A Deep Learning-Based Deflection Response Matching Method for Bridge Groups with the Same Load Source

Abstract

This application discloses a deep learning-based method for matching the deflection response of a bridge group under the same load source, comprising: collecting the deflection response caused by vehicles passing over each bridge in the bridge group; extracting the waveform, time sequence, and phase of the deflection response of each heavy vehicle; calculating the time difference and phase difference of the deflection response of the same heavy vehicle passing over each bridge; constructing a deflection response feature dataset based on the waveform, time difference, and phase difference; assigning category labels to the deflection response feature dataset; constructing a matching model of the same source wave based on a multi-input convolutional neural network model; selecting a reference deflection response data segment and candidate deflection responses; truncating the candidate deflection responses into several candidate deflection response data sub-segments according to the duration of the reference deflection response data segment; calculating the matching probability between each candidate deflection response data sub-segment and the reference deflection response data segment; and using the candidate deflection response data sub-segment corresponding to the maximum matching probability as the matching deflection response of the reference deflection response data segment.

Description

Technical Field

[0001] This application relates to the field of bridge structural health monitoring technology, specifically to a deep learning-based method for matching the deflection response of a group of bridges under the same load source. Background Technology

[0002] With the increasing density of regional road networks and the continuous growth of traffic flow, the coordinated monitoring of the structural safety and service performance of bridge groups, as components of key transportation hubs, is becoming increasingly important. While bridge monitoring systems have been successfully applied to individual bridge operations, the expansion of bridges and the coverage of monitoring indicators makes it difficult for traditional single-bridge monitoring methods to reveal the interactions or common response patterns between bridges. The cluster effect of monitoring data remains to be explored, posing a significant challenge to the accurate and efficient identification and assessment of the structural state and differences among bridges within a group. It is worth noting that heavy-duty vehicles, as the main live loads causing significant responses in national and provincial highway bridges, contain rich structural information during their passage.

[0003] The invention patent application with publication number CN120892975A proposes a lightweight bridge monitoring method and system applicable to bridge group monitoring. This technical solution predicts the deflection of the bridge to be monitored based on deflection monitoring data of a specific bridge within a bridge group, and achieves accurate diagnosis of the bridge's structural condition through quantitative evaluation indicators. The method selects deflection time-history data as the monitoring indicator, leveraging the characteristic that heavy-duty vehicles are the main live loads causing significant responses on national and provincial highway bridges to fully explore the rich structural information contained in their passage. Secondly, by introducing deflection data from "bridges that have passed first" as a basic term and innovatively designing a difference term extraction mechanism, it effectively utilizes monitoring data from adjacent bridges within the bridge group, overcoming the limitations of single-bridge analysis, enhancing the correlation of the dynamic response of the bridge group, and overcoming the large errors caused by the lack of perception of the bridge's own state and external loads in traditional prediction models.

[0004] In the process of "difference extraction", this technical solution usually requires manual collection of measured deflection data of the same historical heavy vehicle passing through various bridges, and manual marking and matching of the deflection response of the same historical heavy vehicle.

[0005] Traditional matching methods primarily rely on manual feature extraction and threshold determination to match the deflection response of heavy vehicles within a bridge complex. The core process involves manually extracting basic temporal physical features such as amplitude range, duration, and rise slope from the deflection time-history curves collected by various sensors. Then, based on threshold conditions preset by expert experience, the matching result is confirmed by determining whether the characteristics of the deflection response waveforms of each bridge simultaneously meet the preset threshold conditions within a reasonable time window. However, this method heavily depends on prior expert knowledge, resulting in lengthy deployment and parameter tuning cycles. It requires tedious, one-to-one threshold calibration and setting for each bridge in the bridge complex, and even for each possible traffic flow state, relying on the experience of senior engineers. This is an extremely time-consuming and non-automated process, failing to meet the requirements of rapid deployment and agile response in modern infrastructure operation and maintenance. Furthermore, when faced with a new bridge complex, this method lacks sufficient adaptability and generalization ability. Summary of the Invention

[0006] To address the problems existing in the prior art, this application proposes a deep learning-based method for matching the deflection response of a bridge group under the same load source. This method can effectively correlate the vertical displacement responses of each bridge under the same vehicle load, and deeply explore the correlation and differences of signals from the same source in the time and frequency domains. This provides a deeper level of information mining and correlation analysis for the overall working status assessment of the bridge group.

[0007] A deep learning-based method for matching the deflection response of a group of bridges under the same load source includes:

[0008] The system collects vehicle-induced deflection responses as vehicles pass over each bridge in a bridge complex; identifies the deflection responses of the same heavy vehicle passing over each bridge from the vehicle-induced deflection responses; extracts the waveform, time sequence, and phase of the deflection responses of each heavy vehicle; calculates the time difference and phase difference of the deflection responses of the same heavy vehicle passing over each bridge; and constructs a deflection response feature dataset of heavy vehicles based on the waveform, time difference, and phase difference; and assigns category labels to the deflection response feature dataset to construct a deflection response feature dataset with category labels.

[0009] Constructing a homologous wave matching model based on a multi-input convolutional neural network model;

[0010] A reference deflection response data segment and a candidate deflection response are selected; based on the duration of the reference deflection response data segment, the candidate deflection response is truncated into several candidate deflection response data segments; the co-source wave matching model calculates the matching probability between each candidate deflection response data segment and the reference deflection response data segment, wherein the candidate deflection response data segment corresponding to the maximum matching probability is the matching result of the reference deflection response data segment.

