Bridge support abnormity detection method

Abstract

The invention provides a bridge support anomaly detection method, and belongs to the technical field of bridge detection based on deep learning. According to the method, a high-precision sensor array is deployed at a key position of a support, and vibration, displacement, strain and environmental parameters are synchronously collected; a one-dimensional convolutional neural network and a self-attention mechanism are innovatively adopted to extract local and global features of a vibration signal, ST-GCN is utilized to capture spatial distribution and time evolution laws of displacement and strain, and an environment long-term trend is modeled through GRU; meanwhile, a multi-modal fusion method based on an adversarial auto-encoder is introduced, heterogeneous features are mapped to a unified and compact potential space, distribution of the heterogeneous features is more regular through adversarial training, and the essential distinguishing capacity of the model for normal and abnormal states is enhanced. According to the invention, comprehensive perception, accurate diagnosis and prospective early warning of the state of the support are realized.

Description

Technical Field

[0001] This invention belongs to the field of bridge inspection technology based on deep learning, and particularly relates to a method for detecting anomalies in bridge bearings. Background Technology

[0002] Bridge bearings, as key components for transmitting loads and adapting to deformation, endure complex dynamic and static stresses and environmental erosion over long periods. Performance degradation or sudden anomalies in bearings can lead to excessive structural displacement, vibration instability, or even localized damage. If not identified and addressed promptly, these issues will gradually accumulate into overall safety risks. With bridge monitoring technology evolving towards intelligence and systematization, traditional methods relying on manual inspections and single-parameter thresholds are no longer sufficient to address the diagnostic challenges of hidden bearing damage and the coupling effects of multiple factors.

[0003] The existing methods for testing bridge bearings are as follows: Chinese invention patent application number CN202411617015.2 discloses a real-time automated detection and positioning method for typical damage of bridge bearings, including: jacking the main beam, arranging ultrasonic sensors on the upper surface of the bearing, sending an initial excitation signal to obtain the first detection signal, using the time reversal method to convert the damage signal extracted from the first detection signal into a secondary excitation signal to obtain the corresponding inverted focusing signal, then calculating the damage characteristic index, judging the bearing damage according to the damage characteristic index, when the damage is judged, obtaining the probability value of damage at the point on the upper surface of the bearing by establishing a probabilistic damage model, and performing bearing damage imaging based on the probability value; Chinese invention patent application number CN202311728169.4 relates to a bridge bearing deformation detection mechanism, including a base and a support box fixedly connected to the top of the base. A detection mechanism is provided on one side of the support box, and the detection mechanism includes a bonding plate disposed on one side of the support box. A detection piece is fixedly connected to the side of the bonding plate away from the support box. A support rod is fixedly connected to the top of the base. This bridge bearing deformation detection mechanism, by setting the bonding plate and the detection piece, can detect the deformation of the bridge bearing. Furthermore, by setting a display, the detection results are displayed on the screen, providing a more intuitive view of the detected structure for the operator. This eliminates the need for operators to use measuring rulers, improving work efficiency. Additionally, the inclusion of a rotating shaft allows the display to rotate, facilitating viewing from different angles and expanding the device's usability in various environments.

[0004] However, the above method has the following limitations: 1) Limited detection methods make it difficult to comprehensively perceive complex fault modes. Existing technologies mostly rely on single types of sensors, which can only capture information in one dimension of the support state and cannot simultaneously acquire multi-physical field data such as vibration, strain, and displacement. This results in insufficient ability to identify complex anomalies such as potential fatigue damage, micro-slippage, and stress redistribution. 2) Weak intelligent analysis and early warning capabilities. Existing methods mostly rely on preset thresholds or simple signal processing algorithms for damage identification, lacking the ability to perform deep feature mining and pattern recognition on multi-source, heterogeneous time-series data. The model cannot adaptively learn the complex dynamic behavior differences of the support under normal and abnormal states, resulting in a high rate of missed and false alarms for early, weak, or novel abnormal patterns. 3) Insufficient consideration of environmental interference factors leads to poor model robustness. Bridge bearings are exposed to changing environmental conditions such as temperature and humidity for extended periods, which can interfere with their response signals. Existing technical solutions generally lack simultaneous monitoring and modeling of environmental parameters, failing to effectively isolate the impact of environmental fluctuations in the analysis. This makes the test results prone to bias, and the stability and reliability under different seasons or climatic conditions are difficult to guarantee. 4) Lack of continuous learning and adaptive optimization mechanisms. Once deployed, existing systems typically employ static detection models that cannot self-update using the constantly generated new data during online monitoring. When bridge bearings experience material aging, changes in load patterns, or the occurrence of unprecedented anomalies, the performance of the static model gradually declines, failing to meet the requirements of continuous model evolution for bridge lifecycle health monitoring. Summary of the Invention

[0005] To address the aforementioned issues, this invention proposes a method for detecting anomalies in bridge bearings. In this method, a high-precision sensor array is deployed at key locations on the bearing to simultaneously collect vibration, displacement, strain, and environmental parameters. An intelligent diagnostic model that integrates deep adversarial autoencoders and deep support vector data descriptions is then constructed to achieve comprehensive perception, accurate diagnosis, and proactive early warning of the bearing's condition.

