A pre-trained bridge damage identification method

Abstract

This invention discloses a pre-trained bridge damage identification method, comprising: acquiring and preprocessing vibration response data and temperature change data collected by a bridge sensor system; generating samples from the preprocessed data through a sliding window and dividing them into model training data and data to be predicted; constructing a pre-trained neural network model and pre-training it using the model training data; using the encoder of the optimal pre-trained model as the encoding layer and further training it using the model training data; inputting the data to be predicted into the trained model and outputting the bridge damage identification result. This invention improves the Transformer algorithm by combining environmental temperature changes to mine inherent structural information of the bridge, making the model more sensitive to bridge damage. It also utilizes a two-stage attention mechanism to improve the model's ability to mine spatiotemporal features of the response. This invention has higher accuracy and good generalization ability, and can effectively reduce labor costs.

Description

Technical Field

[0001] This invention relates to the technical field of bridge structural damage identification, and in particular to a bridge damage identification method based on pre-training. Background Technology

[0002] As man-made structures, bridges are subject to various natural and man-made forces outdoors, leading to damage such as deck cracking and pier subsidence. This affects the overall stability and safety of the bridge, shortening its service life. If not maintained promptly, the accumulated damage can lead to bridge collapse, causing severe loss of life and property. Based on the above analysis, researching bridge health monitoring to enable timely and rapid bridge maintenance is crucial.

[0003] In recent years, bridges have often been equipped with sensor systems. These systems continuously collect and transmit data, allowing managers to perform precise analysis using computer technology. Previous technologies, however, suffered from low reliability, limited capacity to handle massive amounts of data, and an inability to address data sparsity. Furthermore, they often required specialized personnel for methodology design, placing high demands on managers. But with the rapid development of deep learning technology, which has demonstrated powerful information capture and learning capabilities in areas such as weather forecasting and anomaly detection, bridge health monitoring is actively utilizing deep learning to achieve more effective bridge management.

[0004] However, current common deep learning methods for bridge damage identification still use some basic models, such as one-dimensional convolutional neural networks (1DCNN) and long short-term memory neural networks (LSTM). These models perform well on some bridges, but they also have significant limitations. For example, they cannot simultaneously extract the spatial and temporal features of sensor response data; or they are only applicable to certain simple bridge structures, with lower performance for other complex structures; or they do not consider the influence of environmental factors and the role of the bridge's structural information, thus preventing the methods from being effectively generalized. Therefore, proposing a more effective bridge damage identification method is extremely crucial. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings and deficiencies of existing technologies and provide a pre-trained bridge damage identification method. This method fully utilizes data collected by sensor systems deployed on bridges to construct a pre-trained neural network model capable of identifying damage. It extracts hidden high-dimensional information from the data to accurately identify the location and extent of bridge damage. This method can combine sensor response data and environmental temperature changes for damage identification. Pre-training with model training data allows the model to be widely used in various structural scenarios, solving the problems of current neural network models being unable to be applied to various bridge structures, unable to combine bridge structural information and environmental influences, and having poor damage prediction results.

[0006] To achieve the above objectives, the technical solution provided by this invention is: a bridge damage identification method based on pre-training, comprising the following steps:

[0007] Step 1: Obtain vibration response data and temperature change data collected by the bridge sensor system, and perform outlier and zero-mean standardization on the data. Extract samples from the processed data through a sliding window to generate model input data, and divide it into model training data and data to be predicted.

[0008] Step 2: Construct a pre-trained neural network model using the model training data. The designed pre-trained neural network model is an improvement on the Transformer network, including an encoder and decoder. A two-stage attention mechanism and a feature communication module are introduced to improve the model's encoding ability for vibration response data. The vibration response data in the model training data is randomly masked, and the masked part is reconstructed during the pre-training process. The randomly masked vibration response data and temperature change data are input into the pre-trained neural network model. The two-stage attention mechanism extracts the feature information of the vibration response data, called the first feature information, and the multi-head attention mechanism extracts the feature information of the temperature change data, called the second feature information. The first and second feature information are then exchanged through the feature communication module to obtain the fused feature information. The fused feature information is passed through two parallel linear transformation layers to obtain the encoding of the vibration response data and temperature change data. The obtained vibration response encoding is then input into the decoder to obtain the reconstructed vibration response data. During the training process, the loss value of the reconstructed vibration response data and the original vibration response data is calculated using MSE. After multiple iterations, the optimal network model is obtained by minimizing the loss value.

