Evaluation of bridge structure damage based on deep learning identification model

Abstract

Safety monitoring is an important part of bridge engineering construction and operation. At present, there is room for promoting the health monitoring and evaluation of small and medium-sized concrete bridges. In view of this, the study first models the spatial model and physical parameters of the bridge, and then builds the data of vehicle load and vehicle type. To reduce the complexity of data mapping, wavelet packet decomposition is used to analyze the data structure. And the physical field effect analysis is abandoned to directly mine the data relationship at both ends by using deep neural network. The data decomposition results show that the method can discard the temperature-induced effect. And the local decomposition results of the data meet the input of the neural network. The data measured by the sensor is added to the depth learning model for fitting. The overall and local fitting rates are more than 92%. The loss function converges quickly, and there is no gradient explosion. The model predicts the bridge structural damage caused by vehicle stress of four load categories, and the results show that the average fitting rate is 89.72%. Therefore, the identification path of the proposed deep learning model has positive significance for the evaluation of bridge structural damage. The main contribution of the study is to propose a deep learning-based method for bridge structural damage assessment. By modeling the spatial model and physical parameters of the bridge and combining data from vehicle load and vehicle type, the data structure was analyzed using the wavelet packet decomposition method to eliminate temperature-induced effects and data from sensor measurements were added to the deep learning model for fitting. This finding has positive implications for bridge structural damage assessment and can provide effective pathways and methods to monitor and evaluate the health status of small and medium-sized concrete bridges. This has important practical application value for the construction and operation of bridge engineering.

Keywords

Deep learning bridge structural damage assessment role wavelet packet decomposition

1. Introduction

Cross-regional cooperation and international economic communication have deepened with the opening up. As one of the important factors affecting economic cooperation, transportation has become a key area of infrastructure construction, and bridge construction is one of them [1]. The Hong Kong-Zhuhai-Macao Bridge and the Yangtze River Bridge in the urban area have demonstrated the construction technology as large-scale bridges. The construction and maintenance of small and medium-sized concrete bridges are the main business of bridge engineering. Due to the short construction period and long service time, the load of the bridge itself has a critical value, and the increase of traffic is easy to cause the overload of the bridge. With the passage of time, potential safety hazards have accumulated and developed into safety accidents. To solve the problem of bridge safety, the state has issued a general specification for carrying loads. However, insufficient supervision of bridges in remote areas and the impact of geological disasters have become inevitable objective factors. Therefore, the development of effective bridge health detection has become another way to solve the problem of bridge safety [2]. The upgrading of hardware technologies such as sensors and data processing methods such as big data and artificial intelligence provide ideas for intelligent detection of bridges. Massive data and mature model framework provide processing solutions for bridge segments with missing data [3]. As a method of automatic feature extraction, deep learning can mine the relationship between features in existing or similar scenes. And it can meet the background management requirements of remote monitoring and bridge health data [4]. At the same time, it can conduct a reasonable evaluation of the bridge status, reveal the impact of structural response during bridge service, as well as the process of damage accumulation and the inherent laws of mutual influence, thereby better improving the accuracy of bridge health monitoring and evaluation, making bridge monitoring more digital, information-based, and intelligent. In view of this, the research will take the identification model of deep learning as the main framework, and build the adaptive model of load and space calculation for the bridge. The overall framework is optimized by combining various deep learning networks to solve the actual nonlinear relationship construction. Finally, the effect of prediction and evaluation of the built bridge damage structure is analyzed by using big data drive. It is expected to provide ideas for bridge safety construction and reduce the occurrence of relevant safety accidents.

