This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution license (http://creativecommons.org/licenses/by/3.0/).
This paper describes the procedures for development of signal analysis algorithms using artificial neural networks for Bridge WeighinMotion (BWIM) systems. Through the analysis procedure, the extraction of information concerning heavy traffic vehicles such as weight, speed, and number of axles from the time domain strain data of the BWIM system was attempted. As one of the several possible pattern recognition techniques, an Artificial Neural Network (ANN) was employed since it could effectively include dynamic effects and bridgevehicle interactions. A number of vehicle traveling experiments with sufficient load cases were executed on two different types of bridges, a simply supported prestressed concrete girder bridge and a cablestayed bridge. Different types of WIM systems such as highspeed WIM or lowspeed WIM were also utilized during the experiments for crosschecking and to validate the performance of the developed algorithms.
The concept of road vehicle weighinmotion (WIM) was first introduced in the 1950s in the United States. The purpose of this system was to overcome the drawbacks of static weighing and acquiring traffic information such as weight, speed, passing lane, axle spacing, and type of vehicle without interference with the traffic flow. To achieve these objectives, various sensors are installed beneath pavement layers or on a bridge superstructure and the acquired sensor signals are analyzed and saved.
Earlier WIM systems were developed as a lowspeed WIM system which can be applied for vehicles at speeds less than 20 km/h, and were mainly utilized for overweight vehicle detection. Later, highspeed WIM systems were developed to improve WIM systems, but the development suffered difficulties in attaining acceptable accuracy due to the sensitive dynamic interactions between vehicles and pavement surfaces. In Korea, a highspeed WIM system was developed by the Korea Highway Corporation, and is now operating on the Central Inland Highway for preselecting overweight vehicles.
The first Bridge WeighinMotion (BWIM) system can be traced back to Moses and Peters [
Since most of the existing BWIM systems are developed on the static influence line theory, the accuracy can be compromised due to dynamic behaviors or dynamic interactions between the bridges and vehicles. To solve this problem, a number of researches have been performed on measured influence line [
The application of artificial neural networks (ANN) to the BWIM was attempted in 2003 by Gonzalez
As a continuation of the previous research, the ANN method is additionally applied to a prestressed concrete girder bridge, and the results of the two applications are discussed. Unlike previously developed BWIM algorithms, the gross vehicle weight (GVW) is calculated first and then the axle weights are calculated by distributing GVW using axle weight distribution factors (AWDFs) in this study. The ANN is utilized to calculate both GVW and axle weights sequentially.
Two different types of bridges, namely a orestressed concrete (PSC) girder bridge and a cablestayed bridge were employed in this study. The first bridge, a prestressed concrete girder bridge, is the four simple spans' part of Geumdang Bridge (30+3@40m) which is located on the Central Inland Highway in Korea. Strain gauges of the BWIM system are installed on the 1st span of 30 m length. Geumdang Bridge consists of four PSC girders and a concrete deck of 12.6 m width which carries two traffic lanes.
The second bridge, a cablestayed bridge with a steelconcrete composite deck, is Seohae Bridge (
The sensors used to develop the BWIM system are all strain gauge, and the sensors can be categorized into weightmeasuring sensors and axledetecting sensors. Weightmeasuring sensors are installed on the lower surface of main girders and/or cross beams, and axledetecting sensors are installed on the lower surface of concrete deck.
In the case of the Geumdang Bridge, weightmeasuring sensors were installed on both girder and cross beams to compare the performances of each case, but in Seohae Bridge, only a cross beam was instrumented with weightmeasuring sensors since the strain from the main girder would not give sufficient accuracy due to its structural characteristics. Sensor dispositions are depicted in
Experimental trials using test trucks which were statically measured their axle and gross weights were performed on the Geumdang Bridge and Seohae Bridge. The test trucks repeatedly traveled over a lane of the bridges at a predefined speed several times, and the strain signals were analyzed by the suggested BWIM system. In case of Geumdang Bridge, three, four and fiveaxle dump trucks were employed as test trucks. The driving speed was varied from 5 km/h to 90 km/h, and the test trucks traveled 10 occasions for the 60 km/h and 90 km/h driving speed.
Similar types of test trucks such as three, four, and fiveaxle dump trucks were used on Seohae Bridge as well. The axle weights and representative test cases are depicted in
Data of various vehicle types with widely spread weight distribution are required for the training of ANN and validation of the output weights calculated by the trained ANN. This data also must to be acquired independently from other WIM systems simultaneously. The highspeed WIM system which was developed by a research project conducted by Korean Highway Corporation, and lowspeed the WIM system for overweight selection at the nearest tollgate were utilized for Geumdang Bridge and Seohae Bridge, respectively.
Representative BWIM signals are illustrated below, and the characteristics of the signals are discussed in this section.
As proven in the previous research [
Since Seohae Bridge is a cablestayed bridge with a 470 mlong main span, utilizing strains of the main girder makes weight calculation difficult since the number of vehicles simultaneously existing on the considering span increased due to structural characteristics. Moreover, stay cables anchored with a regular spacing of 12.3 m make the derivation of theoretical influence line complex. Therefore, strains of cross beams were believed to be more suitable for calculation of weights for this bridge.
Typical BWIM signals of Geumdang Bridge and Seohae Bridge are shown in
The conventional influence line theory first calculates axle weight directly from the measured signal and then the GVW is derived by summing axle weights. In this study, the contrary procedure is suggested. First, the GVW is calculated basically using strain readings of main girders and/or cross beams by GVW calculating ANN, then axle weights are resulted by multiplying GVW and axle weight distribution factors (AWDFs) which is the output of another ANN. The major input parameter of the AWDF calculating the ANN is the peak strain values of concrete deck which corresponds to each axle of a passing vehicle.
The principal input parameters of the GVW calculating ANN are: (1) the peak strain readings of main girders and/or (2) the peak strain readings of cross beams. Six channels of strain signals which consist of three channels of main girders and three channels of cross beams were available in Geumdang Bridge. In contrast, signals only from three channels of cross beams corresponding to the direction of traveling vehicle were utilized for Seohae Bridge. Vehicle speed, summation of axle spacing and summation of peak strains of the deck were additionally included in the ANN input parameters to increase accuracy.
