A bridge health monitoring method and system based on deep learning

Abstract

The present invention discloses a bridge health monitoring method and system based on deep learning, including: reading the signal on the bridge structure and performing signal preprocessing; obtaining wavelet coefficients after wavelet decomposition, denoising the signal through a threshold function, and reconstructing the signal through the wavelet coefficients; building and training a neural network through a deep learning framework to obtain a deep learning model, wherein the neural network includes a TFNet damage prediction model and an MCSANet damage classification model; optimizing the calculation graph of the deep learning model; deploying the deep learning model after the calculation graph optimization to perform bridge structure health monitoring. The present invention solves the problem that the automation, real-time performance, and accuracy of bridge health monitoring need to be enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of bridge structure health monitoring, and in particular to a bridge health monitoring method and system based on deep learning. Background Art

[0002] Health monitoring and damage identification of bridges are important technical means to ensure the safety of bridge structures and extend their service life. Traditional monitoring methods often rely on manual inspections and simple sensor data analysis, which have problems such as strong subjectivity and low efficiency. Structural health monitoring (SHM) is a very broad research field, but most SHM applications have been focused on damage detection. Traditionally, structural system pattern recognition technology has been widely used to identify damage by utilizing changes in the vibration characteristics of the structure. Judging from the current status of structural damage diagnosis research, it is still a very difficult task to perform reasonable and accurate damage diagnosis on large civil engineering structures. It is still in the theoretical exploration stage and requires further cross-disciplinary development. At present, there are several deficiencies in the field of structural damage diagnosis and comprehensive early warning: (1) Insufficient sensitivity of damage indicators: Existing technologies have limitations in extracting indicators that can accurately reflect the damage status of structures, resulting in insufficient accuracy and reliability of diagnostic results; (2) Limitations of signal processing methods: Traditional time domain and frequency domain analysis methods cannot effectively process nonlinear and non-stationary signals, and it is difficult to fully capture the complex characteristics of structural vibration signals; (3) Lack of accurate computational models: In the field of damage diagnosis, there is insufficient research on computational models for various damage conditions, which affects the accuracy and efficiency of damage diagnosis; (4) Challenges in data preprocessing: Data acquisition, fusion, purification and feature extraction are subject to interference from complex external factors, and existing technologies are difficult to effectively eliminate these inaccuracies; (5) Limited degree of intelligence: Although theoretically feasible, existing methods have practical limitations in realizing a fully automated damage diagnosis process, including data acquisition, preprocessing, feature extraction, etc.; (6) Difficulty in multi-indicator fusion evaluation: The indicators generated by different data sources and damage diagnosis methods vary greatly, and how to integrate these indicators to form a comprehensive judgment remains an unresolved problem. Summary of the invention

[0003] 1. Technical issues to be resolved

[0004] Based on the above problems, the present invention provides a bridge health monitoring method and system based on deep learning to solve the problem that the automation, real-time performance and accuracy of bridge health monitoring need to be enhanced.

[0005] (II) Technical solution

[0006] Based on the above technical problems, the present invention provides a bridge health monitoring method based on deep learning, comprising:

[0007] S1, reading monitoring signals on the bridge structure and performing signal preprocessing, wherein the monitoring signals include tilt signals, settlement signals and offset signals;

[0008] S2, after wavelet decomposition, wavelet coefficients are obtained, the signal is denoised by a threshold function, and then the signal is reconstructed by the wavelet coefficients;

[0009] S3. Building and training a neural network through a deep learning framework to obtain a deep learning model, wherein the neural network includes a TFNet damage prediction model and an MCSANet damage classification model;

[0010] S4. Optimizing the computational graph of the deep learning model;

[0011] S5. Deploy the deep learning model after computational graph optimization to perform bridge structure health monitoring.

[0012] Further, in S2, the threshold function is:

[0013]

[0014] In the formula, η represents the threshold, q j，k and q j，k 'represent the signal data before and after denoising, δ and ε are constants.

[0015] Furthermore, in S3, the TFNet damage prediction model includes: input data enters the linear layer, and then enters multiple stacked TFBlock modules before output, the linear layer includes MatMul fusion matrix and Add addition; one path of the TFBlock module extracts features on the variable dimension through the linear layer, and the other path first extracts features on the time dimension through the LSTM long short-term memory model, and then transforms the features extracted on the time dimension to the variable dimension through the Transpose transfer matrix, and finally fuses the two features through the MatMul fusion matrix to fuse the information on the time dimension and the variable dimension, and after the Tanh activation function, superimposes the initial value input to the TFBlock module to obtain the final output value.