[0011] In one possible implementation, the step of extracting the waveform, timing, and phase of the deflection response of each heavy vehicle, calculating the time difference and phase difference of the deflection response of the same heavy vehicle passing over each bridge, and constructing a deflection response feature dataset of the heavy vehicles based on the waveform, time difference, and phase difference includes:

[0012] Definition of the first i The first heavy vehicle to cross the bridge group j bridge 、 No. k The waveform of the deflection response caused by the bridge is as follows: , , , ，i This represents the ordinal index of the loaded vehicles. j、k This represents the bridge ordinal index. j ≠ k , N Represents the number of loaded vehicles. M This represents the number of bridges in the same bridge group; a waveform dataset of the deflection response of all loaded vehicles with respect to the current bridge group is constructed. ;

[0013] Definition of the first i The first heavy vehicle to cross the bridge group j bridge 、 No. k The timing sequence of the deflection response caused by the bridge is as follows: , Then the first i The first heavy vehicle to cross the bridge group j The bridge and the k The time difference between the deflection responses of the two bridges is... Construct a time difference dataset of the deflection response of all heavy vehicles with respect to the current bridge group. ;

[0014] Definition of the first i The first heavy vehicle to cross the bridge group j bridge 、 No. k The phase of the deflection response caused by the bridge is , Then the first i The first heavy vehicle to cross the bridge group j The bridge and the k The phase difference between the deflection responses of the two bridges is... ;

[0015] Construct a phase difference dataset of the deflection response of all heavy vehicles with respect to the current bridge group. ; Construct the first i Dataset of deflection response characteristics of heavy vehicles on the current bridge group , Indicates the first i The number of heavy vehicles in the current bridge group j The bridge and the k Data set of deflection response characteristics of a bridge.

[0016] In one possible implementation, assigning category labels to the deflection response feature dataset to construct a deflection response feature dataset with category labels includes:

[0017] Definition of the first i Dataset of deflection response characteristics of heavy vehicles on the current bridge group The category label is Construct a class label dataset of the deflection response of all heavy vehicles with respect to the current bridge group. ;

[0018] Construct a dataset of deflection response features of all heavy vehicles with respect to the current bridge group, labeled with categories. .

[0019] In one possible implementation, the construction of the same-origin wave matching model based on the multi-input convolutional neural network model includes:

[0020] A multi-input convolutional neural network model is constructed. The multi-input convolutional neural network model includes three input branches, each of which is connected to a waveform feature extraction layer, a time difference feature extraction layer, and a phase difference feature extraction layer, respectively. The outputs of the waveform feature extraction layer, the time difference feature extraction layer, and the phase difference feature extraction layer are respectively connected to a feature fusion layer, and the output of the feature fusion layer is connected to a classification output layer.

[0021] The deflection response feature dataset with category labels is divided proportionally to obtain a training set, a validation set, and a test set;

[0022] The partitioned training set is input into the multi-input convolutional neural network model to train and optimize the network model parameters. The trained network model is then used as a matching model for the same source wave.

[0023] In one possible implementation, the waveform feature extraction layer includes a convolutional module, a convolutional block attention module, and a long short-term memory network connected in sequence; wherein, the convolutional module includes a first convolutional unit and a second convolutional unit connected in sequence, the first convolutional unit is composed of a first one-dimensional convolutional layer, a first batch normalization layer, and a first pooling layer connected in sequence, and the second convolutional unit is composed of a second one-dimensional convolutional layer, a second batch normalization layer, and a second pooling layer connected in sequence, and the output of the first pooling layer is connected to the input of the second one-dimensional convolutional layer;

[0024] The waveform feature extraction layer extracts features from the waveform through a convolutional module and outputs a first feature map U. The convolutional block attention module sequentially applies attention weights to the first feature map U in both the channel and spatial dimensions and outputs a second feature map U'. The long short-term memory network performs global information aggregation on the feature sequence of the second feature map U' to obtain the temporal feature vector of the waveform. .

[0025] In one possible implementation, the convolutional block attention module sequentially performs attention weighting on the first feature map U in both the channel dimension and the spatial dimension, and outputs a second feature map U', including:

[0026] Global average pooling and global max pooling operations are performed on the first feature map U to generate a first channel description vector and a second channel description vector; the first channel description vector is used to characterize the global average response, and the second channel description vector is used to characterize the global maximum response.

[0027] The first channel description vector and the second channel description vector are respectively input into a shared multilayer perceptron, and the first channel description vector and the second channel description vector are respectively subjected to nonlinear transformation by the multilayer perceptron;

[0028] The first and second channel description vectors after nonlinear transformation are fused and processed by an activation function to generate a normalized channel weight matrix.

[0029] Multiply the channel weight matrix with the first feature map U channel by channel to obtain the channel-weighted intermediate feature map U. c .

[0030] Furthermore, the convolutional block attention module sequentially performs attention weighting on the first feature map U in both the channel dimension and the spatial dimension, and outputs the second feature map U', and also includes:

[0031] For the intermediate feature map U c Perform average pooling and max pooling operations respectively to generate a first spatial feature description map and a second spatial feature description map;

[0032] The first spatial feature description map and the second spatial feature description map are spliced ​​together to obtain a multi-channel feature map;

[0033] The multi-channel feature map is input into a convolutional layer for feature extraction, and then processed by an activation function to generate a spatial weight matrix.

[0034] The spatial weight matrix and the intermediate feature map U are then compared. c Pixel-wise multiplication yields the second feature map U', which is weighted by pixels.

[0035] In one possible implementation, the time difference feature extraction layer and / or phase difference feature extraction layer employ a multilayer perceptron structure.