[0006] This invention provides a method for detecting anomalies in bridge bearings, comprising the following steps: S1, synchronously collect vibration acceleration dataset, displacement dataset, strain dataset, and environmental parameter dataset at key locations of bridge bearings, and perform data preprocessing; S2, input the four preprocessed datasets into the bridge bearing feature extraction module; A vibration depth feature extraction network based on a one-dimensional convolutional neural network and multi-head self-attention extracts features from vibration acceleration data to obtain depth vibration features; a displacement-strain joint feature extraction network based on spatiotemporal graph convolution extracts features from displacement and strain data to obtain displacement-strain joint spatiotemporal features; an environment-multimodal fusion feature extraction network based on gated recurrent and cross-modal attention extracts features from environmental parameter data to obtain long-term environmental dependence features; a cross-modal attention fusion network fuses depth vibration features, displacement-strain joint spatiotemporal features, and long-term environmental dependence features to obtain preliminary fused multimodal features; S3. The pre-fused multimodal features are input into the bridge bearing anomaly detection model. First, multimodal deep fusion is performed through the bridge bearing multimodal adversarial autoencoder to obtain latent spatial features. Then, the latent spatial features are input into the hybrid anomaly detection model. The anomaly score is calculated by combining the deep support vector data description module and the bridge bearing temporal anomaly capture convolutional network. The anomaly scores are then fused to output whether the bearing status is abnormal and the anomaly type.

[0007] Preferably, the vibration acceleration dataset is the vibration acceleration time series data of the support under different loads such as vehicle load and environmental excitation, which reflects the dynamic response characteristics of the support structure and the vibration anomalies caused by potential damage; The displacement dataset consists of the displacement and deformation time series data of the support in both vertical and horizontal dimensions, monitoring the displacement offset and settlement deformation status of the support. The strain dataset is time-series data of strain distribution in the support, used to assess the stress concentration and fatigue damage evolution within the support. The environmental parameter dataset consists of time-series environmental data for the monitored area, including temperature and humidity.

[0008] Preferably, the vibration depth feature extraction network is specifically: The preprocessed vibration acceleration time-series data is reconstructed into a one-dimensional sequence feature input to a one-dimensional convolutional neural network for bridge bearing vibration feature extraction. This network comprises four convolutional blocks and three pooling layers. Each convolutional block consists of a one-dimensional convolutional layer, a batch normalization layer, and an ELU activation function. The kernel sizes are set to 3, 5, 7, and 9, respectively, progressively expanding the receptive field to capture local vibration features at different scales. Max pooling is used in the pooling layers to achieve feature dimensionality reduction and abstraction. An adaptive vibration frequency enhancement module is embedded after each convolutional block. By learning the frequency distribution characteristics of the vibration signal, the feature weights in different frequency ranges are dynamically adjusted to obtain the local vibration features. ; Local vibration characteristics The input multi-head self-attention mechanism module captures the global dependencies of vibration signals in the time dimension through multiple parallel attention heads, with each attention head focusing on the correlation of vibration modes in different time periods. The multi-head self-attention mechanism calculates global vibration features through the following steps: First, each attention head calculates its attention weight by performing a dot product operation on the query vector, key vector, and value vector. Then, the value vectors are weighted and summed to obtain the output of a single attention head. Next, the outputs of all attention heads are concatenated, and multi-head information is fused through a linear projection layer to finally obtain the global vibrational features. ; Local vibration characteristics With global vibration characteristics The features are then stitched and fused together using a fully connected layer and a BatchNorm layer to obtain the deep vibration features. .

[0009] Preferably, the displacement-strain joint feature extraction network is specifically: The preprocessed displacement data is fused with deformation time series data and strain distribution time series data to construct displacement-strain joint data. And based on the spatial deployment of sensors, a spatiotemporal map structure of bridge supports is constructed. , where the set of nodes For each sensor deployment location on the corresponding support, each node contains displacement and strain data for that location; edge set The feature matrix is ​​constructed based on the spatial distance and mechanical relationship between nodes, with larger edge weights between nodes that are closer and have shorter mechanical transmission paths; Displacement-strain combined data It consists of spatiotemporal data information of all nodes; spatiotemporal diagram structure The input spatiotemporal graph convolutional network for bridge bearings consists of four spatiotemporal graph convolutional blocks and two temporal pooling layers. Each spatiotemporal graph convolutional block comprises a spatial graph convolutional submodule and a temporal convolutional submodule. The spatial graph convolutional submodule employs graph convolution operations with a graph attention mechanism. This mechanism calculates a dynamically changing attention weight for each pair of sensor nodes connected by edges in the graph. This weight value originates from the node's current feature vector. The node's current feature vector is first transformed and activated by a neural network layer, then the feature similarity between the two is calculated, and finally, it is obtained through normalization. During spatial information aggregation, the features of each neighboring node are multiplied by its corresponding dynamic weight and then weighted and fused. The temporal convolution submodule employs one-dimensional causal convolution operations to capture the temporal evolution of the same node at different times. Its convolution kernel slides sequentially along the time axis for computation: when calculating the output at any given time, only the input data of that time and all previous historical times are allowed to be used; by stacking multiple such convolutional layers, the field of view of the convolution kernel on the time axis is increased layer by layer. Temporal pooling layers are embedded between the spatiotemporal graph convolutional blocks, and adaptive temporal pooling is used to reduce the dimensionality of temporal features while preserving feature information of key time nodes. The output of the spatiotemporal graph convolutional network for bridge supports aggregates the spatiotemporal features of all nodes through a global average pooling layer to obtain joint displacement-strain spatiotemporal features. .

[0010] Preferably, the environment-multimodal fusion feature extraction network is specifically: Normalized environmental time-series data is input into the gated recurrent unit for bridge bearing environmental feature extraction. This unit utilizes its internal update and reset gate mechanisms to regulate the fusion of historical information and current information. Through the gating mechanism, the long-term temporal dependencies of environmental parameters are adaptively captured. Finally, the final hidden state of the gated recurrent unit for bridge bearing environmental feature extraction incorporates the contextual information of the entire input sequence, and the final hidden state represents the long-term environmental dependency features. .