2. Related works

From human observation to intelligent detection system, bridge health management has made great progress. Scholars have combined physics, geography and other disciplines to conduct comprehensive research and application. Liang believes that bridge health detection after extreme events is crucial. Therefore, he developed a post-disaster detection method for reinforced concrete bridges based on three-level images. In this method, the convolution neural network is used to classify the fault and locate the damage level, and the hyper-parameter is used to optimize the model. This model shows superior performance under Bayesian optimization [5]. Omer is inspired by the game to build a new bridge detection method. This method first uses light detection and ranging to digitize the bridge, and then uses the virtual reality toolkit used by the game industry to conduct virtual generation. They evaluated a brick and stone bridge, and the results showed that the method can achieve key detection under safe conditions. At the same time, the price of the tool kit is within the affordable range, which improves the practicability [6]. To avoid the risk and extra time cost caused by visual judgment, Xie’s team developed a robot detection system. The system uses an industrial camera as the front end of visual sensing to capture the visual information of the bridge. It uses the combination of point and line features of two-dimensional and three-dimensional images to reduce the image mosaic error. During information processing, it is grouped according to the image characteristics, and then a complete panoramic image is formed according to the multi-band hybrid algorithm. The results of field images show that this method has advantages and can be optimized [7]. Similarly, the visual detection method developed by Saleem research team uses robot and computer vision to quantify and locate faults. When visual detection is combined with geographic markers and three-dimensional coordinate positioning, convolutional neural network is used to automatically identify bridge damage and generate global maps. Compared with the actual detection results, this method is effective [8]. Ye and other scholars have designed an auxiliary program combined with robots to enhance the status evaluation of bridge defects. This method extends the defect features with specific storage organization to achieve quantitative analysis. The data acquisition of digital images first classifies the structure and takes pictures in sequence. The defect recognition principle is the color anomaly processing of the image. The test results of Italian railway data show that this method has certain advantages and can be continuously improved [9]. Nguyen and other scholars have designed an automatic detection robot that can climb on steel structures. Their robot can climb on reciprocating structures and magnetic roller chains, and adapt to various steel structures. The prototype is made by integrating Hall effect and various sensing technologies, and the robot is strictly controlled according to dynamics and working condition analysis. The stability experiment results show that the robot can complete the real scene shooting task with a camera [10].

Deep learning can process nonlinear relationships to extract features from data, so scholars have developed fault diagnosis models applicable to various fields. Zhang and other scholars believe that the characteristics learned by traditional intelligent encoders will produce shift mutation characteristics and error classification. Therefore, they use sparse self-coder to join the diagnosis. At this time, when the original signal is input into the diagnosis network, meaningful features can be obtained locally. Then, the displacement invariant features are extracted from the feature area. The results show that their optimization path has development significance [11]. Wang’s research team will use deep learning for fault diagnosis of chemical engineering, and improve the data loss caused by traditional neural network compression features. It is proposed that the deep neural network uses the original data and the hidden feature set as the input of the restricted Boltzmann machine during training. Then the fault classifier is constructed according to the dynamic nature of the data. The real test results show that their optimization scheme has better classification performance [12]. Zhao and other scholars proposed a new deep learning method for high-precision fault diagnosis. The deep residual shrinkage network is used as the feature extraction of high noise vibration signal, and the soft threshold is used as the nonlinear conversion layer to eliminate redundant features. The selection of threshold value uses a special neural network as a trainable module. The simulation test results of different noises show that this method is worth popularizing [13]. The research team of Saufi optimized the data acquisition problem during deep learning and fault analysis. They designed a stacked sparse self-encoder to process sample data to solve the data types generated by sensors that are not conducive to analysis. At the same time, they also developed a diagnosis system based on time-frequency image pattern recognition. The actual results show that the system can achieve accurate fault prediction under limited samples. Compared with the traditional model, the performance is improved by 10%–20% [14]. Yu and other scholars believe that the labels of data sets and training sets in the deep learning model of fault diagnosis are consistent, which is not conducive to the diagnosis of unknown faults. Therefore, they have developed shared domain open set fault diagnosis and cross-domain open set fault diagnosis under mechanical conditions. At the same time, the output of the additional domain classifier is used to construct the bilateral weighted confrontation network. The test results confirm the superiority of this method on different data sets [15].

Scholars have made remarkable achievements in theoretical research and application of bridge fault diagnosis. They took into account the hardware upgrade of the actual scene and the technical improvement of the background management. In the field of deep learning used for fault diagnosis, scholars have optimized and upgraded the problems in the actual scene. This optimization scheme and idea provide ideas for the improvement of the diagnosis model of the research. The research will carry out adaptive construction of structural damage based on the existing bridge fault diagnosis combined with the deep learning improvement scheme, in order to achieve higher evaluation results.