[0009] Step 3: Further training for the bridge damage identification task. The encoder of the pre-trained optimal network model is used as the encoding layer, followed by a convolutional long short-term memory network as the spatiotemporal feature mining layer, and a fully connected network as the output layer. The vibration response data in the model training data is input into the encoding layer to obtain vibration response codes and temperature codes. The vibration response codes are then input into the spatiotemporal feature mining layer and the output layer to obtain identification labels for bridge damage. The optimal network model is obtained through iteration using MSE as the loss function.

[0010] Step 4: Input the data to be predicted into the optimal network model obtained in Step 3, output the identification results of bridge damage, and match the identification results with the bridge to complete the identification of the location and degree of bridge damage.

[0011] Furthermore, step 1 includes the following steps:

[0012] 1.1) The 3sigma principle is used to process outliers in the data of each damage condition. Vibration response data with severe deviations are considered to be abnormal and replaced with the average value.

[0013] 1.2) After anomaly handling, the data is standardized to zero-mean to ensure that the processed vibration response data follows a standard normal distribution with a mean of 0 and a standard deviation of 1, as shown in the following formula:

[0014]

[0015] In the formula, x * This represents the original vibration response data. σ represents the mean of the original vibration response data, σ represents the standard deviation of the original vibration response data, and x represents the vibration response data after standardization.

[0016] 1.3) The processed vibration response data is processed using a sliding window technique to obtain input data that meets the requirements. The sliding window size L is determined, and the vibration response data corresponding to each damage condition is divided into multiple samples of length L. Each sample contains n sensor response sequences, and the sample representation is as follows:

[0017]

[0018] In the formula, X represents a certain sample. This represents the vibration response data collected by sensor numbered i at time t-L+1. Similarly, and These represent the vibration response data collected by the sensor numbered i at time tj and time t, respectively. n represents the number of sensors, the superscript i represents the sensor number, and the subscript t represents the time when the data was recorded.

[0019] 1.4) Set data labels. Divide the bridge into m units according to the horizontal axis. The output label should be a vector of length m. The vector elements represent the damage that occurs in the corresponding unit. The elements are equal to 0 in the undamaged state and the elements represent the size of the damage when there is damage. Perform partial manual labeling on the collected vibration response data. After segmentation, use this part of the vibration response data as training data. After training, the model can identify damage in the collected unlabeled data.

[0020] Further, in step 2, the encoder contains k layers, each layer containing a merging layer, a two-stage attention layer, a multi-head attention layer, an interaction layer, and two linear layers; the decoder contains k+1 layers, each layer having a similar structure to the encoder, containing a merging layer, a two-stage attention layer, a multi-head attention layer, an interaction layer, and a linear layer; mask reconstruction is used as a pre-training task, and each sensor response sequence of the sample is masked independently. The method is to mask each sensor response sequence with a fixed mask ratio, and set the data of the masked part to 0. Then, the pre-training process reconstructs the masked sensor response sequence, attempting to restore the masked part. The closer the restoration is to the original response, the better the pre-training effect.

[0021] The pre-training process of the pre-trained neural network model is as follows:

[0022] 2.1) Divide the input data into segments according to the time dimension, encode each segment separately, and obtain the encoded vector matrix X. emb T emb ∈R sn×sl×n , where X emb Represents the vibration response code, T emb This represents the temperature encoding, where R represents the real number space, sn represents the number of segments, sl represents the segment length, and n represents the number of sensors.