[0011] Preferably, the cross-modal attention fusion network is specifically: First, each modal feature is mapped to the same dimension through a two-layer fully connected network. Then, the importance score is obtained by dot product with the shared projection vector. Finally, the attention weight coefficients of the depth vibration features are obtained by normalization using the softmax function. Displacement-strain joint spatiotemporal feature attention weight coefficient Long-term environmental dependence features and attention weight coefficients ; Multimodal feature fusion based on attention weight coefficients yields preliminary fused multimodal features. .

[0012] Preferably, the bridge bearing multimodal adversarial autoencoder consists of three parts: an encoder, a decoder, and a discriminator. The encoder employs a multilayer perceptron structure, introducing a modality adaptive normalization layer to dynamically adjust the distribution of features from different modalities and calculate the latent spatial features. ; The decoder employs a multilayer perceptron structure symmetrical to the encoder, which integrates latent spatial features. The original multimodal feature space is reconstructed and used to calculate the reconstruction loss, ensuring that the latent features can retain the key information of the original multimodal data; The discriminator employs a convolutional neural network structure to distinguish latent spatial features. By using adversarial training with samples of a pre-defined prior distribution, the latent feature distribution generated by the encoder is forced to approximate the prior distribution, thereby enhancing the generalization ability and discriminative power of the features.

[0013] Preferably, the hybrid anomaly detection model includes a deep support vector data description module and a temporal anomaly scoring network module; The deep support vector data description module learns a minimum hypersphere that tightly encloses the fused features of the normal state within the sphere, while the features of the abnormal state fall outside the sphere. First, the center of the hypersphere is learned using samples labeled as normal states in the training set. With radius , center of the ball Initialized to the mean of normal sample features, then iteratively optimized through backpropagation, radius Determined based on the distance distribution from normal sample features to the center of the sphere; for input multimodal deep fusion features Calculate its distance to the center of the ball. The distance is used as part of the outlier score; the final outlier score is obtained. ; The temporal anomaly scoring network module employs the Bridge Abutment Temporal Anomaly Capture Convolutional Network BBTAC-TCN, which contains four causal convolutional blocks. Each causal convolutional block consists of a one-dimensional dilated convolutional layer, a batch normalization layer, a ReLU activation function, and a temporal attention submodule. The dilation coefficient increases progressively to capture dependencies at different time scales. The temporal attention submodule strengthens the feature responses during anomalous periods by calculating the importance weights of features at different time steps, resulting in the attention-weighted version of the first anomaly score. Features at each time step BBTAC-TCN models the weighted time-series features and calculates the anomaly probability at each time step. Thus, abnormal scores are obtained. .

[0014] Preferably, the fusion anomaly score outputs whether the support status is abnormal and the anomaly type. The specific process is as follows: The hybrid anomaly detection model is optimized by a joint loss function, which includes the cross-entropy loss of the deep support vector data description module and the temporal anomaly scoring network module. By minimizing the total loss function through multi-task learning and optimizing the model parameters, the anomaly scores from the two modules are combined to obtain the final anomaly score. ; Set an abnormal score threshold ,when When the support condition is deemed abnormal, it is further classified into displacement exceeding limits, abnormal vibration frequency, or sudden strain change based on the characteristic response pattern; when When the condition is normal, the support is determined to be in normal condition.

[0015] Compared with the prior art, the present invention has the following innovative features and beneficial effects: (1) Achieving multi-dimensional collaborative perception and diagnosis of support status: Existing technologies mostly rely on single-type sensors, which can only capture single-dimensional information such as local damage or macroscopic deformation of the support, making it difficult to comprehensively reflect its health status. This invention constructs a multi-modal perception system by deploying an array of sensors for vibration, displacement, strain, and environmental parameters. It innovatively uses a one-dimensional convolutional neural network and a self-attention mechanism to extract local and global features of vibration signals, uses ST-GCN to capture the spatial distribution and temporal evolution of displacement and strain, and models long-term environmental trends through GRU. This multi-dimensional, collaborative perception and feature extraction method can comprehensively capture comprehensive information from microscopic material fatigue to macroscopic structural deformation, thereby achieving accurate diagnosis of complex anomalies and solving the fundamental problems of single detection perspective and many blind spots in existing technologies; (2) Enhancing the intelligence and robustness of anomaly detection under complex working conditions: Existing methods often rely on preset thresholds or simple signal processing for detection, which are easily affected by environmental fluctuations, resulting in high false alarm and false negative rates under complex working conditions. This invention innovatively introduces a multimodal fusion method based on adversarial autoencoders (AAEs), mapping heterogeneous features to a unified and compact latent space, and making its distribution more regular through adversarial training, thereby enhancing the model's ability to distinguish between normal and abnormal states. Meanwhile, in the anomaly detection module, Deep SVDD, which focuses on global description, is combined with TCN networks, which focus on local temporal anomalies, and a joint loss function is designed for multi-task learning. This design enables the model to not only judge state deviations from the overall perspective, but also to keenly capture local anomaly patterns such as instantaneous changes, significantly improving the intelligence and robustness of detection under noisy environments and variable loads; (3) Overcoming the limitations of static models and endowing the system with continuous evolution and adaptability: Once deployed, the model parameters of existing detection systems are usually fixed and cannot adapt to the state changes of bridge bearings due to long-term service, material aging, or the emergence of unprecedented abnormal modes, resulting in the degradation of model performance over time. This invention constructs an adaptive optimization platform based on real-time feedback and innovatively introduces a meta-learning framework. When the system discovers new anomalies or performance deviations by comparing the model output with the results of manual inspection, it can quickly fine-tune the model using small samples of new data collected online; (4) Promoting the strategic transformation of bridge operation and maintenance towards predictive maintenance: Existing technologies mostly focus on post-event detection or periodic inspections, making it difficult to achieve early warning and prediction of faults. This invention constructs a closed-loop system integrating real-time perception, intelligent diagnosis, and forward-looking early warning by deeply integrating multimodal sensor data and cutting-edge deep learning models. It can not only detect anomalies in a timely manner, but also capture early signs of bearing performance degradation through continuous analysis of deep features such as vibration modes and strain trends. Combined with a visual interface and alarm system, it provides maintenance personnel with sufficient decision support and response time, thereby transforming the operation and maintenance mode from passive response to proactive intervention, effectively avoiding safety accidents, extending the service life of bearings, and realizing the upgrade of bridge operation and maintenance management mode. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the overall technical route of the present invention.