3. Building of bridge structure damage identification model based on deep learning

3.1 Construction of bridge finite element model integrating vehicle load and stress

In actual management, the vehicular load is the main source of bridge structural damage except for natural disasters. Therefore, the finite element model of the bridge first analyzes the vehicle load and constructs a dynamic data detection system. The load of the vehicle is not only related to the total weight, but also to the actual pressure and vehicle speed. Therefore, the data detection is based on the speed, weight and vehicle parameters [16]. Some abnormal data will appear in the test system built by the sensor, of which the abnormal data generated under the condition of exceeding the measurement accuracy and interference is the main part. At the same time, there is also the situation that the system will judge multiple vehicles as one vehicle due to the close distance between vehicles. The measurement distribution built under the condition of abnormal conditions is shown in Fig. 1.

Figure 1.

Bridge dynamic weighing system.

The source of data obtained by the sensor is vehicle load, which is a random data for the bridge in Fig. 1. There are scale differences in vehicle configuration and on-board mass. If each vehicle is analyzed separately, it will lead to low calculation efficiency and high cost. Therefore, it will be calculated according to the equivalent vehicle weight of Eq. (1).

$\displaystyle W_{e}=\left({\sum{f\pi r^{3}}}\right)^{\frac{1}{3}}$ (1)

$W_{e}$ is the equivalent vehicle weight in Eq. (1). $f$ is the frequency of the same vehicle type. $r$ is the wheelbase of the vehicle. The research will take the number of vehicle axles and wheelbase as the main parameters of vehicle type classification, and make statistics based on the vehicle data measured. Vehicles with specifications accounting for less than 1% were excluded and finally divided into 6 categories according to the number of axles. According to the actual parameters of various vehicles, the model will be built in combination with probability and mathematical calculation. The structural response under the load of the bridge needs to be modeled for spatial calculation. The research will use the grillage method under the finite element analysis to build the model. The basic idea is to replace the fixed structure with equivalent grillage or space truss, and the stiffness structure of the box girder section is included in the adjacent equivalent grillage. The advantage of the grillage method is that the modeling is simple and the cost is low. According to the actual application, the centerline of the formed longitudinal member of the grillage will coincide with the true center, which is the minimum standard to ensure the calculation accuracy [17]. The structure of the geometric figure is shown in the Fig. 2.

Figure 2.

Beam lattice method of bridge geometry model cross-section.

In the cross section of the bridge geometric model, a longitudinal beam grid is composed of a box member and two flanges in Fig. 2. And the entire cross section is composed of five longitudinal beam grids. The center of the overall beam grid coincides with the actual center line. A transverse diaphragm construction is added, and the center line coincides with the center of the longitudinal beam construction. If the transverse and longitudinal construction are parallel to the horizontal line, the model can be used for stress analysis. In the actual stress analysis, the model will be meshed. Because the research and measurement is the structural response under vehicular load, a large number of transient load calculations will be involved. The number of grid elements is determined to be 1330 through experimental method rehearsal, with a total of 2645 nodes. The macro application scenario is composed of multiple transients. The probability distribution model of vehicle load needs to be used for reverse reasoning when working conditions are loaded, which can map local features into the overall data space. This process should first determine the vehicle category and vehicle type, and determine the driving lane of the vehicle according to the real lane distribution probability. Then the vehicle speed is calculated using the speed fitting probability model, and the axle data is determined according to the axle load fitting model. After determining the parameters of the vehicle, it will judge whether it meets the standard, and finally generate the working condition data. The local response data and the dynamic displacement global response data are combined to form the data set for diagnosis.