[0023] 2.2) Adding a learnable position code to the vibration response code and temperature code yields a new X. emb T emb ;

[0024] 2.3) Calculate two-stage attention for vibration response encoding. Two-stage attention refers to the process of encoding X... emb Attention is calculated separately for the time and space dimensions. First, the time attention matrix W is calculated. t ∈R sn×sn Each sensor response sequence encoded by vibration response shares this temporal attention matrix, and then the spatial attention matrix W is calculated. s ∈R n×nThe vibration response encoding shares a spatial attention matrix at each time step, and the first feature information is obtained through two steps; the temperature encoding is input into a multi-head attention mechanism to calculate the second feature information.

[0025] 2.4) Set up an interaction layer to sum and linearly transform the first and second feature information after attention calculation to obtain the output of the interaction layer;

[0026] 2.5) The first feature information, the second feature information, and the output of the interaction layer are summed and then linearly transformed. In this way, the vibration response code and temperature code are obtained by one exchange to obtain a new vibration response code and temperature code.

[0027] 2.6) Merge the segmented vibration response codes and temperature codes into segments to reduce the number of segments and facilitate the extraction of long-range features;

[0028] 2.7) Repeat steps 2.3-2.6) to obtain the outputs of all coding layers, and then calculate the output enc of the k-th layer of the encoder with the corresponding layer in the decoder. k The vibration response code and temperature code input to each decoder layer are calculated in the same way as those input to the encoder (steps 2.3-2.6), and the resulting vibration response code is then processed. k and the corresponding encoder layer output enc k Perform multi-head attention score calculation and obtain the output of the decoder layer through linear transformation;

[0029] 2.8) Use the output of the last layer of the decoder as the reconstructed vibration response data;

[0030] 2.9) Compare the reconstructed vibration response data with the original vibration response data to calculate the loss and optimize the model parameters so that the difference between the two is continuously reduced.

[0031] Furthermore, in step 3, the encoder of the pre-trained neural network model from step 2 is used as the encoding layer to encode the vibration response data, and the encoded data is input into the spatiotemporal feature mining layer for training. The spatiotemporal feature mining layer contains a ConvLSTM network and an attention layer, while the output layer consists of two fully connected layers. The training process is as follows:

[0032] 3.1) The vibration response encoding x_enc output by the encoding layer is deformed to fit the input shape of the ConvLSTM network;

[0033] 3.2) Input the deformed encoding into the ConvLSTM network to obtain the output hidden layer vector cl_out;

[0034] 3.3) Flatten cl_out and then perform attention calculation to obtain the weighted and scaled result acl_out;

[0035] 3.4) Input acl_out into the output layer, and obtain the final recognition label after passing through two fully connected layers;

[0036] 3.5) Using mean squared error (MSE) as the loss function, the loss is calculated by combining the model output label with the true label in the training data. Then, the Adam optimizer is used to update the parameters in the model, thereby improving the accuracy of the model in identifying the location and extent of bridge damage.

[0037] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0038] 1. This invention improves the Transformer network by designing a pre-trained neural network model for pre-training, which enhances the encoding capability of sensor data, exposes the high-dimensional features hidden in the data, and improves the recognition effect of the model.

[0039] 2. The encoder of the pre-trained neural network model uses a two-stage attention mechanism, which can pay attention to the sensor response in terms of time and space, thereby improving the ability to extract hidden information from the sensor data.

[0040] 3. The encoder of the designed pre-trained neural network model incorporates sensor response and temperature changes, fully considering environmental factors. Furthermore, based on temperature changes and sensor strain responses, it can calculate the matrix information of bridge structural stiffness. This information intuitively reflects bridge structural damage, indicating that the model can utilize bridge structural information, which is highly beneficial for damage identification tasks. Simultaneously, the ability to incorporate bridge structural knowledge also improves the model's interpretability.

[0041] 4. This invention automatically identifies bridge damage by designing a neural network model, which can greatly reduce labor costs and improve the real-time performance of bridge health monitoring.

[0042] 5. This invention realizes the application of pre-trained algorithms in bridge damage identification tasks, enabling the accumulated massive bridge sensor monitoring data to generate greater value and has broad application prospects. Attached Figure Description

[0043] Figure 1 This is a schematic diagram of the logical flow of the method of the present invention. Detailed Implementation

[0044] The present invention will be further described in detail below with reference to the embodiments and accompanying drawings, but the embodiments of the present invention are not limited thereto.