[0016] Furthermore, in S3, the MCSANet prediction model includes: after the input data passes through multiple CSABlock modules, it passes through the Conv1d one-dimensional convolution layer, the MatMul fusion matrix, and the SoftMax activation function and is then output. The CSABlock module has two parallel convolutional self-attention blocks, and both convolutional self-attention blocks include a Conv1d one-dimensional convolution layer, a BatchNormal normalization layer, and a Relu activation function. Then, one convolutional self-attention block passes through average pooling AvgPool, a MatMul fusion matrix, a second Relu activation function, a MatMul fusion matrix, and a Sigmoid activation function, and is fused with the output of the Relu activation function before being output. Another convolutional self-attention block passes through Mean and Max processing in parallel and is merged through Concat. After passing through the Conv1d one-dimensional convolution layer and the Sigmoid activation function, it is fused with the output of the Relu activation function before being output.

[0017] Furthermore, in S3, the deep learning framework is Pytorch.

[0018] Furthermore, S4 includes: converting the trained pytorch model into ONNX.

[0019] The present invention also discloses a bridge health monitoring system for embedded systems based on deep learning, comprising:

[0020] at least one processor; and at least one memory in communication with the processor, wherein:

[0021] The memory stores program instructions that can be executed by the processor, and the processor can execute the method by calling the program instructions.

[0022] The present invention also discloses a non-transitory computer-readable storage medium, wherein the non-transitory computer-readable storage medium stores computer instructions, wherein the computer instructions enable the computer to execute the method.

[0023] (III) Beneficial effects

[0024] The above technical solution of the present invention has the following advantages:

[0025] (1) The present invention can automatically, in real time, and comprehensively monitor and analyze the health status of bridges through the bridge structure prediction model and the damage classification model. The deep learning model learns the deep features in the data, thereby realizing intelligent identification and early warning of bridge damage, greatly improving the efficiency and accuracy of monitoring. In addition, the TFNet damage prediction model and MCSANet damage classification model used enhance the feature extraction ability of the model by utilizing the feature transfer matrix and spatiotemporal self-attention mechanism, thereby improving the prediction and classification performance of the model, significantly improving the accuracy of intelligent identification of bridge structure damage, and ensuring good reasoning results for data from different scenarios;

[0026] (2) The present invention uses a lightweight, modular and stackable TFNet damage prediction model and MCSANet damage classification model, which greatly reduces the system resource requirements during model reasoning. By optimizing the ONNX model calculation graph and designing a global shared static memory pool, the model reasoning performance can be further guaranteed while minimizing the computing resource usage, thereby meeting the requirements of real-time monitoring, real-time analysis and early warning of data.

[0027] (3) The present invention preprocesses the signal through wavelet transform, effectively distinguishes the signal and noise characteristics, improves the reconstruction quality of the signal, and improves the threshold function and the hierarchical threshold selection rule method, selects the corresponding threshold according to the signal characteristics, thereby improving the effectiveness of filtering and reducing the wavelet error; even when the noise level increases, the algorithm still maintains the best damage identification effect, the identification accuracy changes little, shows a strong ability to resist noise interference, and exhibits good stability and robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] The features and advantages of the present invention will be more clearly understood by referring to the accompanying drawings, which are schematic and should not be construed as limiting the present invention in any way. In the accompanying drawings:

[0029] Figure 1 A schematic diagram of a flow chart of a bridge health monitoring method based on deep learning according to an embodiment of the present invention;

[0030] Figure 2 This is a structural diagram of the TFNet damage prediction model of an embodiment of the present invention;

[0031] Figure 3 This is a structural diagram of the MCSANet damage classification model of an embodiment of the present invention;

[0032] Figure 4 This is a comparison chart before and after the ONNX operator optimization of the LSTM layer in the TFNet model of an embodiment of the present invention;

[0033] Figure 5A memory usage calculation graph of an embodiment of the present invention;

[0034] Figure 6 A comparison diagram of model loss curves under different prediction step sizes in an embodiment of the present invention;

[0035] Figure 7 A comparison chart of model parameter quantities and floating-point calculation quantities in an embodiment of the present invention;

[0036] Figure 8 This is a comparison chart of the average precision and recall rate of the model of an embodiment of the present invention on a multi-classification dataset. DETAILED DESCRIPTION

[0037] The specific implementation of the present invention is further described in detail below in conjunction with the accompanying drawings and examples. The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0038] The embodiment of the present invention is a bridge health monitoring method based on deep learning, such as Figure 1 As shown, the following steps are included:

[0039] S1, read the monitoring signal on the bridge structure and perform signal preprocessing;

[0040] Read the monitoring signals on the bridge structure, and obtain monitoring data such as the function, response, change, and environment of the bridge structure from various intelligent sensors installed on the bridge, including tilt signals, settlement signals, and offset signals.

