchronized through postprocessing, making it possible
successfully extended to another lane by transfer learning, to extract the vehicle-induced responses for each specifc
showing the practicality of the proposed method. While vehicle. In this manner, the bridge acceleration responses,
there are BWIM proposals based on deep learning on bridge the vehicle’s entering and leaving time, the vehicle’s weight,
response data [16], this paper is unique in the sense that only driving speed, and other vehicle information are obtained
girder-end acceleration signals are employed. A 2-D net- for each vehicle, forming a database for the training of the
work structure with a wavelet transform is shown to improve deep learning network.
the performance, and transfer learning is shown to be
efective.
Tis paper is organized as follows: In Section 2, the 2.2. Brief Analysis of Collected Data. Typical impact accel-
experimental setup for the feld test to obtain various types of eration responses induced by a four-axle passing vehicle
dataset is described and some representative data are pro- with a weight of 11.4 t and a driving speed of 75.5 km/h are
vided. In Section 3, a neural network is proposed to detect depicted in Figure 2. Te distances between the axles of this
the passing vehicle from the measured bridge responses and vehicle are 1.76 m, 4.24 m, and 1.14 m, respectively. It is
to divide the passing vehicle into several classifcations shown that the vertical impulse acceleration is a local re-
according to their passing lanes. Based on the vehicle de- sponse within 1 second. Te impulse response of the four
tection from Section 3, Section 4 proposes a 1-D regression axles can be roughly detected, as shown by the dashed lines
Structural Control and Health Monitoring 3
74 m
Lane II Accelerometer II
Camera Joint
Lane I
Accelerometer I
(a)
(b)
Figure 1: Sketch and sensor layout of the test bridge: (a) top view and sensor layout of the test bridge; (b) lateral view of the test bridge.
in the fgure. Because this vertical impulse acceleration In the proposed network, the input is a matrix with 2 rows
mostly contains high-frequency components, sensors that and 300 columns, representing the time histories measured
have good performance in high-frequency ranges are pre- by using Accelerometers I and II within a 3-second time
ferred. Intuitively, the peak values of the joint acceleration window. Te output of this CNN structure is a vector
should increase together with the weight of the passing containing three values representing the three categories,
vehicles. However, a time history of around 400 seconds namely, “vehicle passing on Lane I,” “vehicle passing on
recorded by using Accelerometer I, which is labeled with Lane II,” and “not detected.” Te structure of the CNN is
camera detection and measured vehicle weights, shows that given in Figure 5, and the labels of the categories are listed
a heavier vehicle weight does not necessarily result in a larger in Table 1.
peak acceleration value, as shown in Figure 3. Tis phe- For the purpose of the classifcation, a softmax function,
nomenon indicates that the vehicle’s weight may be related which is expressed in equation (1), is adopted for the last
to some other factors with a more complex relationship. fully connected layer [21]:
Te above phenomenon is further illustrated in Figure 4.
eajc
Te maximum acceleration is plotted against the vehicle’s Sjc � , (1)
weight and driving speed for a total of 5900 vehicle passages. c eajc
Each vehicle is presented by a data point in the fgure. From where ajc represents the cth output element of the last fully
Figure 4(a), it is observed that, with an increase in the vehicle’s connected layer of the jth sample and Sjc is the corresponding
weight, the maximum acceleration also becomes larger. Te normalized value. Trough this function, the output vector
data are found to be very scattered, indicating that large contains values within the range between 0 and 1.
inaccuracy will occur if we use this linear trend to predict the Te loss function is defned by cross entropy and is
vehicle’s weight from the maximum joint acceleration. Tis is expressed in equation (2) to quantify the error level between
because the impact acceleration response is also afected by the predicted categories and the real categories:
driving speed, number of axles, axle distance, the vehicle’s
passing route, and so on, presenting a highly nonlinear re- Loss � − yjc lnSjc , (2)
lationship. In Figure 4(b), the maximum acceleration shows j c
no signifcant relationship with the vehicle’s driving speed.