[0045] like Figure 1 As shown, this embodiment provides a pre-trained bridge damage identification method. It utilizes a sensor system deployed on the bridge to collect data, designs a pre-trained neural network model as the damage identification model, and extracts hidden high-dimensional information from the data to accurately identify the location and extent of bridge damage. This method can combine sensor response data and environmental temperature changes for damage identification. Pre-training with model training data allows the model to be widely used in various structural scenarios, solving the problems of current neural network models being unable to be applied to multiple bridge structures, unable to combine bridge structural information and environmental influences, and having poor damage prediction results.

[0046] The specific implementation of this bridge damage identification method includes the following steps:

[0047] Step 1: Obtain vibration response data and temperature change data collected by the bridge sensor system, and perform outlier and zero-mean standardization on the data. Extract samples from the processed data through a sliding window to generate model input data, and divide it into model training data and data to be predicted.

[0048] Step 2: Construct a pre-trained neural network model using the model training data. The designed pre-trained neural network model is an improvement on the Transformer network, including an encoder and decoder. A two-stage attention mechanism and a feature communication module are introduced to improve the model's encoding ability for vibration response data. The vibration response data in the model training data is randomly masked, and the masked part is reconstructed during the pre-training process. The randomly masked vibration response data and temperature change data are input into the pre-trained neural network model. The two-stage attention mechanism extracts the feature information of the vibration response data, called the first feature information, and the multi-head attention mechanism extracts the feature information of the temperature change data, called the second feature information. The first and second feature information are then exchanged through the feature communication module to obtain the fused feature information. The fused feature information is passed through two parallel linear transformation layers to obtain the encoding of the vibration response data and temperature change data. The obtained vibration response encoding is then input into the decoder to obtain the reconstructed vibration response data. During the training process, the loss value of the reconstructed vibration response data and the original vibration response data is calculated using MSE. After multiple iterations, the optimal network model is obtained by minimizing the loss value.

[0049] Step 3: Further training for the bridge damage identification task. The encoder of the pre-trained optimal network model is used as the encoding layer, followed by a convolutional long short-term memory network as the spatiotemporal feature mining layer, and a fully connected network as the output layer. The vibration response data in the model training data is input into the encoding layer to obtain vibration response codes and temperature codes. The vibration response codes are then input into the spatiotemporal feature mining layer and the output layer to obtain identification labels for bridge damage. The optimal network model is obtained through iteration using MSE as the loss function.

[0050] Step 4: Input the data to be predicted into the optimal network model obtained in Step 3, output the identification results of bridge damage, and match the identification results with the bridge to complete the identification of the location and degree of bridge damage.

[0051] Specifically, step 1 includes the following steps:

[0052] 1.1) Obtain strain response data from the bridge sensor system. This data consists of data from 20 sensors installed on the bridge (10 at the top and 10 at the bottom, evenly distributed). Use the 3 sigma principle to process outliers in the data for each damage condition. Values ​​of vibration response data that deviate significantly are considered abnormal and replaced with the average value.

[0053] 1.2) After anomaly processing, the data is standardized to zero-mean, ensuring that the processed vibration response data follows a standard normal distribution with a mean of 0 and a standard deviation of 1. The formula is as follows:

[0054]

[0055] Where, x * This represents the original vibration response data. σ represents the mean of the original vibration response data, σ represents the standard deviation of the original vibration response data, and x represents the vibration response data after standardization.

[0056] 1.3) The processed vibration response data is processed using a sliding window technique to obtain the required input data. The sliding window size is set to 128 time steps. The vibration response data for each damage condition is divided into multiple samples of shape (128, 20), where the sample at time t is represented as follows:

[0057]

[0058] Where X represents a sample. This represents the vibration response data collected by sensor number i at time t. The superscript i indicates the sensor number, and the subscript t indicates the time when the data was recorded.