[0041] The signal is segmented and refined to achieve the purpose of subdividing the frequency at low frequency and the time at high frequency, improving the ability to distinguish signal and noise characteristics and realizing bridge data reconstruction.

[0042] Assuming that the actual bridge structure damage data signal is r(a), the signal length is S, the noisy signal is u(a), the noise data is o(a), and it obeys the normal distribution of S(0, 1), then the expression formula of the noisy signal is as follows:

[0043] u(a)=r(a)+αo(a)

[0044] α represents the standard deviation of the noise coefficient, which is used to describe the noise quality in the signal. After obtaining the noisy bridge structure damage signal, it is necessary to decompose it by wavelet transform, and realize the spatial transformation of the signal by decomposing the noisy signal into multiple scales.

[0045] S2, after wavelet decomposition, wavelet coefficients are obtained, the signal is denoised by a threshold function, and then the signal is reconstructed by the wavelet coefficients;

[0046] Construct a threshold function, get the wavelet coefficient component after wavelet decomposition, denoise the signal by constructing the threshold function, and select the corresponding threshold according to the signal characteristics by improving the threshold function and selecting the threshold rule in a hierarchical manner to improve the effectiveness of the threshold and reduce the wavelet error.

[0047] Traditional threshold functions include hard threshold functions and soft threshold functions. The threshold in the hard threshold function is discontinuous, which is easy to retain irregular oscillation characteristics during the denoising process, affecting the smoothness of the reconstructed signal; the soft threshold function has a constant deviation problem, and the reconstructed signal is prone to errors and quality problems, affecting the denoising effect. To this end, the study combines the advantages of the traditional threshold function and improves the threshold function. Its expression is as follows:

[0048]

[0049] η represents the threshold, q j，k and q j，k 'represent the signal data before and after denoising, δ and ε are constants used to adjust the denoising error.

[0050] After the correlation denoising process, the estimated wavelet coefficients are obtained. On this basis, the reverse operation is performed to realize the wavelet coefficient reconstruction, complete the separation of pure signals and noise signals, realize the extraction of pure signals, avoid data errors and other problems, and improve the accuracy of intelligent identification of bridge structure damage.

[0051] S3. Build and train a neural network through the deep learning framework Pytorch to obtain a deep learning model, wherein the neural network includes a TFNet damage prediction model and an MCSANet damage classification model;

[0052] The floating-point computing power and memory resources of embedded platforms and edge devices are very limited. In order to enable deep learning models to run stably on them, how to reduce the computational complexity of the model is the primary issue that needs to be considered when designing the model structure. In order to achieve the best inference performance under extreme hardware limitations, the design uses lightweight, modular and stackable TFNet damage prediction model and MCSANet damage classification model.

[0053] TFNet model such as Figure 2As shown in the figure, the design idea is as follows: the model consists of a linear layer and multiple stacked TFBlock modules. The linear layer inputs the data into the TFNet model for initial prediction, that is, a rough estimate of the predicted data is made, and then it is input into TFBlock for accurate prediction. The idea of ​​boosting algorithm similar to ensemble learning is adopted to gradually improve the prediction performance by stacking multiple TFBlock modules. As the core part of the TFNet model, the main working principle of the TFBlock module is as follows: LSTM is used to extract features in the time dimension, and the linear layer (MatMul+Add) is used to extract features in the variable dimension. The features extracted in the time dimension are transformed to the variable dimension through the transfer matrix. Finally, the information in the time dimension and the variable dimension are fused through the fusion matrix and the initial value input to the TFBlock module is superimposed to obtain the final output value. The essence of the TFNBlock module is to predict the residual value between the input value and the target value by fusing the spatiotemporal features of the sequence. Therefore, by stacking multiple TFBlock modules, the residual value between the predicted value and the target value is gradually reduced, and finally the ideal prediction output is achieved.