Terefore, an algorithm that can well refect the nonlinear where
relationship among the vehicle’s weight, driving speed, and ⎨ 1,
⎧ if the jth sample belongs to category c,
joint acceleration, and many other factors are needed. yjc � ⎩ (3)
0, if not.
3. Vehicle Detection Based on Joint Acceleration For the training data labeled as Category I and Category
II, the acceleration time histories at Accelerometers I and II
Correct detection of vehicles entering and leaving the corresponding to vehicle passage over the joint were
bridge provides the foundation for vehicle weight esti- extracted based on the recordings from the camera. Each
mation. In this section, a CNN-based classifcation algo- category has 9730 labeled time histories for training. For
rithm is proposed to determine whether a vehicle is passing Category III, acceleration was extracted from the time pe-
across the bridge joint within a time window and to de- riods in which vehicles are not at the joint, including the case
termine from which lane the vehicle is entering the bridge. of ambient vibration and the case where vehicles are on other
4 Structural Control and Health Monitoring
×10-3
4
2
Acceleration (m/s2)
0
-2
-4
0 0.2 0.4 0.6 0.8 1
Time (s)
Acceleration
Axle passing
Figure 2: Impact vertical acceleration at the bridge joint induced by a passing vehicle.
0.01
23.9 t
0.005
Acceleration (m/s2)
21.1 t 15.6 t
22 t 11.4 t
19.5 t
-0.005
-0.01
0 50 100 150 200 250 300 350
Time (s)
Acceleration
Camera detection
Figure 3: A time history of impact acceleration labeled with measured vehicle weight.
0.025 0.025
Maximum acceleration (m/s2)
0.02 0.02
0.015 0.015
0.01 0.01
0.005 0.005
0 0
10 20 30 50 60 70 80 90 100
Vehicle weight (t) Vehicle speed (km/h)
Data Data
Fitting trend Fitting trend
(a) (b)
Figure 4: Relation of maximum acceleration against the vehicle’s weight and speed: (a) with vehicle weight; (b) with vehicle driving speed.
Structural Control and Health Monitoring 5
·· ·· 50
Acc at joints · 20 · 10
Lane I
Lane II 3
(2 x 100 x 20) (2 x 50 x 10)
passsing time Conv & Relu Conv & Relu Full Full
Δt
Table 1: Categories for vehicle detection. 4.1. 1-D CNN Structure for Vehicle Weight and Speed
Category I II III
Identifcation. From Section 3, the acceleration signals
represented in Figure 2 can tell whether a vehicle is passing
Description Passing lane I Passing lane II Not detected
over the joint. However, in addition to the presence of the
vehicle, further details about the passing vehicle can also be
extracted from the acceleration signals, i.e., the vehicle’s
parts of the bridge. Te amount of data in Category III is weight and driving speed. In this regard, a 1-D CNN
made equal to the number of the other two categories. structure is constructed, as shown in Figure 7, where the
Te labeled data were used to train the network by recorded signal of one accelerometer corresponding to the
a stochastic gradient descent method. Before starting detected lane of the vehicle is adopted as the input of the
training, 80% of the data were randomly selected as training network, and the output vector contains the passing vehicle’s
data, while the other 20% were used as validation and test weight and driving speed. Before being fed into the neural
data. Te accuracy and the loss function are plotted in network, the time history undergoes a time-shifting process
Figure 6 for each iteration of the training process. Validation to position the maximum absolute value of the input ac-
data are used to test the accuracy of the network every 50 celeration at the center of the time history. Tis step aims to
iterations. Te training process is terminated when the loss reduce the network input’s complexity and enhance the
function is larger than the previously smallest values three accuracy of vehicle weight and speed identifcation.