During the operation of the bridge health monitoring system, it needs consider the engineering design and sensor layout requirements. In the above bridge geometric model, the cross sections in the middle of the left and right spans should be selected as the detection points. The detection point can set 6 temperature measuring points, 52 strain measuring points and 10 dynamic monitoring points. The dynamic strain and displacement detection data are the core data for calculating the demand. In the deep learning analysis model, these data are used for the transient internal force information of the bridge. Exponential anomaly is usually generated when the bridge structure is subjected to abnormal load. At this time, the local response changes, thus affecting the overall state of the bridge. When the load exceeds the threshold, visual damage will occur and even lead to safety accidents [18]. Therefore, the response of the local state will be the key item of the global prediction. The displacement data collected in the actual operation is in the situation of high noise and high spikes. The reason is that the complex physical environment makes multiple factors accumulate, so it is necessary to deconstruct and reduce the dimension of the original data for feature extraction.

3.2 Bridge data mapping and damage assessment model construction based on deep learning

The physical environment and data acquisition logic of bridge structure damage assessment have been clarified above, but the complexity of the real scene makes the transmitted data unable to be directly used for detection. In view of this, the research will build a data deconstruction method to process the data. Firstly, wavelet analysis is used to decompose the time domain information contained in the original data to enhance the resolution of the data [19]. In the analysis, it is assumed that function $f(x)$ satisfies the condition of Eq. (2).

$\displaystyle\left\{{\begin{array}[]{l}\int_{-\infty}^{\infty}{|{f(x)}|^{2}dx<% \infty}\\ f(x)\in L^{2}(R)\\ \end{array}}\right.$ (2)

$d x$ is the frequency variable of high frequency band in Eq. (2). $L^{2}(R)$ is a square integral real number space. $R$ indicates the time-domain information contained by the original data. At this time, the defined function $g(x)$ also satisfies $L^{2}(R)$ , and the function after Fourier transform satisfies the linear decomposition operation of Eq. (3).

$\displaystyle C_{g}=\frac{\int_{R}{|{{g}^{\prime}(x)}|^{2}}}{|{x}^{\prime}|}d{% x}^{\prime}$ (3)

$C_{g}$ is a linear decomposition operation in Eq. (3). The original function $g(x)$ is the base wavelet, and the definition of the resulting wavelet family C is shown in Eq. (4).

$\displaystyle g_{jk}(x)=2^{\frac{j}{2}}g({2^{j}x-k}),j,k\in Z$ (4)

$Z$ of Eq. (4) is an integer set. $k$ represent the different members of the wavelet family. If the wavelet family generated by the base wavelet satisfies Eq. (5), it is called orthogonal wavelet.

$\displaystyle\langle{g_{jk},g_{lm}}\rangle=\delta_{j,{j}^{\prime}}\delta_{k,{k% }^{\prime}}$ (5)

$\delta$ in Eq. (5) is the frequency variation. At this time, any function can be converted as shown in Eq. (6).

$\displaystyle f(x)=\sum\nolimits_{j,k=-\infty}^{\infty}{C_{j,k}}g_{jk}(x)$ (6)

After the initial signal is decomposed by wavelet, the high frequency and low frequency results of the actual signal are obtained. This will continue to decompose the low-frequency signal by wavelet, and then decompose the low-frequency signal repeatedly. If the faher wavelet is $\mu_{0}^{0}(t)$ and the mother wavelet is $\mu_{1}^{0}(t)$ , there is a function $\mu_{n}$ as shown in Eq. (7).

$\displaystyle\left\{{\begin{array}[]{l}\mu_{2\alpha}(x)=\sum\nolimits_{k}{S_{k% }\mu_{\alpha}({2x-k})}\\ \mu_{2\alpha+1}(x)=\sum\nolimits_{k}{T_{k}\mu_{\alpha+1}({2x-k})}\\ \end{array}}\right.$ (7)

In Eq. (7), $\alpha$ represents a non-negative integer. $S_{k}$ and $T_{k}$ are low-pass and high-pass filters. At this time, the wavelet packet is decomposed and reconstructed according to Eq. (8).