[0059] 1.4) Set data labels: Divide the bridge into 50 units along the horizontal axis. The output label should be a vector of length 50. The vector elements represent the damage in the corresponding unit. In the undamaged state, the element is equal to 0; when there is damage, the element represents the magnitude of the damage. For example, if the damage level at unit number 20 is 60%, the element at position 20 in the vector should be set to 0.6. After all working conditions are segmented, manually label the damage. Use this segmented data as training data. The trained model will then be able to identify damage from the collected unlabeled data.

[0060] Specifically, the encoder comprises three layers, each containing a merging layer, a two-stage attention layer, a multi-head attention layer, an interaction layer, and two linear layers. The decoder comprises four layers, each with a similar structure to the encoder, including a merging layer, a two-stage attention layer, a multi-head attention layer, an interaction layer, and a linear layer. Mask reconstruction is used as a pre-training task. Each sensor response sequence of the samples is independently masked by setting 30% of the data in a sequence to 0, with a maximum mask length of 4. The pre-training process then reconstructs the masked sensor response sequences, attempting to partially restore the masked portion. The closer the restored response is to the original, the better the pre-training effect.

[0061] The pre-training process of a pre-trained neural network model is as follows:

[0062] 2.1) Divide the input data into segments according to the time dimension, encode each segment separately, and obtain the encoded vector matrix X. emb ,T emb , where X emb Represents the vibration response code, T emb The temperature codes are represented by shapes (8, 16, 20), which correspond to the number of segments, the length of the segments, and the number of sensors, respectively.

[0063] 2.2) Adding a learnable position code to the vibration response code and temperature code yields a new X. emb ,T emb .

[0064] 2.3) Calculate two-stage attention for vibration response encoding. Two-stage attention refers to the process of encoding X... emb Attention is calculated separately for the time and space dimensions. First, the time attention matrix W is calculated. t Each sensor response sequence encoded by vibration response shares this temporal attention matrix, and then the spatial attention matrix W is calculated. s The vibration response encoding shares a spatial attention matrix at each time step, and the first feature information is obtained after two steps; the temperature encoding is input into a multi-head attention mechanism to calculate the second feature information.

[0065] 2.4) Set up an interaction layer to sum and linearly transform the first and second feature information after attention calculation to obtain the output of the interaction layer.

[0066] 2.5) The first feature information, the second feature information and the output of the interaction layer are summed and then linearly transformed. In this way, the vibration response code and temperature code are obtained by one exchange to obtain a new vibration response code and temperature code.

[0067] 2.6) Merge the segmented vibration response code and temperature code to reduce the number of segments and facilitate the extraction of long-range features. For example, if the output dimension of the previous layer is (8,16,20), the merged code dimension is (4,32,20).

[0068] 2.7) Repeat steps 2.3-2.6) to obtain the outputs of all coding layers, and then perform calculations with each corresponding layer in the decoder. For example, the output of the second layer of the encoder participates in the calculation of the second layer of the decoder. The vibration response code and temperature code input to each decoder layer are processed in the same way as the encoder, following steps 2.3-2.6). The obtained vibration response code and the corresponding encoder layer output are then used to calculate the multi-head attention score, and the decoder layer output is obtained through a linear transformation.

[0069] 2.8) Use the output of the last layer of the decoder as the reconstructed vibration response data.

[0070] 2.9) Compare the reconstructed vibration response data with the original vibration response data to calculate the loss and optimize the model parameters so that the difference between the two is continuously reduced.

[0071] Specifically, in the damage identification task of step 3, the encoder of the pre-trained neural network model from step 2 is used as the encoding layer to encode the vibration response data, and the encoded data is input into the spatiotemporal feature mining layer for training. The spatiotemporal feature mining layer consists of a ConvLSTM network and an attention layer, while the output layer consists of two fully connected layers. The training process is as follows:

[0072] 3.1) The vibration response encoding x_enc output by the encoding layer is deformed to fit the input shape of the ConvLSTM network. The deformed shape of x_enc is (128,2,10,1).

[0073] 3.2) Input the deformed encoding into the ConvLSTM network to obtain the output hidden layer vector cl_out.

[0074] 3.3) Flatten cl_out and then perform attention calculation to obtain the weighted and scaled result acl_out.