[0017] Figure 2 This is a structural diagram of a bridge bearing anomaly detection model based on multimodal fusion and hybrid detection.

[0018] Figure 3 This is a diagram of the overall system framework of the present invention.

[0019] Figure 4 The dataset shown in this example contains a partial sample.

[0020] Figure 5 This is a comparison chart of the recall rates of abnormal features in the examples.

[0021] Figure 6 This is a comparison chart of feature discriminant performance in the embodiments.

[0022] Figure 7 This is a comparison chart of anomaly type recognition rates in the examples. Detailed Implementation

[0023] This invention proposes a method for detecting anomalies in bridge bearings, the overall process of which is as follows: Figure 1 As shown: S1. Firstly, addressing the problem that single sensor data is insufficient to comprehensively perceive the health status of bridge bearings in condition monitoring, this invention deploys arrays of vibration acceleration sensors, displacement sensors, strain sensors, and environmental parameter sensors at key locations on the bridge bearings to collect multi-source time-series data. The collected data is then cleaned, timestamped, and standardized, and normal and abnormal states are labeled based on design specifications and expert experience. Finally, the vibration acceleration dataset, displacement dataset, strain dataset, environmental parameter dataset, and state label set are integrated to construct a multimodal dataset for bridge bearings. In S2, the multimodal dataset obtained in S1 is input into the feature extraction module based on multimodal collaboration. First, deep vibration features are extracted through a one-dimensional convolutional neural network and a multi-head self-attention mechanism for bridge bearing vibration feature extraction. Second, displacement-strain joint spatiotemporal features are extracted through a spatiotemporal graph convolutional network for bridge bearings. Then, long-term environmental dependency features are extracted through a gated recurrent unit for bridge bearing environmental feature extraction. Finally, the three types of features are dynamically weighted and fused through a cross-modal attention fusion module to form preliminary fused multimodal features. S3. The preliminary fused multimodal features obtained in S2 are input into the anomaly detection model based on multimodal fusion and hybrid detection. First, multimodal deep fusion is performed through the bridge bearing multimodal adversarial autoencoder to obtain latent spatial features. Then, the deep fused features are input into the hybrid anomaly detection model. The anomaly score is calculated by combining the deep support vector data description module and the bridge bearing temporal anomaly capture convolutional network. The final anomaly score is then fused and output to determine whether the bearing status is abnormal and the type of anomaly. S4. After model deployment, a real-time feedback closed-loop system is built. The detection deviation is calculated by comparing the model prediction results with the actual state. When the deviation exceeds the threshold, optimization is triggered. A bridge bearing anomaly adaptation meta-learning framework is introduced to quickly fine-tune the model based on small sample labeled data and update the model parameters. The continuous optimization mechanism enables the model to adapt to new anomaly patterns and environmental changes. S5, during system implementation, involves installing a multi-source sensor array according to the deployment plan and connecting it to the monitoring center server via a data acquisition unit and communication module; deploying a software system integrating preprocessing, feature extraction, anomaly detection, and adaptive optimization modules; achieving real-time data acquisition, preprocessing, feature extraction, anomaly detection, and result display; and triggering an early warning mechanism when an anomaly is detected to send warning information to maintenance personnel.

[0024] The specific implementation process of the present invention will be further described below with reference to specific embodiments.

[0025] I. Construction of a Multimodal Dataset for Bridge Bearings Multi-source sensor deployment and data acquisition: High-precision sensor arrays are deployed at key locations of bridge bearings to achieve comprehensive dynamic perception of bearing status. Key locations of bridge bearings include the bearing contact surface, edge stress concentration area, and key vertical displacement monitoring points. Vibration Acceleration Sensor (VAS) is used to collect real-time vibration acceleration data of the bearing under different loads such as vehicle loads and environmental excitations, reflecting the dynamic response characteristics of the bearing structure and vibration anomalies caused by potential damage. Displacement Sensor (DS) uses laser displacement measurement to collect real-time displacement and deformation data of the bearing in both vertical and horizontal dimensions, monitoring the bearing's displacement and settlement deformation. Strain Sensor (SS) uses fiber optic sensing technology, deployed along the critical path of bearing stress distribution, to collect real-time strain distribution data of the bearing, assessing the stress concentration and fatigue damage evolution within the bearing. Environmental Parameter Sensor (EPS) simultaneously collects time-series data on temperature and humidity in the monitoring area, providing data support for eliminating the interference of environmental factors on the detection results. The sampling frequencies of the four types of sensors are uniformly set according to the dynamic response characteristics and monitoring requirements of the bridge bearing, ensuring the synchronization and integrity of data acquisition. The collected raw data are recorded as vibration acceleration datasets. Displacement dataset Strain dataset Environmental parameter dataset .