$\displaystyle\left\{{\begin{array}[]{l}d_{k}^{2n}=\frac{1}{2}\sum\nolimits_{L}% {d_{L}^{n}S_{L-2k}}\\ d_{k}^{2n+1}=\frac{1}{2}\sum\nolimits_{L}{d_{L}^{n}T_{L-2k}}\\ d_{k}^{n}=\sum\nolimits_{L}{d_{\alpha}^{2n}S_{k-2\alpha}\sum\nolimits_{L}{d_{% \alpha}^{2n+1}S_{k-2\alpha}}}\\ \end{array}}\right.$ (8)

$L$ is the length of each stage of the wave signal in Eq. (8). $n$ denotes constant. The dynamic capture formed by acceleration integral during signal decomposition cannot extract the temperature-induced static behavior. Therefore, the low frequency region where the signal is located is processed. According to the above wavelet packet decomposition principle, the signal of available frequency band is reconstructed. And the decomposition scale should meet the requirement that the decomposition result is slightly larger than half of the first frequency band outside the available frequency band. The deformation of bridges is usually caused by many factors. To accurately evaluate the results, multidimensional data will be mapped. The relationship between the longitudinal displacement $y$ under load $F$ in structural mechanics and the action position is shown in Eq. (9).

$\displaystyle y=\sum\nolimits_{m=1}^{M}{\frac{F}{ul\omega_{m}}}\left[{\frac{2}% {({\omega_{m}^{2}-\beta_{m}^{2}})}\omega_{m}\sin\beta_{m}t-\beta_{m}\sin\omega% _{m}t}\right].\sin\frac{m\pi x}{l}$ (9)

$\omega_{m}$ is the $m$ -order natural frequency of constant section in Eq. (9). $u$ is the quality. $l$ is the length. $F$ represents the load in the structural mechanics. $\beta$ is defined in Eq. (10).

$\displaystyle\beta_{m}=\frac{m\pi v}{l}$ (10)

In the Eq. (10), $v$ represents the displacement velocity. In mechanics of materials, when a straight beam is bending, the normal stress of any cross section is calculated like Eq. (11).

$\displaystyle\gamma=\frac{Wy_{h}}{I}$ (11)

$I$ in Eq. (11) is the interface inertia moment. $W$ is the material bending moment under stress. $y_{h}$ is the height from the section edge to the neutral axis. Normal stress can be linearly mapped to strain, and interface bending moment can be linearly mapped to deflection. Therefore, there is a nonlinear mapping between the interface strain and displacement of the bridge. From the perspective of causal logic, the structural response of the bridge has a complex factor space. The strain caused by vehicle load can also cause complex structural changes, which makes it difficult to calculate by formula. The research will build the mapping relationship between vehicle load data and bridge structural parameters through deep learning. The advantage of this construction method is that it does not need to explore the field effect, but directly carries out end-to-end relationship mapping. Recurrent Neural Network (RNN) and Long Short-Term Memory (LSTM) are commonly used to add time series information to variable mapping. The generative effect of RNN has the possibility of gradient explosion and disappearance. And LSTM can compensate for this disadvantage in the form of unit state. Figure 3 shows the specific operation logic.

Figure 3.

LSTM operation structure diagram.

In Fig. 3, there are always three modules in the network structure of LSTM, namely the input layer, output layer, and hidden layer. The input layer mainly performs dimensional transformation on the data, the hidden layer extracts features from the processed results, and the output layer’s data is the predicted results of the entire model. The specific operating logic of LSTM is to first filter the useless parts of the input information through input gates, update some data, and keep the rest unchanged. When the input gate determines the information and stores it, it passes it to the forgetting function, and then discards the state that needs to be forgotten through the forgetting gate to avoid gradient explosion. Then perform a sigmoid operation on the information from the previous moment of the hidden layer, determine whether to forget it, and finally update the cell state [20]. Adaptively construct the bridge data according to the principle of LSTM. The overall network has 6 layers, and the structure is shown in Fig. 4.

Figure 4.

LSTM training structure based on bridge damage data.

Figure 5.

LSTM multi-scale mapping process.

In addition to the response layer of the control module, the two full connection layers of LSTM are all 1 except the first input dimension of 50 in Fig. 4. Data preprocessing will be carried out between the response layer and the LSTM output layer, which will be normalized according to Eq. (12).

$\displaystyle x_{i}=\frac{X_{i}-\bar{X}}{\lambda}$ (12)

$\lambda$ is the standard deviation of the sample set in Eq. (12). $X_{i}$ is the original data. $\bar{X}$ is the mean value of the sample set. The representation during training is reflected by the loss function, and the research will be calculated by the mean square error of Eq. (13).