[0054] The MCSANet model is as follows Figure 3As shown in the figure, the idea is as follows: the main part of the model consists of CSABlock modules, each of which has two parallel convolutional self-attention blocks. The convolutional self-attention block on one side consists of a convolutional block plus a channel self-attention block, and the convolutional self-attention module on the other side consists of a convolutional block plus a spatial self-attention block. The convolutional blocks on both sides contain a Conv1d one-dimensional convolutional layer, a BatchNormal normalization layer, and a Relu activation function. The difference between the convolutional blocks on both sides is that the convolutional kernels have different sizes, and different convolutional kernels of different sizes are used to extract features of different frequencies from the data. Different channels and different feature space regions in the input feature map (data) have different effects on subsequent layers, so the channel self-attention module and the spatial self-attention module are used to adaptively recalibrate the feature map from two different dimensions. This adaptive rescaling is crucial to improving the performance of the model. Therefore, specifically, the MCSANet prediction model includes: after the input data passes through multiple CSABlock modules, it passes through the Conv1d one-dimensional convolution layer, the MatMul fusion matrix, and the SoftMax activation function before output. The CSABlock module has two parallel convolutional self-attention blocks, and both convolutional self-attention blocks include the Conv1d one-dimensional convolution layer, the BatchNormal normalization layer, and the Relu activation function. Then, one convolutional self-attention block passes through the average pooling AvgPool, the MatMul fusion matrix, the second Relu activation function, the MatMul fusion matrix, and the Sigmoid activation function, and then is fused with the output of the Relu activation function and output. The other convolutional self-attention block passes through the Mean mean and Max maximum value processing in parallel and is merged through Concat. After passing through the Conv1d one-dimensional convolution layer and the Sigmoid activation function, it is fused with the output of the Relu activation function and output.

[0055] Pytorch is a popular deep learning framework that provides complete tools and libraries from data preprocessing to model training in deep learning tasks. The TFNet damage prediction model and the MCSANet classification model both use Pytorch for model building and training. During the training process, TFNet uses mean squared error (MSE) as the loss evaluation function, and the optimizer algorithm uses stochastic gradient descent (SGD). CSSANet uses cross entropy (Cross Entropy) as the loss function, and the optimizer algorithm uses the Adam (Adaptive Moment Estimation) adaptive learning rate adjustment algorithm. Both use the ReduceLROnPlateau learning rate adjustment strategy, which automatically reduces the learning rate when the validation set loss continues to stop decreasing.

[0056] S4. Optimizing the computational graph of the deep learning model;

[0057] The computation graph is a directed acyclic graph that represents the computation of each layer in the model as computation nodes and connects them together in sequence. The computation nodes are each rectangular box in the graph, and the computation order of the computation graph is indicated by the arrows in the graph.

[0058] In order to facilitate the deployment and optimization of the model, the pytorch model obtained through the above training needs to be converted into an ONNX (Open Neural Network Exchange) model. Due to the fragmentation of the operators in the ONNX model itself and the invalid glue operators generated during the conversion process, the efficiency of model reasoning will be reduced. In order to accelerate reasoning deployment, the automatic optimization of the calculation graph is achieved through preset optimization templates such as operator fusion, deletion, and replacement, which accelerates the operation reasoning of the model calculation graph, including:

[0059] Traverse each computing node in the computational graph, and then determine whether there is an optimization method for the computing node. If so, select the optimized node or skip the node based on the existing optimization method; the optimization methods include redundant node elimination, constant node folding, and computing node fusion, etc. Therefore, whether each node is optimized requires the prior calculation graph optimization experience setting, and then the specific optimization process is written as code.

[0060] like Figure 4 Figure (a) shows the schematic diagram of the ONNX operator of the LSTM layer in the TFNet model before optimization. Figure 4Figure (b) shows the schematic diagram of the optimized ONNX operator of the LSTM layer in the TFNet model. The operators related to dimensional calculation in the calculation graph are deleted, that is, shape, Gather, Unsqueeze, Concat and ConstantOfShape are deleted, which speeds up the operation of the calculation graph by reducing memory access and calculation amount.