[0026] Data preprocessing and annotation: Systematic preprocessing was performed on the four types of multi-source time-series data collected. During data cleaning, the 3σ criterion was used to remove outliers from various sensor data. Missing values ​​generated during data acquisition were filled using linear interpolation to ensure data continuity. Based on a unified clock synchronization mechanism, [the following was also performed]. , , , Precise timestamp alignment was performed to establish a one-to-one correspondence between different modal data at the same time. The aligned data types were then standardized to eliminate the impact of dimensional differences. Based on bridge bearing design specifications and maintenance technical standards, the preprocessed dataset was labeled with status information, categorized into normal and abnormal states. Abnormal states were further divided into three types: displacement exceeding limits, abnormal vibration frequency, and sudden strain change. Displacement exceeding limits was determined by the bearing displacement exceeding the design allowable range; abnormal vibration frequency was determined by the vibration signal's dominant frequency deviating from the normal operating range; and sudden strain change was determined by the instantaneous rate of change of strain data exceeding a set threshold. This process resulted in a labeled multimodal dataset. ,in The corresponding status label set.

[0027] Multimodal dataset integration: integrating labeled vibration acceleration datasets Displacement dataset Strain dataset Environmental parameter dataset With status label set The data samples were structured and integrated in chronological order. Each sample contained four modalities of data within the same time window and their corresponding state labels. This constructed a multimodal dataset of bridge bearings for training, validation and testing of anomaly detection models, providing high-quality data support for subsequent deep feature extraction and anomaly detection model construction.

[0028] II. Design of Bridge Support Feature Extraction Module Based on Multimodal Collaboration 1. Vibration depth feature extraction network based on one-dimensional convolutional neural network and multi-head self-attention Vibration acceleration dataset The vibration acceleration time series data were preprocessed. First, wavelet threshold denoising was used to suppress environmental noise and measurement noise. The calculation formula for wavelet threshold denoising is as follows: ; in Represents the wavelet coefficients after denoising. This represents the first wavelet transform of the original vibration signal. Layer Wavelet coefficients, The wavelet threshold, dynamically calculated based on the signal-to-noise level using an adaptive noise estimation method, is used to distinguish between signal coefficients and noise coefficients. After denoising, outliers are removed again using the 3σ criterion to ensure data reliability, followed by standardization.

[0029] The preprocessed vibration acceleration time-series data is reconstructed into a one-dimensional sequence feature input to a one-dimensional convolutional neural network (BBVFE-1D-CNN) for bridge bearing vibration feature extraction. This network is specifically designed for bridge bearing vibration signals and contains four convolutional blocks and three pooling layers. Each convolutional block consists of a one-dimensional convolutional layer, a batch normalization layer, and an ELU activation function. The convolutional kernel sizes are set to 3, 5, 7, and 9, respectively, progressively expanding the receptive field to capture local vibration features at different scales. The pooling layers use max pooling (pooling kernel size 2×2) to achieve feature dimensionality reduction and abstraction. To enhance the ability to capture anomalous frequency components, an adaptive vibration frequency enhancement module is embedded after each convolutional block. This module learns the frequency distribution characteristics of the vibration signal, dynamically adjusts the feature weights of different frequency ranges, highlights the feature responses corresponding to anomalous frequencies, and improves the discriminative power of local features.

[0030] Local vibration characteristics output by BBVFE-1D-CNN The input module is a Multi-Head Self-Attention (MHSA) mechanism. This module captures the global dependencies of vibration signals over time using multiple parallel attention heads, each focusing on the correlation of vibration modes at different time periods. The MHSA mechanism calculates global vibration features through the following steps: First, each attention head calculates its attention weight by performing a dot product operation on the query vector, key vector, and value vector. Then, the value vectors are weighted and summed to obtain the output of a single attention head. Next, the outputs of all attention heads are concatenated, and the multi-head information is fused through a linear projection layer to finally obtain the global vibration features. The calculation formula is as follows: ; ; ; ; ; in , , They represent the first The query vector, key vector, and value vector of each attention head. , , They represent the first The learnable weight matrix for each attention head. This represents the dimension of the query vector and the key vector. Indicates the number of attention heads. This indicates a feature concatenation operation. This represents the concatenated linear projection weight matrix. This represents the global vibrational characteristics of the multi-head self-attention mechanism output.

[0031] Local vibration features extracted by BBVFE-1D-CNN Global vibrational characteristics of multi-head self-attention mechanism output The features are then stitched and fused together using a fully connected layer and a BatchNorm layer to obtain the deep vibration features. This feature contains both local anomaly frequency components and global time dependencies, providing strong discriminative support for subsequent anomaly detection.

[0032] 2. Displacement-Strain Joint Feature Extraction Network Based on Spatiotemporal Graph Convolution For displacement dataset and strain dataset The displacement and deformation time-series data, as well as the strain distribution time-series data, are preprocessed. First, the data are spatially aligned according to the spatial deployment coordinates of the sensors to ensure the spatial correspondence of data from sensors at different locations. A sliding window method is then used to segment the aligned time-series data, with the window size set according to the time evolution period of the support deformation. Each window corresponds to a time segment of displacement-strain data. Normalization is then performed. Finally, the displacement and deformation time-series data, along with the strain distribution time-series data, are fused using feature channel stitching. That is, at the same time step, the preprocessed displacement and deformation time-series data are directly stitched together along the feature dimension to construct joint displacement-strain data. ,in Indicates the first Each time step Indicates the first Normalized displacement eigenvectors at each time step Indicates the first Normalized strain feature vectors at each time step, and a spatiotemporal diagram structure of bridge supports is constructed based on sensor spatial deployment. , where the set of nodes For each sensor deployment location on the corresponding support, each node contains displacement and strain data for that location; edge set The feature matrix is ​​constructed based on the spatial distance and mechanical relationship between nodes, with larger edge weights between nodes that are closer and have shorter mechanical transmission paths; Displacement-strain combined data It consists of spatiotemporal data information of all nodes.