$\displaystyle\textit{Loss}=\frac{1}{2}\sum\nolimits_{j=1}^{\Omega}{({y_{j}-{y}% ^{\prime}_{j}})}^{2}$ (13)

$y_{j}$ and ${y}^{\prime}_{j}$ are actual and expected corresponding values in Eq. (13). $\Omega$ is the total number of responses generated. After the model training is completed, it needs to be adjusted with the help of the verification set. The research will use the root mean square error ratio of Eq. (14) as the verification basis.

$\displaystyle\text{Re}=\sqrt{\frac{1}{N}\sum\nolimits_{k-1}^{N}{({y_{k}-r_{k}}% )}}$ (14)

$N$ is the capacity of the model verification set in Eq. (14). $y_{k}$ and $r_{k}$ are the corresponding actual value and model regression value of $k$ . The amplitude will be defined as shown in Eq. (15).

$\displaystyle U=\frac{\text{Re}}{A}$ (15)

$A$ in Eq. (15) is the sequence amplitude of the target data. In the actual test, the average value of $U$ is less than expected, and all sequences $U<0.05$ . The error level is within the allowable range. At this time, the model construction and adjustment will be output as the final result. The global logic is shown in Fig. 5.

The LSTM model will be judged manually after training in Fig. 5. It needs validate the performance of the network model in multiple ways to ensure the fitting degree. If any link is unqualified, multiple parameters will be adjusted, including the number of layers, learning rate, forgetting rate and training set size of each module. Since the model will carry out multi-channel and multi-dimensional nonlinear mapping, the performance of the network model needs to be repeatedly verified.

4. Performance analysis of bridge damage identification model based on deep learning

4.1 Data processing and training analysis

In order to verify the performance advantages of LSTM model, traditional recurrent neural networks, RNN, Deep Belief Networks (DBN), and Gated Recurrent Unit (GRU) are selected for the study, and error results are compared with them. At the same time, because hyper parameter have a great impact on the prediction accuracy of the model, Therefore, through multiple cross validation, the best super parameter value was selected for comparative experiments. The specific comparison results are shown in the Fig. 6.

Figure 6.

Comparison of errors between different models.

As shown in the figure, the error indicators of the LSTM model are 4.86%, 2.51%, and 0.62%, respectively. Compared to the RNN model, it has decreased by 10.63%, 9.84%, and 2.66%, respectively. Compared to the DBN model, it decreased by 2.98%, 4.43%, and 1.89%, respectively. Compared to the GRU model, it decreased by 4.15%, 4.8%, and 2.25%, respectively. The LSTM model has higher prediction accuracy compared to other models. Bridge structural damage involves load data processing and structural data processing. Because of the complex logical relationship between the two data sets, the deep learning of the research structure abandons the field effect and directly maps the relationship between the data. The performance test of the deep learning bridge damage identification model includes three aspects. That is to identify the original data structure of the sensor, the mapping relationship of the model, and evaluate the damage magnitude of the bridge structure. During data deconstruction, the temperature-induced static effect is eliminated, and the results are shown in Fig. 7.

Figure 7.

Exclusion of temperature effect by wavelet packet decomposition.

Figure 7a shows the original data column measured by the sensor. The strain of the data fluctuates and cannot be directly analyzed. Figure 7b shows the effect of wavelet packet decomposition. In the actual scenario, the factors that cause the bridge structure strain include vehicle-induced factors and temperature-induced effects. When the temperature effect is removed from the data set, the deformation caused by pure vehicle load will remain, which will eliminate the interference for subsequent mapping analysis. The ambient temperature cannot be corrected during preventive application. Therefore, it is necessary to exclude this part as random fixed effects, which is helpful for physical regulation. The premise of nonlinear mapping is that there is a relationship between the data at both ends. Finding and proving the existence of the relationship will be a prerequisite for LSTM. The research will enlarge part of the vehicle-induced variation data to obtain the content shown in Fig. 8.

Figure 8.

Mapping relational datasets and local enlargement.