[0061] S5. Deploy the deep learning model optimized by the computational graph, perform bridge structure health monitoring based on the reconstructed signal, and output the bridge damage diagnosis results and damage types;

[0062] After completing the calculation graph optimization of the ONNX model, the final model packaging and deployment can be started. The TFNet damage prediction model and the MCSANet damage classification model are used to monitor the health of the bridge structure signal, predict whether it is damaged, and determine the damage type.

[0063] In this embodiment, the conversion process from the ONNX model to the deployment model is to parse the structure and parameters of the original ONNX, complete the C language code implementation of the operator in the model calculation graph, design and simplify the ONNX model serialization storage data structure, store the parameters in the ONNX model in files according to the specified storage logic, and finally complete the automatic assembly and generation of the C deployment code through an automated script.

[0064] In addition, on embedded platforms and edge devices with limited memory resources, how to reduce the memory usage during model inference has become a key issue for whether the model can be deployed on the corresponding platform. In order to further reduce the memory usage during model inference, a global shared static memory scheduling algorithm is adopted. The input and output data of each operator in the calculation graph are assigned a correct offset address in the shared static memory through the memory scheduling algorithm. At the same time, the size of the shared static memory pool is equivalent to the memory usage of a computing node with the largest occupancy in the calculation graph. The process of calculating the maximum memory usage is briefly explained through a test calculation graph. In the model inference scenario, the storage size of the model parameters and input data is fixed. Therefore, the memory usage of the input and output data of any computing node in the calculation graph during the model inference process is known. To describe the two dimensions of data retention time and memory usage during the inference process, we abstract the data in two dimensions according to time and memory size, such as Figure 5 As shown in the figure, the horizontal axis represents memory consumption, the vertical axis represents time nodes, and the upper left corner corresponds to the coordinate origin. Each rectangular color block corresponds to an input and output data, and the number represents the memory usage of the data. Through calculation, it can be known that the maximum memory usage of the test calculation graph = 64+36+16+8.

[0065] This embodiment uses the operating system Ubuntu 22.04, the CPU model Core (TM) i5-10400Inter (R), and the memory 32G. Actual bridge monitoring data and damage classification data are selected for model performance analysis. In the two data sets, 60% are randomly selected as test sets, 20% as validation sets, and 20% as training sets. The mean square error MSE (Mean Squared Error) and the mean absolute percentage error MAPE (Mean Absolute Percentage Error) are used as evaluation indicators for the TFNet damage prediction model, and the MCSANet model uses precision and recall as evaluation indicators. Figure 6 As shown, Figure 6 Figure (a) shows the comparison of the MSE loss curves of the models under different prediction step sizes. Figure 6 Figure (b) shows the comparison of the MAPE loss curves of the models under different prediction step sizes. TFNet has very low MSE and MAPE under different prediction step sizes. Figure 7 As shown, Figure 7 Figure (a) shows the comparison of model parameters. Figure 7 Figure (b) shows the comparison of floating-point calculations. MCSANet has the smallest number of parameters and floating-point calculations. Figure 8 As shown, Figure 8 Figure (a) shows the average accuracy comparison of the model in multi-classification datasets. Figure 8 Figure (b) shows the average accuracy comparison of the model in the multi-classification data set. The MCSANet model has the highest accuracy and recall rate while maintaining the lowest number of parameters and computation. Therefore, the TFNet model and the MCSANet model of this embodiment ensure the reasoning performance while minimizing the computing resource usage.

[0066] Finally, it should be noted that the above-mentioned control method can be converted into software program instructions, which can be implemented by using a control system including a processor and a memory, or by computer instructions stored in a non-transitory computer-readable storage medium. The above-mentioned integrated unit implemented in the form of a software functional unit can be stored in a computer-readable storage medium. The above-mentioned software functional unit is stored in a storage medium, including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) or a processor (processor) to perform some steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (Read-Only Memory, ROM), random access memory (Random Access Memory, RAM), disk or optical disk and other media that can store program codes.

[0067] In summary, the above-mentioned bridge health monitoring method and system based on deep learning have the following beneficial effects:

[0068] (1) The present invention can automatically, in real time, and comprehensively monitor and analyze the health status of bridges through the bridge structure prediction model and the damage classification model. The deep learning model learns the deep features in the data, thereby realizing intelligent identification and early warning of bridge damage, greatly improving the efficiency and accuracy of monitoring. In addition, the TFNet damage prediction model and MCSANet damage classification model used enhance the feature extraction ability of the model by utilizing the feature transfer matrix and spatiotemporal self-attention mechanism, thereby improving the prediction and classification performance of the model, significantly improving the accuracy of intelligent identification of bridge structure damage, and ensuring good reasoning results for data from different scenarios;

[0069] (2) The present invention uses a lightweight, modular and stackable TFNet damage prediction model and MCSANet damage classification model, which greatly reduces the system resource requirements during model reasoning. By optimizing the ONNX model calculation graph and designing a global shared static memory pool, the model reasoning performance can be further guaranteed while minimizing the computing resource usage, thereby meeting the requirements of real-time monitoring, real-time analysis and early warning of data.