[0075] 3.4) Input acl_out into the output layer, and obtain the final recognition label after passing through two fully connected layers.

[0076] 3.5) The mean squared error (MSE) is used as the loss function. The loss is calculated by combining the model's output label with the true labels in the training data. Then, the Adam optimizer is used to update the parameters in the model, improving the accuracy of the model in identifying the location and extent of bridge damage. The learning rate is set to 0.001, and an early stopping mechanism is used for training. Training is stopped if the loss of the training data does not decrease after 10 rounds of training.

[0077] After training is complete, the unlabeled data to be identified is input into the obtained optimal model, which will output the bridge damage identification results.

[0078] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.

Claims

1. A bridge damage identification method based on pre-training, characterized in that, Includes the following steps: Step 1: Obtain vibration response data and temperature change data collected by the bridge sensor system, and perform outlier and zero-mean standardization on the data. Extract samples from the processed data through a sliding window to generate model input data, and divide it into model training data and data to be predicted. Step 2: Construct a pre-trained neural network model using the model training data. The designed pre-trained neural network model is an improvement on the Transformer network, including an encoder and a decoder. A two-stage attention mechanism and a feature communication module are introduced to improve the model's encoding ability for vibration response data. The vibration response data in the model training data is randomly masked, and the masked part is reconstructed during the pre-training process. The randomly masked vibration response data and temperature change data are input into the pre-trained neural network model. The two-stage attention mechanism extracts the feature information of the vibration response data, called the first feature information, and the multi-head attention mechanism extracts the feature information of the temperature change data, called the second feature information. The first and second feature information are then exchanged through the feature communication module to obtain the fused feature information. The fused feature information is passed through two parallel linear transformation layers to obtain a new encoding of the vibration response data and temperature change data, which is then input into the next layer of the encoder. Finally, the vibration response encoding obtained by the encoder is input into the decoder to obtain the reconstructed vibration response data. During training, MSE is used to calculate the loss value of the reconstructed vibration response data and the original vibration response data. After multiple iterations, the optimal network model is obtained by minimizing the loss value. The encoder comprises k layers, each containing a merging layer, a two-stage attention layer, a multi-head attention layer, an interaction layer, and two linear layers; the decoder comprises k+1 layers, each with a similar structure to the encoder, including a merging layer, a two-stage attention layer, a multi-head attention layer, an interaction layer, and a linear layer; mask reconstruction is used as a pre-training task. Each sensor response sequence of the sample is masked independently using a fixed masking ratio, with the masked data set to 0. The pre-training process then reconstructs the masked sensor response sequence, attempting to restore the masked portion. The closer the restored response is to the original response, the better the pre-training effect. Step 3: Further training for the bridge damage identification task. The encoder of the pre-trained optimal network model is used as the encoding layer, followed by a convolutional long short-term memory network as the spatiotemporal feature mining layer, and a fully connected network as the output layer. The vibration response data in the model training data is input into the encoding layer to obtain vibration response codes and temperature codes. The vibration response codes are then input into the spatiotemporal feature mining layer and the output layer to obtain identification labels for bridge damage. The optimal network model is obtained through iteration using MSE as the loss function. Step 4: Input the data to be predicted into the optimal network model obtained in Step 3, output the identification results of bridge damage, and match the identification results with the bridge to complete the identification of the location and degree of bridge damage.

2. The bridge damage identification method based on pre-training according to claim 1, characterized in that, Step 1 includes the following steps: 1.1) The 3sigma principle is used to process outliers in the data of each damage condition. Vibration response data with severe deviations are considered to be abnormal and replaced with the average value. 1.2) After anomaly handling, the data is standardized to zero-mean to ensure that the processed vibration response data follows a standard normal distribution with a mean of 0 and a standard deviation of 1, as shown in the following formula: In the formula, x * This represents the original vibration response data. σ represents the mean of the original vibration response data, σ represents the standard deviation of the original vibration response data, and x represents the vibration response data after standardization. 1.3) The processed vibration response data is processed using a sliding window technique to obtain input data that meets the requirements. The sliding window size L is determined, and the vibration response data corresponding to each damage condition is divided into multiple samples of length L. Each sample contains n sensor response sequences, and the sample representation is as follows: In the formula, X represents a certain sample. This represents the vibration response data collected by sensor numbered i at time t-L+1. Similarly, and These represent the vibration response data collected by the sensor numbered i at time tj and time t, respectively. n represents the number of sensors, the superscript i represents the sensor number, and the subscript t represents the time when the data was recorded. 1.4) Set data labels. Divide the bridge into m units according to the horizontal axis. The output label should be a vector of length m. The vector elements represent the damage that occurs in the corresponding unit. The elements are equal to 0 in the undamaged state and the elements represent the size of the damage when there is damage. Manually label some of the collected vibration response data and use this part of the vibration response data as training data after segmentation. After training, the model can identify damage in the collected unlabeled data.