Figure 8a shows the measured vehicle-induced strain data set. According to the above wavelet packet decomposition, it is locally amplified to get Fig. 8b. The existence of fluctuations in the partial enlarged view belongs to the category of easy analysis. Therefore, there is a direct correspondence between the data at both ends of the vehicle-induced strain. And the collected data will be trained to form a one-to-one non-linear mapping scenario. At this time, the loss convergence process of the model is shown in Fig. 9.

Figure 9.

Loss convergence of single-to-single nonlinear mapping models.

The loss at the beginning of the model is more than 0.2, and the other loss values are less than 0.2 in Fig. 9. And the loss value changes regularly and steadily. This shows that the convergence performance of the loss function is good and the error can be controlled.

4.2 Performance analysis of the bridge damage prediction model

The above experiments prove the rationality of the model construction and the feasibility of optimization. At this time, the collected data set will be divided into training set and verification set test set to fit the damage prediction of the bridge. Figure 10 shows the results.

Figure 10.

Long-term signal trend prediction fitting.

The predicted results are basically consistent with the actual results, with a fitting rate of 92.1% in Fig. 10. This shows that the performance of nonlinear mapping model is dominant in the global view and has certain positive significance for long-term monitoring of bridges. At this time, the model will be used for local feature prediction to verify the evaluation performance of instantaneous strain and response. Figure 11 shows the results.

Figure 11.

Local prediction of nonlinear mapping.

The actual value is basically consistent with the predicted value, and the fitting degree reaches 93.3% in Fig. 11. Under the condition of superior global fitting performance, the model can accurately predict the instantaneous strain value of the bridge. Therefore, the model can diagnose the transient fault. At this time, the regression displacement and the measured displacement are fitted. And the decision coefficient of the model is used to determine to explore the mining performance of the mapping relationship of the end test data set. Figure 12 shows the results.

Figure 12.

Map data scatter plot.

In the mapping results of Fig. 12, some data points are relatively scattered. Considering the structural utility between vehicles, the errors caused by these data are within a controllable range. In addition, the slope between the predicted data and the overall data is close to 1, the mapping relationship is relatively accurate, and the judgment result of the judgment coefficient is 0.94. The mapping relationship mining performance of the model is good, and the judgment results show that the model can judge the relationship between bridge structure and load accurately. Under the premise of good prediction performance of individual data, the model is used to simulate the stress of vehicles with four load categories. Figure 13 shows the results.

Figure 13.

Model stress prediction for vehicles with various loads.

Figure 13 shows the stress prediction of four load types of vehicles in the data set. The real data is calculated based on the physical field effect of vehicle driving. The stress predicted by the model for each type of vehicle is close to the real value. The existing individual offset data is caused by the structural impact between vehicles, with an average fitting rate of 89.72%. This shows that the deep learning neural network can directly simulate the mapping relationship from the end to the side.

5. Conclusion

With the development of logistics industry and the popularity of vehicles, the increase of load makes the bridge safety face new challenges. To evaluate the damage of the bridge structure, the research takes the neural network commonly used in fault diagnosis as the main framework for diagnosis. According to the need of in-depth learning input data, the research first constructs the physical model of the load and the spatial model of the vehicle type of the bridge, and then digitizes the parameters at both ends. To solve the complex characteristics of the original data, wavelet packet decomposition is adopted to eliminate the temperature-induced effects in the data. According to the complex mapping between the vehicle-induced strain data and the bridge displacement, the research will abandon the field effect analysis and adopt LSTM to directly map the nonlinear relationship of the data at both ends. And the judgment coefficient is also constructed to test the performance. The results showed that the error indicators of the LSTM model were 4.86%, 2.51%, and 0.62%, respectively, which were reduced by 10.63%, 9.84%, and 2.66% compared to the RNN model. And the method can remove the temperature-induced effect of the original data and obtain qualified vehicle-induced load data, and partially enlarge the data. The enlarged data can be used as the input data for training. The loss function of deep learning for bridge structure damage prediction is not more than 0.2. The global fitting rate of the model is 92.1%, and the local fitting rate is 93.3%. This shows that the bridge structure damage diagnosis model proposed in the study has positive significance, and the damage assessment of multi-channel bridges will be studied in the follow-up study. Current studies focus on the analysis of single-channel data, while actual bridge structures often have multiple channels, such as sensor data in different locations and directions. A more comprehensive assessment of the bridge’s structural damage is possible by considering data from multiple channels simultaneously. So future studies will consider multi-channel bridge damage assessment.