[0070] (3) The present invention preprocesses the signal through wavelet transform, effectively distinguishes the signal and noise characteristics, improves the reconstruction quality of the signal, and improves the threshold function and the hierarchical threshold selection rule method, selects the corresponding threshold according to the signal characteristics, thereby improving the effectiveness of filtering and reducing the wavelet error; even when the noise level increases, the algorithm still maintains the best damage identification effect, the identification accuracy changes little, shows a strong ability to resist noise interference, and exhibits good stability and robustness.

[0071] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the embodiments of the present invention are described in conjunction with the accompanying drawings, those skilled in the art may make various modifications and variations without departing from the spirit and scope of the present invention, and such modifications and variations shall fall within the scope defined by the appended claims.

Claims

1. A bridge health monitoring method based on deep learning, characterized in that: include: S1, reading monitoring signals on the bridge structure and performing signal preprocessing, wherein the monitoring signals include tilt signals, settlement signals and offset signals; S2, after wavelet decomposition, wavelet coefficients are obtained, the signal is denoised by a threshold function, and then the signal is reconstructed by the wavelet coefficients; S3. Building and training a neural network through a deep learning framework to obtain a deep learning model, wherein the neural network includes a TFNet damage prediction model and an MCSANet damage classification model; The TFNet damage prediction model includes: input data enters the linear layer, and then enters multiple stacked TFBlock modules for output, the linear layer includes MatMul fusion matrix and Add addition; one path of the TFBlock module extracts features on the variable dimension through the linear layer, and the other path first extracts features on the time dimension through the LSTM long short-term memory model, and then transforms the features extracted on the time dimension to the variable dimension through the Transpose transfer matrix, and finally fuses the two features through the MatMul fusion matrix, fuses the information on the time dimension and the variable dimension, and after the Tanh activation function, superimposes the initial value input to the TFBlock module to obtain the final output value; the MCSANet prediction model includes: after the input data passes through multiple CSABlock modules, it passes through the Conv1d one-dimensional Convolution layer, MatMul fusion matrix, SoftMax activation function and then output, the CSABlock module has two parallel convolution self-attention blocks, both of which include Conv1d one-dimensional convolution layer, BatchNormal normalization layer and Relu activation function, then one convolution self-attention block is subjected to AvgPool average pooling, MatMul fusion matrix, second Relu activation function, MatMul fusion matrix, Sigmoid activation function, and then fused with the output of the Relu activation function and then output, the other convolution self-attention block is processed in parallel by Mean and Max and then merged by Concat, and after passing through Conv1d one-dimensional convolution layer and Sigmoid activation function, it is fused with the output of the Relu activation function and then output; S4. Optimizing the computational graph of the deep learning model; S5. Deploy the deep learning model after computational graph optimization, perform bridge structure health monitoring based on the reconstructed signal, and output bridge damage diagnosis results and damage types.

2. The bridge health monitoring method based on deep learning according to claim 1 is characterized in that: In S2, the threshold function is: In the formula, η represents the threshold, q j，k and q j，k 'represent the signal data before and after denoising, δ and ε are constants.

3. The bridge health monitoring method based on deep learning according to claim 1 is characterized in that: In S3, the deep learning framework is Pytorch.

4. The bridge health monitoring method based on deep learning according to claim 3 is characterized in that: The S4 includes: converting the trained pytorch model into ONNX.

5. A bridge health monitoring system for embedded systems based on deep learning, characterized in that: include: at least one processor; and at least one memory in communication with the processor, wherein: The memory stores program instructions executable by the processor, and the processor can execute the method according to any one of claims 1 to 4 by calling the program instructions.

6. A non-transitory computer-readable storage medium, characterized in that: The non-transitory computer-readable storage medium stores computer instructions that cause the computer to execute the method according to any one of claims 1 to 4.

Images