[0036] In one possible implementation, the steps include selecting a reference deflection response data segment and candidate deflection responses; truncating the candidate deflection responses into several candidate deflection response data segments based on the duration of the reference deflection response data segment; and calculating the matching probability between each candidate deflection response data segment and the reference deflection response data segment using the same-source wave matching model. The candidate deflection response data segment corresponding to the highest matching probability is the matching result of the reference deflection response data segment, including:

[0037] A reference deflection response data segment is selected; this segment is triggered when the target heavy vehicle passes over the first bridge in the bridge group, and its duration is [duration missing]. t 0;

[0038] Candidate deflection responses are selected and preprocessed; the candidate deflection responses are the deflection responses to be matched induced by various vehicles passing over the second bridge in the bridge group, and the monitoring time of the candidate deflection responses is [missing information]. T , T > t 0;

[0039] With t A sliding time window of equal length traverses the preprocessed candidate deflection response data sequence and extracts several candidate deflection response data segments.

[0040] Extract the waveform, timing, and phase of the reference deflection response data segment and each candidate deflection response data segment, and then calculate the time difference and phase difference between the reference deflection response data segment and each candidate deflection response data segment;

[0041] The matching model of the same source wave calculates the matching probability of each candidate deflection response data segment with the reference deflection response data segment, wherein the candidate deflection response data segment corresponding to the maximum matching probability is the matching result of the reference deflection response data segment.

[0042] In one possible implementation, a quantitative evaluation metric is used to evaluate the trained network model, wherein the quantitative evaluation metric includes at least one of accuracy, precision, recall, and F1 score.

[0043] Compared with the prior art, this application has the following beneficial effects:

[0044] 1. By focusing on extensively monitored deflection time history data, we can deeply mine the rich structural response information contained in heavy vehicle traffic, thereby replacing the reliance on expensive additional sensors and complex feature engineering, significantly reducing deployment and implementation costs. At the same time, it fundamentally reduces the reliance on expert prior knowledge, realizing a paradigm shift from "manual interpretation" to "intelligent recognition", greatly improving the level of automation and engineering universality, and demonstrating extremely high value for promotion and application.

[0045] 2. Employing a multi-input convolutional neural network architecture, this application fundamentally avoids the underfitting problems caused by insufficient feature extraction in simple models and the overfitting problems caused by excessive memorization of training data in complex models through the collaborative design of multidimensional perception and dynamic focusing. The model in this application constructs a multimodal perception space with complementary information by inputting time-domain waveforms, frequency-domain phase, and statistical time difference features in parallel. On this basis, an attention mechanism is introduced as an internal control center, enabling it to adaptively recalibrate and collaboratively optimize various features, thereby dynamically strengthening key discriminative information and suppressing redundant noise. This structure not only enables the model to deeply capture complex spatiotemporal patterns in waveforms, but also significantly improves the classification accuracy and generalization efficiency of homologous deflection responses through an intelligent feature selection mechanism.

[0046] 3. By training with a large number of labeled samples, the model can grasp the inherent laws and complex nonlinear modes of deflection response under various complex working conditions, thus getting rid of the dependence on fixed thresholds and having excellent adaptive ability. When deployed to a new bridge group, only a small amount of local data is needed to optimize some of its parameters, so that the model can quickly adapt to the specificity of the new structure while retaining general knowledge, thus overcoming the fatal defect of insufficient generalization ability of traditional methods. Attached Figure Description

[0047] Figure 1 A flowchart of a deep learning-based method for matching the deflection response of a bridge group with the same load source, provided for the implementation of this application;

[0048] Figure 2 A process for constructing a deflection response feature dataset provided for embodiments of this application;

[0049] Figure 3 A matching process for reference deflection response and candidate deflection response is provided for the implementation of this application;

[0050] Figure 4 Deflection time history curves of different bridges in the same bridge group provided for the implementation of this application;

[0051] Figure 5 The present application provides a structure for a multi-input convolutional neural network model.

[0052] Figure 6 A category label generation process provided for embodiments of this application;

[0053] Figure 7 A loss value versus accuracy curve for a training set provided for an embodiment of this application;

[0054] Figure 8 A loss value versus accuracy curve for a validation set provided for an embodiment of this application;

[0055] Figure 9 A test set confusion matrix provided for the implementation of this application. Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of this application clearer, the application will now be described in further detail with reference to the accompanying drawings.

[0057] It should be understood that the described embodiments are merely some, not all, of the embodiments in this application. All other embodiments obtained by those skilled in the art based on the embodiments in this application without inventive effort are within the scope of protection of this application.

[0058] The terminology used in the embodiments of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application. The singular forms "a," "the," and "the" as used in the embodiments of this application are also intended to include the plural forms unless the context clearly indicates otherwise.

[0059] Furthermore, where the terms "first" and "second" appear, these terms are for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined with "first" or "second" may explicitly or implicitly include at least one of that feature.

[0060] It should be understood that the term "and / or" used in this article is merely a description of the same field for related objects, indicating that three relationships can exist.

[0061] like Figures 1-3 As shown, a deep learning-based method for matching the deflection response of a group of bridges under the same load source includes:

[0062] S1: Construction of the deflection response feature dataset;

[0063] S11: Set up measuring points on each bridge of the bridge group and synchronously collect the vehicle-induced deflection response caused by vehicles passing through each bridge. The vehicle-induced deflection response includes the deflection response caused by multiple heavy vehicles passing through each bridge in sequence.

[0064] S12: Preprocess the vehicle-induced deflection response;

[0065] S13: Identify the deflection response caused when the same heavy vehicle passes over each bridge in the bridge group;

[0066] S14: Extract the waveform, timing, and phase of the deflection response caused by each heavy vehicle passing through each bridge in the bridge group; calculate the time difference and phase difference of the deflection response between two different bridges (not limited to adjacent bridges) when the same heavy vehicle passes through each bridge based on the timing and phase; construct the deflection response feature dataset of the heavy vehicle based on the waveform, time difference, and phase difference.