An individual ANN was constructed separately for calculating axle weight distribution factors (AWDFs) which are used to calculate weight of each axle by multiplying to the resulting GVW. The input values of this ANN are peak strain values of the concrete deck corresponding to the axle of passing vehicle and axle spacings, and the output values of this ANN would be the AWDF for each axle. When AWDFs are obtained from the ANN, axle weights can be simply calculated by multiplying the GVW and AWDFs. The structure of ANN for AWDF calculation is depicted in
Independent data was acquired from an adjacent WIM system (i.e., highspeed WIM or lowspeed WIM for overweight selection) as mentioned previously.
The acquired data set was appropriate for training and testing the ANN since the data set contains of vehicles with various numbers of axles (from 2 to 6 axles) and total weights (from 100 to 400 kN), as shown in
When compared to the case of utilizing static weights, utilizing the WIM data of random trucks leads to relatively low convergence accuracy since the target values which were extracted from the WIM data contain certain portion of error from the former. However, it has advantages from the viewpoint of amount, distribution, acquisition time and cost of data when other WIM systems are available.
Calculated output weights of ANN for the remaining data which were not used for training were compared to the target values as a validation test. As shown in
Although the error statistics can be calculated with these results, discussion on the accuracies of the systems is not appropriate since the target values are not static weights and they initially contained errors. Therefore, accuracy of the systems will be discussed later with the results of test trucks which were statically weighed and driven.
Statically preweighed test trucks of 3, 4, and 5axles were driven repeatedly. Then the calculated dynamic weight from the suggested BWIM systems were compared to the static weights, and finally the accuracy classes of the European WIM specification, which were established by the WAVE project were determined from statistical analysis of the relative errors between the dynamic and static weights.
The resulting errors of the dynamic weights and accuracy determination results of Geumdang Bridge are presented in
Comparing the results in
Though a lower accuracy class than the existing lowspeed WIM system resulted from the ANN method, similar accuracy classes to the Geumdang Bridge cases could be achieved.
Existing influence line methods calculate individual axle weights first, and then the gross vehicle weight (GVW) is calculated by summing up the axle weights. In contrast, the suggested ANN method calculates GVW first, and then axle weights are calculated with axle weight distribution factors (AWDFs). Data acquired on the Geumdang Bridge is utilized to compare the performance of the axle weight calculations. The structure of the ANN for axle weight calculation is as mentioned previously, and the result is compared to the influence line method in this section.
The training set of the axle weight calculating ANN was prepared from the highspeed WIM system which is located near the Geumdang Bridge as in the previous case. According to the number of axles of passing vehicles, individual ANNs for 3, 4, and 5axle trucks were constructed and trained using random vehicles' data. When training was completed, the validation test followed using the remaining data that were not used for training.
Results of training and validation test are shown in
Determination of the axle weight accuracy class is carried out with the results of a 10 times' repeated experimental test of preweighed test trucks. The speed of the vehicles was controlled to closely maintain 90 km/h. The following figure shows the results of axle weight calculation with respect to the corresponding static weights. It can be confirmed that an overall ±20% error bound is satisfied.
As the case of GVW calculation, resulting accuracy classes of axle weights are compared with the results of a recent research using influence line method [
In this study, the applicability of artificial neural networks (ANN) is investigated for the improvement of conventional BWIM systems so that it can be implemented on longspan bridges (such as cablestayed bridges) where the application of influence line theory had difficulties.
The proposed algorithm mainly consists of two separately developed stages which are calculations from the gross vehicle weight (GVW) and the distribution of this GVW into individual axle weights, and ANNs for each stage. The ANN for the 1st stage calculates GVW by analyzing the dynamic strain signal measured from the main girders and/or cross beams, and the 2nd ANN calculates GVW distribution factors using peak strain values of concrete deck, axle spacings and speed of the passing vehicle. Finally, individual axle weights can simply be calculated from GVW and GVW distribution factors.
Data acquired from adjacent independent WIM systems (lowspeed WIM or highspeed WIM) were utilized for the training and validation tests of the ANNs. Experimental test data from three, four, and fiveaxle trucks whose their static weights were previously measured were also utilized for the accuracy class calculation that is established by the WAVE project.
The proposed method is applied to two different types of bridges (a prestressed concrete girder bridge, Geumdang Bridge, and a steelconcrete composite cable stayed bridge, Seohae Bridge) and the results compared to those of the conventional method.
For the GVW calculation, the proposed ANN method and conventional influence line method show similar accuracy classes. Therefore, the ANN method can be considered as an alternative to the influence line method for longspan bridges where it cannot be applied easily.
Moreover, the ANN method results in higher accuracy class than the influence line method for axle weight calculation. Consequently, the combination of both the influence line method and the ANN method is also possible for improving the axle weight calculation accuracy of existing BWIM systems.
The authors wish to express their gratitude for the support of Seoul National University of Technology, Dankook University, Korea Expressway Corporation and Hanyang University.
(a) Geumdang Bridge site. (b) Main girders, cross beams and a concrete deck. (c) Typical section (unit: m).
Seohae Bridge.
Sensor disposition of Geumdang Bridge (unit: m).
Sensor disposition of Seohae Bridge (unit: m).
(a) Experimental test on Seohae Bridge. (b) Static weighing of test trucks for Geumdang Bridge.
Representative BWIM signals of (a) Geumdang Bridge. (b) Seohae Bridge.
Geumdang Bridge random truck cases' histogram of (a) Number of axles. (b) Gross vehicle weight (GVW).
Results of training of (a) Geumdang Bridge—cross beam strain. (b) Seohae Bridge—north bound 3^{rd} lane.
Results of validation test of (a) Geumdang Bridge—cross beam strain. (b) Seohae Bridge—north bound 3^{rd} lane.
Performance comparison between ANN input parameters (Geumdang Bridge).
ANN construction for Geumdang Bridge (5axle random trucks) (a) Results of training. (b) Results of validation test.
Performances of trained ANN (a) 3axle test truck. (b) 4axle test truck. (c) 5axle test truck.
Weight of test trucks and test cases of Geumdang Bridge.