times in series. In Figure 6(a), the accuracy, which is defned Similar to the training process for vehicle detection
as the ratio of the count for correct samples to the number of described in Section 3, 80% of the data serve as training data,
all samples, keeps increasing for both training and validation while the other 20% are adopted as the validation and test
data with the number of iterations. After around 900 iter- data. Before training, the input matrices are normalized by
ations, the abovementioned termination criteria are met, dividing the maximum values of the weights and speeds of
and the training process is thus stopped. Te fnal accuracy all passing vehicles. Te output of the network is also
of the trained network reaches 94.93%. normalized as follows:
To evaluate the classifcation accuracy of the trained
network, a confusion matrix is constructed, as shown in wi − min w
wi,norm � ,
Table 2, which includes the classifcation results of all the test max w − min w
data. Te values of the diagonal elements indicate the (4)
s − min s
number of results correctly classifed, while other values si,norm � i .
indicate the wrongly classifed results. Te accuracy is cal- max s − min s
culated and listed at the bottom of the matrix for each
In this manner, the output values of the network are all
category, showing that the presence and the passing lane of
in the range between 0 and 1.
the coming vehicle can be detected with acceptable accuracy.
To quantify the prediction error of the vehicle weight and
Once a vehicle passing across the joint is detected, the
driving speed, the loss function for the regression CNN
identifcation of the vehicle weight and the driving speed is
network is defned as follows:
the next step, which is discussed in the following sections.
1 2 2
Loss � wpre,i − wtar,i + spre,i − star,i , (5)
2 i
4. Vehicle Weight and Speed
Identification through 1-D CNN where wi and si stand for the weight and driving speed of the
ith vehicle sample and the subscripts “pre” and “tar” indicate
In this section, the impact acceleration responses at bridge the predicted and target values, respectively.
joints are used to give an estimation of the vehicle’s weight
and driving speed using 1-D CNN for regression. Te labeled
training data include the vehicle-induced joint vertical 4.2. Training Process and Test Results for 1-D CNN. Te
impulse acceleration, the vehicle’s weight, and instant evolution process of the loss function is shown in Figure 8.
driving speed. Te weight and driving speed are extracted As the training process starts, the loss function starts to
from a weighing scale and a camera, respectively. decrease from a high value. Same as in Section 3, the
6 Structural Control and Health Monitoring
100
80
Accuracy (%)
60
40
20
0 100 200 300 400 500 600 700 800 900
Iteration
Training data
Validation data
(a)
1
Loss
0.5
0
0 100 200 300 400 500 600 700 800 900
Iteration
Training data
Validation data
(b)
Figure 6: Training process of the network of classifcation: (a) accuracy; (b) loss.
·· ·· 10
· 20 · 10
Acc at joints
Lane I
2
(1 x 100 x 20) (1 x 50 x 10)
passsing time Conv & Relu Conv & Relu Full Full
Δt
Figure 7: 1-D CNN network for vehicle weight and speed estimation.
Structural Control and Health Monitoring 7
validation data are applied to the network every 50 itera- explained in the next section, whose performance is
tions, and the training process stops when the loss function evaluated and compared with the one given by the 1-
is larger than the previously smallest values three times in D CNN.
series to prevent overftting.
For the estimated values of vehicle weights and driving 5. Vehicle Weight and Speed
speeds, the target value and predicted value from the net- Identification through 2-D CNN
work for all test data are plotted in Figures 9(a) and 9(b),
respectively. A reference line, y � x, is plotted in the fgure. When utilizing CNNs to complete a task, the selection of the
Ideally, the data points represented by circles in grey should input matrix is important and may have a signifcant in-
be on the reference line. However, due to many factors, fuence on the accuracy of the prediction. Although a strong
including training error and measurement noise, the data and powerful network has the capability to extract the high-
points are scattered around the reference line. To quantify dimensional relation between the input and the output, it is
the scatter level of these data points, a correlation coefcient always expected that the relevant properties are not buried
(CC) is calculated following the defnition of equation (6) too deep in the input signals. For the identifcation task of
and is shown together in the fgure: the vehicle weight and driving speed, although it is still
i xi − x y i − y possible to use the raw signals measured by using the ac-
r � ������������������� �, (6) celerometers in the same way as in Section 3, it is better to
2 2
i xi − x i yi − y apply preprocessing to expose some key properties of the
joint acceleration response. In this section, a wavelet
where xi, and yi are substituted by the target and predicted transform is adopted and briefy reviewed herein.