[0067] S15: Assign category labels to the deflection response feature dataset to construct a deflection response feature dataset with category labels; the principles for assigning category labels are as follows:

[0068] The waveforms, time differences, and phase differences of the deflection responses originating from the same heavy vehicle are associated as a homogeneous sample set and assigned a first category label; the waveforms, time differences, and phase differences of the deflection responses originating from different heavy vehicles are associated as a non-homogeneous sample set and assigned a second category label.

[0069] In one implementation, step S13 identifies the deflection response caused by the same heavy vehicle passing through each bridge in the bridge group based on the waveform and timing. Specifically, since the difference in deflection response caused by heavy vehicles and non-heavy vehicles is quite significant, in continuous monitoring data, obvious downward deflection in the waveform is usually regarded as the deflection response caused by heavy vehicles. Secondly, based on the bridge length and the speed of the heavy vehicle, the time difference interval from the heavy vehicle to the next bridge can be estimated. Therefore, based on the waveform and timing of historical vehicle-induced deflection responses, the deflection response caused by the same heavy vehicle passing through different bridges can be identified. Figure 4 The diagram shows the deflection time history curves of two bridges—Bridge A and Bridge B—in one embodiment. The red curve is the deflection time history curve collected from a measuring point in Bridge A. This curve shows three distinct downward deflections, indicating that three heavy vehicles passed through this measuring point in sequence. The blue curve is the deflection time history curve collected from a measuring point in Bridge B. This curve also shows three distinct downward deflections in sequence, and the three downward deflections in the two deflection time history curves exhibit a one-to-one correspondence of time difference characteristics. The time differences are all within the reasonable time difference range for a heavy vehicle to travel from the previous bridge to the next bridge. Based on the above principle, the deflection response caused by the same heavy vehicle passing through different bridges in the same bridge group can be identified and marked.

[0070] In one implementation, step S14 includes the following:

[0071] Definition of the first i The first heavy vehicle to cross the bridge group j bridge 、 No. kThe waveform of the deflection response caused by the bridge is as follows: , , , ，i This represents the ordinal index of the loaded vehicles. j、k This represents the bridge ordinal index. j ≠ k , N Represents the number of loaded vehicles. M This represents the number of bridges in the same bridge group; a waveform dataset of the deflection response of all loaded vehicles with respect to the current bridge group is constructed. ;

[0072] Definition of the first i The first heavy vehicle to cross the bridge group j bridge 、 No. k The timing sequence of the deflection response caused by the bridge is as follows: , Then the first i The first heavy vehicle to cross the bridge group j The bridge and the k The time difference between the deflection responses of the two bridges is... Construct a time difference dataset of the deflection response of all heavy vehicles with respect to the current bridge group. ;

[0073] Definition of the first i The first heavy vehicle to cross the bridge group j bridge 、 No. k The phase of the deflection response caused by the bridge is , Then the first i The first heavy vehicle to cross the bridge group j The bridge and the k The phase difference between the deflection responses of the two bridges is... ;

[0074] Construct a phase difference dataset of the deflection response of all heavy vehicles with respect to the current bridge group. ; Construct the first i Dataset of deflection response characteristics of heavy vehicles on the current bridge group , Indicates the first i The number of heavy vehicles in the current bridge group j The bridge and the k Data set of deflection response characteristics of bridges, , They represent the first i The first heavy vehicle to cross the bridge group j bridge 、 No.k The waveform of the deflection response caused by the bridge. , They represent the first i The first heavy vehicle to cross the bridge group j The bridge and the k The time difference and phase difference of the deflection response of the two bridges.

[0075] In one implementation, step S14 calculates the analytical deflection response using Hilbert transform and extracts the instantaneous phase information of the deflection response caused by each heavy vehicle passing through each bridge in the bridge group.

[0076] In one implementation, step S15 includes the following:

[0077] Definition of the first i Dataset of deflection response characteristics of heavy vehicles on the current bridge group The category label is Optionally, for the homologous sample set, the label value of the first category label is set to 1; for the non-homologous sample set, the label value of the second category label is set to 0; construct a category label dataset of the deflection response of all heavy vehicles with respect to the current bridge group. ;

[0078] Construct a dataset of deflection response features of all heavy vehicles with respect to the current bridge group, labeled with categories. .

[0079] S2: Construction of the matching model for the same source wave;

[0080] S21: Construct a multi-input convolutional neural network model based on one-dimensional convolutional layers (Conv1D), convolutional block attention modules (CBAM), long short-term memory networks (LSTM), fully connected layers (Dense), feature fusion layers (concat), and classification output layers;

[0081] Specifically, the multi-input convolutional neural network model includes three input branches, each taking the aforementioned waveform, time difference, and phase difference as inputs. These three input branches are respectively connected to a waveform feature extraction layer, a time difference feature extraction layer, and a phase difference feature extraction layer for feature extraction of the waveform, time difference, and phase difference. The outputs of the three feature extraction layers are respectively connected to a feature fusion layer (concat), which combines the output features of the three feature extraction layers. The output of the feature fusion layer is connected to a classification output layer to output the prediction result (i.e., the matching result).

[0082] S22: The deflection response feature dataset with category labels... The dataset is divided proportionally to obtain a training set, a validation set, and a test set.

[0083] S23: Input the partitioned training set into the multi-input convolutional neural network model, train and optimize the network model parameters based on the deflection response samples covering various working conditions in the training set, and use the trained network model as a matching model of the same source wave.