3. The bridge damage identification method based on pre-training according to claim 1, characterized in that, In step 2, the pre-training process of the pre-trained neural network model is as follows: 2.1) Divide the input data into segments according to the time dimension, encode each segment separately, and obtain the encoded vector matrix X. emb T emb ∈R sn×sl×n , where X emb Represents the vibration response code, T emb This represents the temperature encoding, where R represents the real number space, sn represents the number of segments, sl represents the segment length, and n represents the number of sensors. 2.2) Adding a learnable position code to the vibration response code and temperature code yields a new X. emb T emb ; 2.3) Calculate two-stage attention for vibration response encoding. Two-stage attention refers to the attention given to X. emb Attention is calculated separately for the time and space dimensions. First, the time attention matrix W is calculated. t ∈R sn×sn Each sensor response sequence encoded by vibration response shares this temporal attention matrix, and then the spatial attention matrix W is calculated. s ∈R n×n The vibration response encoding shares a spatial attention matrix at each time step, and the first feature information is obtained after two steps. Temperature-encoded input is processed using a multi-head attention mechanism to calculate the second feature information. 2.4) Set up an interaction layer to sum and linearly transform the first and second feature information after attention calculation to obtain the output of the interaction layer; 2.5) The first feature information, the second feature information, and the output of the interaction layer are summed and then linearly transformed. In this way, the vibration response code and temperature code are obtained by one exchange to obtain a new vibration response code and temperature code. 2.6) Merge the segmented vibration response codes and temperature codes into segments to reduce the number of segments and facilitate the extraction of long-range features; 2.7) Repeat steps 2.3-2.6) to obtain the outputs of all coding layers, and then calculate the output enc of the k-th layer of the encoder with the corresponding layer in the decoder. k The vibration response code and temperature code input to each decoder layer are calculated in the same way as those input to the encoder (steps 2.3-2.6), and the resulting vibration response code is then processed. k and the corresponding encoder layer output enc k Perform multi-head attention score calculation and obtain the output of the decoder layer through linear transformation; 2.8) Use the output of the last layer of the decoder as the reconstructed vibration response data; 2.9) Compare the reconstructed vibration response data with the original vibration response data to calculate the loss and optimize the model parameters so that the difference between the two is continuously reduced.

4. The bridge damage identification method based on pre-training according to claim 1, characterized in that, In step 3, the encoder of the pre-trained neural network model from step 2 is used as the encoding layer to encode the vibration response data. The encoded data is then input into the spatiotemporal feature mining layer for training. The spatiotemporal feature mining layer consists of a ConvLSTM network and an attention layer, while the output layer is composed of two fully connected layers. The training process is as follows: 3.1) The vibration response encoding x_enc output by the encoding layer is deformed to fit the input shape of the ConvLSTM network; 3.2) Input the deformed encoding into the ConvLSTM network to obtain the output hidden layer vector cl_out; 3.3) Flatten cl_out and then perform attention calculation to obtain the weighted and scaled result acl_out; 3.4) Input acl_out into the output layer, and obtain the final recognition label after passing through two fully connected layers; 3.5) Using mean squared error (MSE) as the loss function, the loss is calculated by combining the model output label with the true label in the training data. Then, the Adam optimizer is used to update the parameters in the model, thereby improving the accuracy of the model in identifying the location and extent of bridge damage.

Images