[0033] spatiotemporal diagram structure The input is the Bridge Support Spatiotemporal Graph Convolutional Network (BBS-TGCN), designed specifically to capture the spatiotemporal dependencies of support displacement and strain. It comprises four spatiotemporal graph convolutional blocks and two temporal pooling layers. Each spatiotemporal graph convolutional block consists of a spatial graph convolutional submodule and a temporal convolutional submodule. The spatial graph convolutional submodule employs a graph convolution operation improved with a graph attention mechanism, calculated as follows: ; in Indicates the first The layer in the first Output of spatial characteristics at time step. Represents a node The set of neighboring nodes, Represents a node With nodes The attention weights between nodes are calculated using node feature similarity and are used to enhance feature transmission for key spatial connections. Indicates the first The weight matrix of layer-space graph convolution. Indicates the first Layer nodes eigenvectors, This represents the corresponding bias vector. This represents the ELU activation function.

[0034] The temporal convolution submodule uses one-dimensional causal convolution operations to capture the temporal evolution of the same node at different times. The calculation formula is as follows: ; in Indicates the first Layer nodes The time-series feature output, This represents a one-dimensional causal convolution operation. Indicates the first The weight matrix of the time-series convolution. This represents the corresponding bias vector.

[0035] Temporal pooling layers are embedded between the convolutional blocks of the spatiotemporal graph, and adaptive temporal pooling is used to reduce the dimensionality of temporal features while preserving feature information of key time nodes. The output of BBS-TGCN aggregates the spatiotemporal features of all nodes through a global average pooling layer to obtain the displacement-strain joint spatiotemporal features. This feature comprehensively depicts the spatial distribution pattern and temporal evolution law of support deformation, providing a spatiotemporal dimension for anomaly detection.

[0036] 3. Environment-Multimodal Fusion Feature Extraction Network Based on Gated Recurrent and Cross-Modal Attention Environmental parameter dataset The temperature and humidity time-series environmental data were preprocessed. First, a moving average filter was used to eliminate seasonal fluctuations and random noise, followed by normalization. The normalized environmental time-series data was then input into a gated recurrent unit (BBEF-GRU) for bridge bearing environmental feature extraction. This unit adaptively captures the long-term temporal dependencies of environmental parameters through a gating mechanism, mitigating the gradient vanishing problem. The final hidden state of the BBEF-GRU represents the long-term environmental dependency features. .

[0037] 4. Cross-modal attention fusion network Deep vibration characteristics Displacement-strain combined spatiotemporal characteristics Long-term dependence on the environment The input is a cross-modal attention fusion network, which dynamically calculates the weight coefficients of different modal features, highlighting modal information that contributes highly to anomaly detection and suppressing redundant information. First, each modal feature is mapped to the same dimension through a two-layer fully connected network, then its importance score is obtained by dot product with a shared projection vector. Finally, the attention weight coefficients of the depth vibration features are obtained by normalization using a softmax function. Displacement-strain joint spatiotemporal feature attention weight coefficient Long-term environmental dependence features and attention weight coefficients .

[0038] Multimodal feature fusion is performed based on attention weight coefficients, and the calculation formula is as follows: ; in This represents the multimodal features after initial fusion. These features integrate key information from three modes—vibration, displacement-strain, and environment—through dynamic weighting, effectively reducing redundancy and improving the discriminative power of the feature representation, thus laying the foundation for subsequent deep multimodal fusion.

[0039] III. Construction of Bridge Bearing Anomaly Detection Model Based on Multimodal Fusion and Hybrid Detection The overall structure of the model is as follows Figure 2 As shown: 1. Multimodal deep fusion based on adversarial autoencoders The pre-fused multimodal features The input bridge bearing multimodal adversarial autoencoder (BBM-AAE) maps multimodal features to a unified latent space, achieving deep fusion and enhancing feature discriminativity. BBM-AAE consists of an encoder, decoder, and discriminator. The encoder employs a multilayer perceptron (MLP) structure, introducing a modality adaptive normalization layer to dynamically adjust the distribution of different modal features, ensuring consistency of features in the latent space. The encoder's calculation formula is as follows: ; in Indicates potential spatial characteristics, Indicates encoder, , This represents the weight matrix of the encoder. , This represents the corresponding bias vector. Represents the ReLU activation function. This represents the hyperbolic tangent activation function. The modality adaptive normalization layer is embedded between two fully connected layers and dynamically adjusts the normalization parameters according to the modality distribution of the input features.

[0040] The decoder employs an MLP structure symmetrical to the encoder, which incorporates latent spatial features. The original multimodal feature space is reconstructed and used to calculate the reconstruction loss, ensuring that the latent features can retain the key information of the original multimodal data. The calculation formula for the decoder is as follows: ; in This represents the reconstructed multimodal features. Indicates decoder, , This represents the weight matrix of the decoder. , This represents the corresponding bias vector.

[0041] The discriminator employs a convolutional neural network structure to distinguish latent spatial features. By using samples with a pre-defined prior distribution, adversarial training forces the encoder to generate a latent feature distribution that approximates the prior distribution, thereby enhancing the generalization ability and discriminative power of the features. The loss function of the discriminator and the formula for calculating the adversarial loss of the encoder are as follows: ; in Indicating resistance to loss, This represents the output probability of the discriminator. This represents the pre-defined Gaussian prior distribution. This represents the distribution of initially fused multimodal features. This indicates the expected operation.

[0042] The total loss function of BBM-AAE is a weighted sum of reconstruction loss and adversarial loss, calculated as follows: ; in This represents the total loss function of BBM-AAE. The weights representing the reconstruction loss are used to balance reconstruction accuracy and adversarial training performance. This represents the square of the L2 norm. The BBM-AAE is trained by minimizing the total loss function, and the final encoder outputs the latent spatial features. This is a multimodal deep fusion feature, which deeply integrates information from different modalities and has strong discriminative and generalization capabilities.