[0084] S3: Deflection response matching;

[0085] S31: Select a reference deflection response data segment; the reference deflection response data segment is induced when the target heavy vehicle passes over the first bridge in the bridge group, and its duration is... t 0;

[0086] S32: Select candidate deflection responses and preprocess them; the candidate deflection responses are deflection responses to be matched within a fixed time period collected from monitoring points on the second bridge in the bridge group, and the deflection responses to be matched include deflection responses caused by various vehicles, including target heavy vehicles; the monitoring duration of the candidate deflection responses is... T , T > t 0; It should be noted that "the first bridge and the second bridge" are not used to define the adjacent position relationship or the order of passage between the two bridges, but only to indicate two different bridges in the same bridge group.

[0087] S33: with t A sliding time window of equal length traverses the preprocessed candidate deflection response data sequence and extracts several candidate deflection response data segments.

[0088] S34: Extract the waveform, timing and phase of the reference deflection response data segment and each candidate deflection response data sub-segment, and then calculate the time difference and phase difference between the reference deflection response data segment and each candidate deflection response data sub-segment;

[0089] S35: The co-origin wave matching model calculates the matching probability between each candidate deflection response data segment and the reference deflection response data segment, wherein the candidate deflection response data segment corresponding to the maximum matching probability is the matching result of the reference deflection response data segment; specifically including the following:

[0090] S351: The waveform feature extraction layer, time difference feature extraction layer, and phase difference feature extraction layer respectively extract features from the waveform of each candidate deflection response data segment, as well as the time difference and phase difference between each candidate deflection response data segment and the reference deflection response data segment, to obtain a time-series feature vector. Time difference feature vector and phase difference eigenvector ;

[0091] S352: Transfer the time-series feature vector Time difference feature vector and phase difference eigenvector The features are concatenated at the feature fusion layer (concat) to form a multimodal feature vector. To achieve multimodal information fusion; the multimodal feature vector The expression is as follows: ;

[0092] This operation preserves the integrity of features from different sources to the greatest extent possible and provides a joint representation for the subsequent classification output layer that combines spatiotemporal characteristics with physical statistical significance.

[0093] S353: Multimodal Feature Vector After being input into the classification output layer, the classification output layer uses an activation function to process the multimodal feature vector. The mapping is a probability distribution vector; each component of the probability distribution vector corresponds to the confidence level of the reference response data segment and each candidate response data segment originating from the same heavy vehicle load; the candidate deflection response data segment corresponding to the maximum probability value in the probability distribution vector is the matching result of the same source wave matching task, that is, the deflection response of the target heavy vehicle on the second bridge matched with the reference deflection response data segment.

[0094] In one implementation, in order to improve the training effect of the convolutional neural network, in steps S12 and S32, the vehicle-induced deflection response and / or candidate deflection response are preprocessed by filtering, Z-score standardization, etc.; through Z-score standardization, the mean of the processed deflection data distribution is 0 and the standard deviation (and variance) is 1, thereby unifying the feature scale of different bridge deflection data.

[0095] Figure 5 This is a schematic diagram of a preferred multi-input convolutional neural network model. In this implementation, the waveform feature extraction layer includes sequentially connected convolutional modules (Conv1D+BN+pooling layers), convolutional block attention modules (CBAM), and long short-term memory networks (LSTM). Generally, the convolutional module includes a one-dimensional convolutional layer (Conv1D), a batch normalization layer (BN), and a pooling layer. Preferably, the convolutional module adopts a two-level convolutional unit structure, including: a first convolutional unit and a second convolutional unit connected in sequence. The first convolutional unit is composed of a first one-dimensional convolutional layer, a first batch normalization layer, and a first pooling layer connected in sequence. The second convolutional unit is composed of a second one-dimensional convolutional layer, a second batch normalization layer, and a second pooling layer connected in sequence. The output of the first pooling layer is connected to the input of the second one-dimensional convolutional layer.

[0096] The time difference feature extraction layer and / or phase difference feature extraction layer adopt a multilayer perceptron structure. The multilayer perceptron includes an input layer, a hidden layer and an output layer. The hidden layer and the output layer are usually implemented by fully connected layers (Dense). Each fully connected layer (Dense) includes a linear transformation unit and a nonlinear activation function.

[0097] Based on the above structure, waveform data is input into the waveform feature extraction layer through the first input branch. After local feature extraction by the convolutional module, a first feature map U is output. The convolutional block attention module (CBAM) receives the first feature map U output by the convolutional module, sequentially applies attention weights to the first feature map U in the channel and spatial dimensions, and outputs a second feature map U' to enhance the feature response of key channels and key time regions. Through this dynamic feature recalibration, the convolutional block attention module enables the model to adaptively focus on the most discriminative information, significantly improving feature quality. The long short-term memory network receives the second feature map U' output by the convolutional block attention module (CBAM), performs global information aggregation on the feature sequence of the second feature map U', and obtains the temporal feature vector of the waveform. .

[0098] Time difference data and phase difference data are input into two independent multilayer perceptrons for processing. The multilayer perceptrons use a hierarchical forward propagation structure to perform deep feature mining on the input time difference data and phase difference data to obtain time difference feature vectors. Phase difference eigenvectors .

[0099] In one implementation, the convolutional block attention module (CBAM) of the waveform feature extraction layer employs channel attention and spatial attention mechanisms, calculating attention weights sequentially along both channel and spatial dimensions, specifically including the following steps:

[0100] (1) Channel attention mechanism:

[0101] Global average pooling and global max pooling operations are performed on the first feature map U to generate a first channel description vector and a second channel description vector; the first channel description vector is used to characterize the global average response, and the second channel description vector is used to characterize the global maximum response.