values of the ith vehicle. For the vehicle weight and driving
speed, this coefcient is calculated to be 0.87 and 0.69,
respectively. 5.1. Wavelet Transform to Generate 2-D Signals. Te con-
An error distribution histogram for vehicle weight and tinuous wavelet transform, which is defned in equation (10),
driving speed prediction is plotted in Figure 10 as the error is widely used in signal processing and vibration analysis:
defnition of ∞
1 t−τ
y − xi C(τ, f) � ���� x(t)Ψ∗ dt, (10)
εi � i , (7) s(f) −∞ s(f)
xi
in which x(t) is the input signal to be analyzed, s(f ) is the
where εi indicates the relative error of the ith vehicle. Te scale parameter related to frequency f, and Ψ∗ indicates the
error distribution has its highest value around zero, in- complex conjugate of the mother wavelet Ψ. In this study,
dicating that the proposed algorithm neither overestimates the mother wavelet Ψ is chosen to be a Morse wavelet. Te
nor underestimates the target. shape of the wavelet is decided by a symmetry parameter and
To further evaluate the entire performance of the net- a time-bandwidth parameter. In this study, the above-
work, a mean absolute error (MAE) and a root mean square mentioned two values are empirically set at 3 and 60,
error (RMSE) are defned in equations (8) and (9), respectively.
respectively, Figure 11 shows the wavelet transform of a time history
1 recorded by using Accelerometer I induced by a two-axle
MAE � yi − xi , (8) vehicle passing on Lane 1. Note that these wavelet coefcients
N i are both time- and frequency-related. In this manner, the
������������� original 1-D signal is converted to a 2-D matrix, which is
1 2 helpful to develop the advantages of CNNs. Moreover, the
RMSE � y i − xi , (9)
N i properties of the recorded signal in both the time and fre-
quency domains are clearly expressed in this matrix, making
it easier for the CNN to extract its hidden features.
where N is the total number of vehicles.
From the abovementioned defnitions, the MAE and
RMSE of the vehicle’s weight and driving speed are calcu- 5.2. CNN Structure for Vehicle Weight and Speed
lated and listed in Table 3. It is observed that the estimation Identifcation. As stated above, the input matrix of the
accuracy is within an acceptable level, while RMSE is usually network for vehicle weight and speed identifcation is the
larger than MAE due to the existence of very large and low, wavelet coefcients. For a 3 s time history with 100 Hz
and thus rare, vehicle weights and driving speeds, as in- sampling frequency as in the current case, the wavelet
dicated in Figure 9. analysis gives 54 values along the frequency axis. Terefore,
Note that, in this network, the recorded acceleration the input matrix in this study has a size of 54 × 301, cor-
responses are put into the network without any pre- responding to the frequency and the time domain. Two
processing technique. Te network needs to fnd out the convolution layers with sizes of 20 × 20 and 10 × 10 follow
deep relation between the input and output data. To in- the input matrix. ReLU function is used after each of the
crease the prediction accuracy of the vehicle’s weight and convolution layers to provide the network with the ability to
driving speed, a 2-D CNN based on wavelet coefcients is extract nonlinear features. Te maximum pooling layer has
8 Structural Control and Health Monitoring
10-1
Loss
10-2
0 100 200 300 400 500 600 700 800 900 1000
Iteration
Training data
Validation data
Figure 8: Loss function for the 1-D network of vehicle weight and speed identifcation.
r = 0.87 r = 0.69
35 100
30
Predicted Speed (km/h)
Predicted Weight (t)
25 80
20
15 60
10
5 40
5 10 15 20 25 30 35 40 60 80 100
Target Weight (t) Target Speed (km/h)
Data Data
Reference Reference
(a) (b)
Figure 9: Test results for vehicle weight and driving speed: (a) vehicle weight; (b) driving speed.