[0043] 2. Hybrid anomaly detection based on deep support vector data description and time-series scoring A hybrid anomaly detection model for bridge bearings was constructed. This model consists of a Deep Support Vector Data Description (Deep SVDD) module and a temporal anomaly scoring network, with multimodal deep fusion features as input. .

[0044] The Deep SVDD module is specifically designed for modeling the normal state of bridge bearings. It learns a minimum hypersphere that tightly encloses the fused features of the normal state within the sphere, while the features of the abnormal state fall outside the sphere. First, the center of the hypersphere is learned using samples labeled as normal states in the training set. With radius , center of the ball Initialized to the mean of normal sample features, then iteratively optimized through backpropagation, radius Determined based on the distance distribution from normal sample features to the center of the sphere. For the input multimodal deep fusion features... Calculate its distance to the center of the ball. The distance, as part of the outlier score, is calculated using the following formula: ; in This represents the anomaly score output by the Deep SVDD module, when... At that time, it was initially determined that the support state corresponding to this feature was abnormal; when At that time, it was initially determined to be normal.

[0045] The temporal anomaly scoring network employs the Bridge Support Temporal Anomaly Capture Convolutional Network (BBTAC-TCN), an improvement upon the Temporal Convolutional Network (TCN) by introducing a temporal attention module to focus on feature changes during key anomaly periods. BBTAC-TCN comprises four causal convolutional layers, each consisting of a one-dimensional dilated convolutional layer, a batch normalization layer, a ReLU activation function, and a temporal attention submodule. The dilation coefficients increase progressively (1, 2, 4, 8) to capture dependencies at different time scales. The temporal attention submodule enhances the feature response during anomaly periods by calculating importance weights for features at different time steps. The calculation formula is as follows: ; ; in Indicates the first Attention weight coefficients for each time step The attention scoring function is implemented using a single fully connected layer. Indicates the length of the time series. Indicates the first The fusion features of each time step Indicates the attention-weighted number of... Each time step feature.

[0046] BBTAC-TCN models the weighted temporal features and outputs the anomaly probability at each time step, calculated using the following formula: ; ; in Indicates the first The anomaly probability corresponding to each time step. This represents the Sigmoid activation function. Indicates will Input BBTAC-TCN for processing. This represents the anomaly score output by the time-series anomaly scoring network; a higher score indicates a greater likelihood of an anomaly.

[0047] We design a joint loss function to optimize the hybrid anomaly detection model. This loss function includes the sphere loss of Deep SVDD and the cross-entropy loss of the temporal anomaly scoring network. The calculation formula is as follows: ; in This represents the total loss function of the hybrid anomaly detection model. This indicates the number of samples in the normal state. This represents the fusion feature set of samples in a normal state. The weights representing the cross-entropy loss are... This represents the total number of samples. Indicates the first The true label of each time step (1 for normal state, 0 for abnormal state).

[0048] By minimizing the total loss function through multi-task learning and optimizing the model parameters, the anomaly scores from the two modules are fused to obtain the final anomaly score. The calculation formula is as follows: ; in Indicates the final abnormal score, The fusion weight coefficient representing the anomaly scores of the two modules is determined through validation set optimization. An anomaly score threshold is set. ,when At that time, the support condition was determined to be abnormal, and further calculations were performed to determine the depth vibration characteristics. Combined spatiotemporal characteristics of displacement-strain The deviation from the Euclidean distance of the feature centers of normal samples, if The largest deviation indicates an abnormal vibration frequency; if If the deviation is the largest, then the normalized average deviation between the preprocessed displacement data and strain data relative to the historical normal mean within the current time window is further calculated. If the deviation of the displacement data is significantly higher than that of the strain data, it is determined that the displacement exceeds the limit; otherwise, it is determined that the strain changes abruptly. When the condition is normal, the support is determined to be in normal condition.

[0049] IV. Construction of an Adaptive Optimization Platform for Bridge Support Anomaly Detection Based on Meta-learning Real-time feedback closed-loop construction: After deploying the trained hybrid anomaly detection model to the online detection environment, a real-time feedback closed-loop system is constructed. This system receives the anomaly determination results output by the anomaly detection model in real time. Simultaneously, the actual status results obtained from manual inspections or professional equipment testing are collected. The prediction accuracy of the model is quantified by calculating the detection deviation, where the anomaly detection result... and actual results All conditions include four states: normal, displacement exceeding limits, abnormal vibration frequency, and sudden strain change. The calculation process for the detection deviation is as follows: the sum of the absolute differences between the predicted and actual results of all samples within the statistical period. This difference calculation considers the correctness of the state classification; that is, when the predicted state does not match the actual state, it is included in the deviation. This is then divided by the total number of samples to obtain the average detection deviation, which reflects the accuracy of the model's prediction.

[0050] Meta-learning framework and few-shot fine-tuning: A meta-learning framework for bridge bearing anomaly adaptation (BBAA-MLF) is introduced. This framework is specifically designed to adapt to novel anomaly patterns and changing working conditions in bearings, and includes two stages: meta-training and meta-fine-tuning. The meta-training stage is based on historical multimodal datasets. We construct multiple small-sample tasks from samples of different anomaly types. Each task contains a support set and a query set. By optimizing the generalization error between tasks, we initialize the model parameters in regions that are sensitive to changes in anomaly patterns.