[0102] The first channel description vector and the second channel description vector are respectively input into a shared multilayer perceptron (MLP), and the first channel description vector and the second channel description vector are respectively subjected to nonlinear transformation by the multilayer perceptron (MLP);

[0103] The first and second channel description vectors after nonlinear transformation are fused and processed by the Sigmoid activation function to generate a normalized channel weight matrix.

[0104] Multiply the channel weight matrix with the first feature map U channel by channel to obtain the channel-weighted intermediate feature map U. c ;

[0105] (2) Spatial attention mechanism:

[0106] For the intermediate feature map U c Perform average pooling and max pooling operations respectively to generate a first spatial feature description map and a second spatial feature description map;

[0107] The first spatial feature description map and the second spatial feature description map are spliced ​​together to obtain a multi-channel feature map;

[0108] The multi-channel feature map is input into a convolutional layer for feature extraction, and then processed by the Sigmoid activation function to generate a spatial weight matrix.

[0109] The spatial weight matrix and the intermediate feature map U are then compared. c Pixel-wise multiplication yields a second feature map U' with pixel weighting, highlighting the feature representation of key regions;

[0110] The calculation formulas corresponding to the above process are as follows:

[0111] ;

[0112] ;

[0113] In the formula, This represents the first feature map output by the convolutional module; Represents the channel weight matrix; Represents the intermediate feature map; Represents the spatial weight matrix; This represents the second feature map output by the attention module of the convolutional block;

[0114] The second feature map U' is further input as a temporal feature sequence into a Long Short-Term Memory (LSTM) network. Through the gating mechanism composed of the input gate, forget gate, and output gate of the LSTM network, the long-range dependencies in the temporal feature sequence are iteratively learned. The hidden state of the LSTM network at the last time step is extracted, or the hidden state sequence of the entire time step is aggregated to generate a temporal feature vector. , as the output of the waveform feature extraction layer;

[0115] In the above process, the Convolutional Block Attention (CBAM) module, through the combined action of channel attention and spatial attention mechanisms, can dynamically recalibrate the first feature map U, enabling the model to adaptively focus on the most discriminative channel and spatial information, and locate the most significant events (such as the critical moment when the vehicle axle load passes through), thereby significantly improving feature quality and model recognition performance. Both mechanisms work together to purify and focus the temporal feature sequence, which is then input into a Long Short-Term Memory (LSTM) network. The LSTM network is used to capture and memorize key temporal patterns in the input temporal feature sequence. This process is crucial for understanding the temporal dynamics of deflection data, such as identifying the entire process of a vehicle passing over a bridge, the interaction between different vehicle events, and the delayed and persistent effects of structural response. Its output temporal feature vector... As a deep temporal feature of this branch, it provides the core basis for subsequent fusion and classification decisions.

[0116] In one implementation, both the time difference feature extraction layer and the phase difference feature extraction layer employ a multilayer perceptron structure to perform deep feature mining on the input time difference data and phase difference data, respectively, to obtain time difference feature vectors. and phase difference eigenvector The specific process is as follows:

[0117] The time difference data are combined into a time difference input feature vector, which is then fed into the input layer of the first multilayer perceptron through a second input branch. The number of neurons in this input layer is equal to the dimension of the time difference input feature vector. A linear transformation is performed on the time difference input feature vector using a weight matrix and a bias vector to obtain a first net input. A nonlinear transformation is then performed on the first net input using a nonlinear activation function. The above linear and nonlinear transformations are iteratively performed in multiple hidden layers, ultimately generating a fixed-dimensional time difference feature vector rich in semantic information at the output layer of the first multilayer perceptron. This serves as the final representation of the time difference feature;

[0118] Similarly, the phase difference data is combined into a phase difference input feature vector, which is then fed into the input layer of the second multilayer perceptron through the third input branch. The number of neurons in this input layer is equal to the dimension of the phase difference input feature vector. A linear transformation is then performed on the phase difference input feature vector using a weight matrix and a bias vector to obtain the second net input. A nonlinear transformation is then performed on the second net input using a nonlinear activation function. These linear and nonlinear transformations are iterated across multiple hidden layers, ultimately generating a fixed-dimensional phase difference feature vector at the output layer of the second multilayer perceptron. ;

[0119] Time difference feature vector Phase difference eigenvectors The expression is as follows:

[0120] ;

[0121] ;

[0122] in, , These represent the ordinal indices of the hidden layers of the first and second multilayer perceptrons, respectively. The first multilayer perceptron represents the first... The time difference output of the hidden layer is used as the input feature vector. The second multilayer perceptron represents the first... The phase difference output from the hidden layer is used as the input feature vector; , Let represent the nonlinear activation functions of the first multilayer perceptron and the second multilayer perceptron, respectively; The first multilayer perceptron represents the first... The weight matrix of the hidden layer, The second multilayer perceptron represents the first... The second multilayer perceptron represents the first... The bias vector of the hidden layer;

[0123] The aforementioned time difference and phase difference data serve as important auxiliary inputs, which are processed using a multilayer perceptron to abstract and combine features layer by layer in order to uncover hidden high-order patterns and interactions in the input feature vector.

[0124] In one implementation, during the model training process, the multi-input convolutional neural network model uses cross-entropy as the loss function to calculate the loss value.

[0125] In one implementation, the classification output layer consists of two fully connected layers (Dense×2) and an activation function, preferably a softmax function.

[0126] In one implementation, multiple trained network models are evaluated based on at least one quantitative evaluation index, wherein the quantitative evaluation index is used to screen convolutional neural networks that meet the requirements to obtain reliable homologous wave matching models.