0.4 0.4
0.3 0.3
Frequency
Frequency
0.2 0.2
0.1 0.1
0 0
-1 -0.5 0 0.5 1 -1 -0.5 0 0.5 1
Estimation error Estimation error
(a) (b)
Figure 10: Error distribution for vehicle weight and driving speed: (a) vehicle weight; (b) driving speed.
Structural Control and Health Monitoring 9
×10-3
Wavelet coefficient
0
40
30 3
20 2
Frequ
enc y 10 1
(Hz) 0 0 Time (s)
(a)
40
30
Frequency (Hz)
20
10
0
0 0.5 1 1.5 2 2.5 3
Time (s)
(b)
Figure 11: Wavelet coefcients of joint responses induced by a two-axle vehicle: (a) 3-D surface for wavelet coefcients; (b) vertical view of
the 3-D surface.
a stride of 2 × 2. Te fully connected layers lead to an output Te relative error distribution is plotted in Figure 15. For
vector containing two elements, which are represented by both the vehicle weight and driving speed, the histograms
vehicle weights and driving speeds, respectively. Te are observed to be sharper and narrower compared to those
structure of the network is depicted in Figure 12. in Figure 10, indicating that the estimation accuracy is in-
creased using the 2-D structure. Table 3 summarizes the
MAE, RMSE, and CC for the two CNN structures. Results
5.3. Training Process and Test Results. Te loss function is show that the 2-D structure performs better than the 1-D
defned in the same way as in Section 4 and is plotted with the structure from the perspective of all accuracy indices.
training process in Figure 13. Tis curve is similar to the one in Te test results are also examined from the aspect of the
Figure 8 that shows the training process for the 1-D CNN. It is vehicle weight histogram, which is of signifcant importance
observed that the fnal value of the loss function is lower than its in the feld of bridge weight-in-motion because it is highly
peer in the 1-D CNN, indicating that the new network pro- related to bridge fatigue life. Such a histogram could be
posed in this section performs better than the previous one. directly adopted in the fatigue analysis. In Figure 16, the
Te target and prediction values are plotted in Figure 14. target and predicted histograms are plotted together for
Similar to the points shown in Figure 9, the data points are comparison. It is observed from the target histogram that
scattered around the reference line. However, the data points are most vehicles are within the ranges of 12–16 t and 20–25 t.
closer to the reference line compared with the results from 1-D Te predicted histogram is shown to coincide well with the
CNN, as is refected by the CCs by comparing with Figure 9. target one.
10 Structural Control and Health Monitoring
·· ·· 20
Acc at Lane I · ·
54 2
(20 x 30 x 20) (10 x 10 x 10)
Conv Conv Full Full
Relu Relu
MaxPool MaxPool
passsing time
Δt
Figure 12: Structure of the network for vehicle weight and speed identifcation.
10-1
Loss
10-2
0 100 200 300 400 500 600 700 800 900
Iteration
Training data
Validation data
Figure 13: Loss function for the 2-D network of vehicle weight and speed identifcation.
r = 0.93 r = 0.79
35 100
30
Predicted Speed (km/h)
Predicted Weight (t)
25 80
20
15 60
10
5 40
5 10 15 20 25 30 35 40 60 80 100
Target Weight (t) Target Speed (km/h)
Data Data
Reference line Reference line
(a) (b)
Figure 14: Test results for vehicle weight and driving speed: (a) vehicle weight; (b) driving speed.
6. Possibility of Algorithm Extension by the vehicles entering Lane 1 of this bridge, where the net-
Transfer Learning work is trained. Te practicality of the proposed network is
thus quite limited. In engineering practice, it is desired that
Te abovementioned training and testing processes are the trained network can be conveniently adopted when
based on Accelerometer I located at the entrance of Lane 1. applied to other bridges. However, although the quantitative
Strictly speaking, the trained networks are only suitable for nonlinear physical relation between the vehicle’s weight and
Structural Control and Health Monitoring 11
0.4 0.4
0.3 0.3
Frequency
Frequency
0.2 0.2
0.1 0.1
0 0
-1 -0.5 0 0.5 1 -1 -0.5 0 0.5 1
Estimation error Estimation error
(a) (b)
Figure 15: Error distribution for vehicle weight and driving speed: (a) vehicle weight; (b) driving speed.