References

Allen

Arkolakis

. The welfare effects of transportation infrastructure improvements. Rev Econ Stud. 2022; 89(6): 2911-2957.

Hajializadeh

. Deep learning-based indirect bridge damage identification system. Struct Health Monit. 2023; 22(2): 897-912.

Salet

TAM

Ahmed

Bos

Laagland

HLM

. Design of a 3D printed concrete bridge by testing. Virtual Phys Prototy. 2018; 13(3): 222-236.

Reagan

Sabato

Niezrecki

. Feasibility of using digital image correlation for unmanned aerial vehicle structural health monitoring of bridges. Struct Health Monit. 2018; 17(5): 1056-1072.

Liang

. Image-based post-disaster inspection of reinforced concrete bridge systems using deep learning with Bayesian optimization. Comput-Aided Civ Inf. 2019; 34(5): 415-430.

Omer

Margetts

Mosleh

Hewitt

Parwaiz

. Use of gaming technology to bring bridge inspection to the office. Struct Infrastruct E. 2019; 15(10): 1292-1307.

Xie

Yao

Liu

Xia

Zeng

. Automatic multi-image stitching for concrete bridge inspection by combining point and line features. Autom Constr. 2018; 90: 265-280.

Saleem

Park

Lee

Jung

Sarwar

. Instant bridge visual inspection using an unmanned aerial vehicle by image capturing and geo-tagging system and deep convolutional neural network. Struct Health Monit. 2021; 20(4): 1760-1777.

Jin

Yun

. A review on deep learning-based structural health monitoring of civil infrastructures. Smart Struct Syst. 2019; 24(5): 567-585.

10.

Nguyen

. A climbing robot for steel bridge inspection. J Intell Robot Syst. 2021; 102: 75.

11.

Zhang

Sun

. Structural damage identification via physics-guided machine learning: A methodology integrating pattern recognition with finite element model updating. Struct Health Monit. 2021; 20(4): 1675-1688.

12.

Wang

Pan

Yuan

Yang

Gui

. A novel deep learning based fault diagnosis approach for chemical process with extended deep belief network. ISA T. 2020; 96: 457-467.

13.

Zhao

Zhong

Tang

Pecht

. Deep residual shrinkage networks for fault diagnosis. IEEE T Ind Inform. 2019; 16(7): 4681-4690.

14.

Saufi

Ahmad

ZAB

Leong

Lim

. Gearbox fault diagnosis using a deep learning model with limited data sample. IEEE T Ind Inform. 2020; 16(10): 6263-6271.

15.

Zhao

Zhang

Liu

Sun

Chen

. Deep-learning-based open set fault diagnosis by extreme value theory. IEEE T Ind Inform. 2021; 18(1): 185-196.

16.

Fan

Chen

Sun

Feng

Hassanin

Sareh

. Machine learning applied to the design and inspection of reinforced concrete bridges: Resilient methods and emerging applications. Structures. 2021; 33: 3954-3963.

17.

Zhang

Chen

. A framework of structural damage detection for civil structures using a combined multi-scale convolutional neural network and echo state network. Eng Comput-Germany. 2023; 39(3): 1771-1789.

18.

Mei

Gül

Boay

. Indirect health monitoring of bridges using Mel-frequency cepstral coefficients and principal component analysis. Mech Syst Signal. 2019; 119: 523-546.

19.

Yang

Zhang

Chen

Wang

Jiang

Zeng

. A hierarchical deep convolutional neural network and gated recurrent unit framework for structural damage detection. Inform Sciences. 2020; 540: 117-130.

20.

Liu

. Nonlinear unsteady bridge aerodynamics: Reduced-order modeling based on deep LSTM networks. J Wind Eng Ind Aerod. 2020; 198: 104116.