[0127] The quantitative evaluation metrics include, but are not limited to, accuracy, precision, recall, and F1 score;

[0128] Accuracy is defined as the proportion of samples correctly predicted by the model in all samples (including positive and negative examples). This evaluation metric measures the model's ability to identify both positive and negative examples.

[0129] Precision is defined as the proportion of all samples that the model predicts as positive, and that are actually positive. This evaluation metric measures the reliability of the model's prediction results.

[0130] Recall is defined as the proportion of samples that are actually positive that are correctly predicted as positive by the model. This evaluation metric measures the model's ability to find positive examples.

[0131] The F1 score is defined as the harmonic mean of precision and recall, and is used to comprehensively evaluate the performance of a model.

[0132] The following example, using a bridge group in a section of a provincial highway bridge health monitoring project in 2024, illustrates in detail the feasibility and effectiveness of this application.

[0133] The bridge group in this section consists of 8 bridges. The 8 bridges are located close to each other, have the same superstructure, and are of the same age. Most of the bridges are in good technical condition, but the traffic volume is high. Therefore, it is still necessary to use photoelectric deflectometers to monitor the vertical displacement (i.e., deflection) of the main beams of all bridges in order to understand the overall alignment and changes of the bridges along the entire route.

[0134] In this embodiment, 188 events from the bridge group in this section in January 2025 were selected as the deflection response feature dataset (containing 752 vertical displacement response waveforms). Figure 6 The process of generating category labels is given, in which the vertical displacement response of different loaded vehicles is labeled as 0, and the vertical displacement response of the same loaded vehicle is labeled as 1; 50% of the data in the deflection response feature dataset with category labels is randomly selected as the training set, 20% as the validation set, and 30% as the test set.

[0135] like Figure 7 , Figure 8 As shown, the accuracy of the first iteration of the co-origin wave identification model has reached 78%. With the increase of the iteration cycle, the loss value and accuracy on the training set and the validation set gradually converge and tend to stabilize. Finally, the training accuracy reaches 100% and the validation accuracy is 99.04%.

[0136] according to Figure 9 It can be seen that for the actual mismatched waveforms, the model predicts 3 waveforms as matching waveforms and 168 waveforms as mismatched waveforms; for the actual matching waveforms, the model predicts 53 waveforms as matching waveforms and 2 waveforms as mismatched waveforms.

[0137] The overall accuracy of the test set reached 97.79%; for the prediction of mismatched waveforms, the precision was 98.82%, the recall was 98.25%, and the F1 score was 98.53%; for the prediction of matched waveforms, the precision was 94.64%, the recall was 96.36%, and the F1 score was 95.50%.

[0138] The results show that the model can accurately identify matched and mismatched waveforms when faced with a large number of signals to be matched. This application provides a new approach for matching the deflection response of a bridge group under the same heavy vehicle load, which is a support and guarantee for subsequent response prediction and structural monitoring.