65.0  91.3  91.6      5  1 
10 ∼ 50  1  

73.6  73.9  80.8  80.8    60  10 
70, 80  1  

61.2  75.5  78.2  98.2  98.5  90  10 
Weight of test trucks and test cases of Seohae Bridge.

67.0  85.3  81.2      60  50 
65  10  

71.5  92.2  71.0  85.8    70  48 
75  8  

59.6  79.3  79.4  88.7  88.6  80  40 
Structures of ANN for GVW calculation.
Input parameters (number of values)  o Peak strain values of Cross beam (3) and/or girder (3)  
o vehicle speed (1)  
o Σ axle distances (1)  
o Σ peak strain values of deck (1)  
Layer (number of node)  Input layer: 1 (6 or 9 nodes)  Input layer: 1 (6 nodes) 
Hidden layer: 2 (10 and 7 nodes)  Hidden layer: 1 (10 nodes)  
Output layer: 1 (1 node)  Output layer: 1 (1 node)  
Transfer function  Pure linear – pure linear – pure linear 
Structure of ANN for AWDF calculation.
Input parameters (number of values)  o Peak strain values of deck (=number of axles) 
Layer (number of node)  Input layer: 1 (2 × {number of axles}  1) 
Transfer function  Pure linear – pure linear – pure linear 
Accuracy results of GVW from various BWIM algorithms (Geumdang Bridge).
ANN  CrossBeam  26  1.62  4.52  94.9  C(15)  15.0  99.0  
Girder  26  4.99  2.38  94.9  B(10)  10.0  94.9  
CrossBeam & Girder  24  1.56  3.77  94.7  B(10)  10.0  95.2 
Accuracy results of GVW from BWIM and Lowspeed WIM (Seohae Bridge).
ANN (Crossbeam)  25  0.58  5.46  94.7  C(15)  15.0  97.1  
Lowspeed WIM  33  3.82  2.71  95.5  B(10)  10.0  96.8 
Accuracy results of axle weights from ANN method.
single axle  40  1.45  6.20  95.6  C(15)  20.0  15.2  99.3  
group of axles  40  0.41  3.81  95.6  B+(7)  10.0  9.2  97.3  
axle of a group  80  0.40  4.99  96.6  B+(7)  14.0  11.9  98.8 
Accuracy results of axle weights from influence line method [
single axle  188  1.31  7.27  93.7  B(10)  15  14.8  94.0  
group of axles  239  0.18  5.26  93.9  B(10)  13  10.6  98.0 