0.3
0.25
0.2
Distribution
0.15
0.1
0.05
0
10 20 30
Vehicle weight (t)
Target
Predicted
Figure 16: Target and identifed vehicle weight distribution.
vertical acceleration may be diferent because of the con- entering from Lane II. Te same training process is con-
struction details of the joints, some basic characteristics may ducted, and the training process is plotted in Figure 17. On
be common. For example, the impact acceleration responses the other hand, as a test to illustrate the efectiveness of
at bridge joints are all impulse-like responses, with the transfer learning, the frst convolutional layers, ReLU layers,
number of peaks roughly determined by the number of axles, and pooling layers are frozen, while only fully connected
and the peak values of the impulse are related to the vehicle layer parameters are trained, which are plotted in the same
weight. Tese similarities provide the basis for the extension fgure. It is clear that the training process with and without
of the proposed network to other lanes or even other bridges. transfer learning behaves much diferently. In the case of no
Te extension is analyzed through a technique known as transfer learning, the termination criterion is satisfed much
transfer learning. If two tasks share similar characteristics, earlier than in the case of transfer learning, and the efect of
the trained network from the frst task is adopted as the overftting is also observed, possibly due to insufcient
initial value of the network for the second task. In this training data. When transfer learning is applied, it takes
manner, the number of training data needed for the second more iterations to reach the termination criterion, and the
network can be much smaller than in the case where the fnal value of the loss function becomes lower than the one
network is randomly initialized. without transfer learning.
In this paper, transfer learning is tested by using vehicle Te target and predicted values of the test data are
data entering Lane II. From measurement records, there plotted in Figure 18. Te results for training with and
were 5160 vehicles entering from Lane 1 and 740 vehicles without transfer learning are plotted together. Because these
12 Structural Control and Health Monitoring
10-1
Loss
10-2
0 100 200 300 400 500 600 700 800 900
Iteration
35
30
Predicted Weight (t)
25
20
15
10
5
5 10 15 20 25 30 35
Target Weight (t)
Without TL
With TL
Reference
Figure 18: Test results for test data with and without transfer learning.
two sets of data have the same number of data points, the Table 4: Accuracy for training with and without transfer learning
scattering of the points without transfer learning clearly for the vehicle’s weight.
indicates that the adoption of transfer learning does have Lanes I II
a good efect on higher estimation accuracy. Number of training data 5160 740
Table 4 shows some details of the training process dis- Transfer learning No No Yes
cussed above. Te number of vehicles passing over Lane II is MAE (t) 1.32 1.90 1.31
much less than the number passing Lane I. Tis reduction in RMSE (t) 1.76 2.55 1.67
the number of training data undoubtedly leads to a larger CC 0.93 0.86 0.94
error, i.e., MAE and RMSE, and a lower CC when estimating
the passing vehicle’s weight. Once transfer learning is in-
troduced, the estimation error decreases to a comparable for bridges where there is no WIM station available, portable
level as in the case where sufcient training data are BWIM systems based on accelerometers can be temporarily
available. adopted on the bridge to obtain the passing vehicle’s weight
Tis section tests the possibility of using transfer learning for training [7, 22]. Once the training process is fnished
when there is not sufcient data to train the network from using the proposed method, most accelerometers on the
the beginning. In this paper, the training data for vehicle bridge can be removed, while only one accelerometer for
weights are obtained from a WIM station nearby. A future each lane at the bridge joint remains on the bridge for long-
BWIM scenario based on this paper’s fndings is as follows: term BWIM purpose.
Structural Control and Health Monitoring 13