[0139] The above embodiments are only for illustrating the technical concept and features of this application, and are intended to enable those skilled in the art to understand the content of this application and implement it accordingly. They should not be used to limit the scope of protection of this application. For those skilled in the art, several improvements and modifications can be made without departing from the principles of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A deep learning-based method for matching deflection responses of a group of bridges from the same load source, characterized in that, The method comprises: collecting vehicle-induced deflection responses caused by vehicles passing through each bridge in a bridge group; identifying deflection responses of the same heavy vehicle passing through each bridge from the vehicle-induced deflection responses; extracting waveforms, time sequences and phases of the deflection responses of each heavy vehicle, calculating time differences and phase differences of the deflection responses of the same heavy vehicle passing through each bridge, and constructing a deflection response feature data set of the heavy vehicle based on the waveforms, time differences and phase differences; assigning a category label to the deflection response feature data set and constructing a deflection response feature data set with a category label; A multi-input convolutional neural network model is constructed; the multi-input convolutional neural network model comprises three input branches, each of which is connected to a waveform feature extraction layer, a time difference feature extraction layer and a phase difference feature extraction layer, respectively, the outputs of the waveform feature extraction layer, the time difference feature extraction layer and the phase difference feature extraction layer are connected to a feature fusion layer, respectively, and the output of the feature fusion layer is connected to a classification output layer; the deflection response feature data set with a category label is divided in proportion to obtain a training set, a validation set and a test set; the training set obtained by the division is input into the multi-input convolutional neural network model, and the network model parameters are trained and optimized, and the trained network model is used as a homologous wave matching model; the waveform feature extraction layer comprises a convolution module, a convolution block attention module and a long short-term memory network connected in sequence; wherein the convolution module comprises a first convolution unit and a second convolution unit connected in sequence, the first convolution unit is composed of a first one-dimensional convolution layer, a first batch normalization layer and a first pooling layer connected in sequence, the second convolution unit is composed of a second one-dimensional convolution layer, a second batch normalization layer and a second pooling layer connected in sequence, and the output of the first pooling layer is connected to the input of the second one-dimensional convolution layer; the waveform feature extraction layer outputs a first feature map U after extracting the features of the waveform through the convolution module; the convolution block attention module performs attention weighting on the first feature map U in the channel dimension and the spatial dimension in turn, and outputs a second feature map U'; the long short-term memory network performs global information aggregation on the feature sequence of the second feature map U' to obtain a time sequence feature vector of the waveform ; the time difference feature extraction layer and / or the phase difference feature extraction layer adopts a multi-layer perceptron structure; selecting a reference deflection response data segment and a candidate deflection response; according to the duration of the reference deflection response data segment, the candidate deflection response is cut into a plurality of candidate deflection response data subsegments; the homologous wave matching model calculates the matching probability of each candidate deflection response data subsegment and the reference deflection response data segment, wherein the candidate deflection response data subsegment corresponding to the maximum matching probability is the matching result of the reference deflection response data segment. 2.The deep learning-based same load source bridge group deflection response matching method according to claim 1, characterized in that, The method comprises: Definition 1 i The deflection response of the i-th heavy vehicle passing through the j-th bridge in the bridge group j The deflection response of the i-th heavy vehicle passing through the j-th bridge in the bridge group 、 The deflection response of the i-th heavy vehicle passing through the j-th bridge in the bridge group k The deflection response of the i-th heavy vehicle passing through the j-th bridge in the bridge group , , , , i denotes the heavy vehicle sequence index, j、k denotes the bridge sequence index, j ≠ k , N denotes the number of heavy vehicles, M denotes the number of bridges in the same bridge group; construct the waveform dataset of the deflection response of all heavy vehicles with respect to the current bridge group ; Definition 1 i The first heavy vehicle passes through the first bridge in the bridge group j The first bridge 、 The first k The timing of the deflection response caused by the first heavy vehicle passing through the first bridge in the bridge group , The time difference of the deflection responses of the first heavy vehicle passing through the first bridge and the second bridge in the bridge group i The first k The time difference of the deflection responses of the first heavy vehicle passing through the first bridge and the second bridge in the bridge group j The first The dataset of the time difference of the deflection responses of all heavy vehicles about the current bridge group ； Definition of the first i vehicle through the first j seat bridge 、 The phase of the deflection response caused by the first k seat bridge , The phase difference of the deflection responses of the first i vehicle through the first j seat bridge and the second k seat bridge is ; constructing a dataset of phase difference data for deflection response of all heavy vehicles with respect to the current bridge group ; constructing a dataset of deflection response characteristic data for the first i heavy vehicle with respect to the current bridge group , representing a dataset of deflection response characteristic data for the first i heavy vehicle with respect to the first j bridge and the second k bridge in the current bridge group. 3.The deep learning-based same load source bridge group deflection response matching method according to claim 2, characterized in that, The method comprises: Definition of the first i deflection response characteristic data set of the heavy vehicle with respect to the current bridge group category label is ; build a category label data set of the deflection response of all heavy vehicles with respect to the current bridge group ; constructing a dataset of deflection response signatures of all heavy vehicles with respect to the current bridge group's band class label . 4.The deep learning-based same load source bridge group deflection response matching method according to claim 1, characterized in that, The convolution block attention module sequentially performs attention weighting on the first feature map U in the channel dimension and the spatial dimension, and outputs a second feature map U', which comprises: performing a global average pooling operation and a global maximum pooling operation on the first feature map U respectively to generate a first channel description vector and a second channel description vector; the first channel description vector is used to represent a global average response, and the second channel description vector is used to represent a global maximum response; the first channel description vector and the second channel description vector are respectively input into a shared multilayer perceptron, and the first channel description vector and the second channel description vector are respectively subjected to nonlinear transformation through the multilayer perceptron; the first channel description vector and the second channel description vector after nonlinear transformation are fused and processed through an activation function to generate a normalized channel weight matrix; multiplying the channel weight matrix with the first feature map U channel by channel to obtain an intermediate feature map U after channel weighting c .

5. The deep learning-based same load source bridge group deflection response matching method according to claim 4, characterized in that, The convolution block attention module sequentially performs attention weighting on the first feature map U in the channel dimension and the spatial dimension, and outputs a second feature map U', which further comprises: For the intermediate feature map U c Perform average pooling and max pooling operations respectively to generate a first spatial feature description map and a second spatial feature description map; the first spatial feature description map and the second spatial feature description map are spliced to obtain a multi-channel feature map; the multi-channel feature map is input into a convolution layer for feature extraction and processed through an activation function to generate a spatial weight matrix; The spatial weight matrix is multiplied with the intermediate feature map U c Pixel by pixel multiplication is performed to obtain the pixel-weighted second feature map U'. 6.The deep learning-based same load source bridge group deflection response matching method according to claim 1, characterized in that, The method comprises: selecting a reference deflection response data segment and a candidate deflection response; according to the duration of the reference deflection response data segment, the candidate deflection response is cut into a plurality of candidate deflection response data subsegments; the homologous wave matching model calculates the matching probability of each candidate deflection response data subsegment and the reference deflection response data segment, wherein the candidate deflection response data subsegment corresponding to the maximum matching probability is the matching result of the reference deflection response data segment. selecting a reference deflection response data segment; the reference deflection response data segment is induced by the target loaded vehicle as it passes over a first bridge in the bridge group, and has a duration of t 0; The candidate deflection response is selected, and the candidate deflection response is preprocessed; the candidate deflection response is a to-be-matched deflection response induced when each type of vehicle passes through a second bridge in a bridge group, and a monitoring time length of the candidate deflection response is T , T > t 0; With t A sliding time window of equal length traverses the preprocessed candidate deflection response data sequence and extracts several candidate deflection response data segments. extracting the waveform, timing and phase of the reference deflection response data segment and each candidate deflection response data sub-segment, and then calculating the time difference and phase difference between the reference deflection response data segment and each candidate deflection response data sub-segment; The homologous wave matching model calculates the matching probability of each candidate deflection response data sub-segment and the reference deflection response data segment, wherein the candidate deflection response data sub-segment corresponding to the maximum matching probability is the matching result of the reference deflection response data segment.

7. The deep learning-based same load source bridge group deflection response matching method according to claim 1, characterized in that, The trained network model is evaluated by using a quantitative evaluation index, and the quantitative evaluation index includes at least one of accuracy, precision, recall and F1 score.

Images