[0051] When a new abnormal pattern is detected or long-term changes in environmental conditions are detected, the system automatically collects small sample labeled data in the new scenario. Based on BBAA-MLF, rapid fine-tuning with small samples is performed to update the parameters of the hybrid anomaly detection model. Parameter updates employ gradient descent, and the calculation process is as follows: The gradient of the meta-loss function with respect to the model parameters is calculated based on the support set. Then, the learning rate is fine-tuned according to meta-learning, updating the model parameters in the reverse direction of the gradient, enabling the model to quickly adapt to new anomaly patterns. The calculation formula is as follows: ; in This represents the model parameters after fine-tuning. This represents the model parameters before fine-tuning. This indicates that the learning rate is fine-tuned during meta-learning and determined through optimization during the meta-training phase. The meta-loss function with respect to the model parameters gradient, Indicates based on support set The calculated meta-loss function adopts the joint loss function form consistent with the hybrid anomaly detection model.

[0052] V. Hardware Integration and Implementation for Bridge Bearing Anomaly Detection Overall architecture as follows Figure 3 As shown: Hardware integration: Vibration acceleration sensors, displacement sensors, strain sensors, and environmental parameter sensors are installed at key locations on the bridge supports. The sensor array is connected to the data acquisition unit via an industrial bus. The data acquisition unit establishes a communication connection with the server of the bridge health monitoring center through a 5G industrial router, enabling remote data transmission and real-time interaction.

[0053] Software Deployment: Anomaly detection software is deployed on the server of the bridge health monitoring center, integrating the aforementioned bridge bearing feature extraction module, hybrid anomaly detection model, and adaptive optimization platform. The software system adopts a modular design, including a data receiving module, a preprocessing module, a feature extraction module, an anomaly detection module, an adaptive optimization module, a result display module, and an early warning module. Each module achieves data interaction and collaborative work through standardized interfaces.

[0054] Real-time anomaly detection process: The anomaly detection software receives multi-source parameter data from the sensor array in real time through the data receiving module, transmits the data to the preprocessing module, and performs preprocessing operations such as denoising, outlier removal, standardization, and spatiotemporal alignment in sequence according to the above method. The preprocessed data is input into the feature extraction module, which extracts deep vibration features, displacement-strain joint spatiotemporal features, and long-term environmental dependence features through BBVFE-1D-CNN with multi-head self-attention mechanism, BBS-TGCN, BBEF-GRU with cross-modal attention module, respectively, and generates preliminary fusion features. After deep fusion of the preliminary fusion features by BBM-AAE, the data is input into the hybrid anomaly detection model, and the final anomaly score is calculated by the Deep SVDD module and the temporal anomaly scoring network to determine whether the support state is abnormal and the specific anomaly type.

[0055] VI. Experimental Verification and Analysis This experiment, based on the proposed multimodal anomaly detection technique for bridge bearings, was validated using a self-built bridge dataset. Some data samples are shown below. Figure 4 As shown. The dataset covers normal operating conditions, environmental disturbances, minor damage evolution, and typical anomalies including displacement exceeding limits, vibration frequency anomalies, and strain abrupt changes. It is divided into training, validation, and test sets in a ratio of 7:1.5:1.5. 1. Validation of the effectiveness of multimodal feature extraction and fusion This study focuses on the feature representation capabilities and fusion effects of three modes: vibration, displacement-strain, and environment. It compares the discriminative performance of single-mode features with that of multi-mode fused features to verify the rationality of the feature extraction module. Evaluation metrics: Feature discriminativeness: The ability of a feature to distinguish between abnormal states; the higher the value, the stronger the discriminativeness. Anomaly feature recall (FRR): The proportion of features that effectively capture anomaly-related information. Feature extraction time: The computational delay of feature extraction for a single sample throughout the entire process.

[0056] Table 1. Experimental results verifying the effectiveness of multimodal feature extraction and fusion.

[0057] 2. Performance Validation of Hybrid Anomaly Detection Model The detection performance of the proposed hybrid detection model (Deep SVDD+BBTAC-TCN) was compared with that of traditional anomaly detection methods to verify the model's ability to identify different anomaly types and its robustness.

[0058] Evaluation indicators: Detection accuracy: The accuracy rate of determining the support status; Anomaly type identification rate: The proportion of accurately distinguishing between displacement exceeding limits, abnormal vibration frequency, and sudden strain changes; False alarm rate: The probability that a normal state is mistakenly identified as abnormal.

[0059] Table 2. Experimental results of performance verification of the hybrid anomaly detection model.

[0060] Table 1 verifies the advantages of the multimodal feature fusion strategy compared with single-modal features in bridge bearing anomaly detection, and evaluates the expressive power and extraction efficiency of features under different working conditions, providing experimental basis for the rationality of the feature extraction module; Table 2 verifies the core advantages of the proposed hybrid detection model combined with meta-learning fine-tuning, and verifies the model's ability to identify different anomaly types and its adaptability to small sample scenarios by comparing it with single models and traditional methods. Figure 5 The performance of anomalous feature recall under three test scenarios—normal operating conditions, environmental disturbances, and minor damage evolution—was demonstrated for four feature types: vibration single mode, displacement-strain single mode, environmental single mode, and multi-mode fusion. Figure 6 The display shows the feature discriminant values ​​of four feature types in three test scenarios, with the color intensity intuitively representing the strength of the discriminant ability. Figure 7 The accuracy of four detection methods—including single Deep SVDD, single TCN, traditional thresholding, and the hybrid detection model of this method—in identifying three anomaly types: displacement exceeding limits, vibration frequency anomalies, and strain abrupt changes is demonstrated.

[0061] Experimental results show that the proposed multimodal feature fusion strategy significantly improves the discriminativeness of abnormal features. The proposed hybrid detection model exhibits higher detection accuracy, abnormality type recognition rate and lower false alarm rate in various abnormal scenarios, which is better than traditional methods and single-modal detection schemes.

[0062] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0063] While the specific embodiments of the present invention have been described above, they are not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for detecting anomalies in bridge bearings, characterized in that, The method comprises the following steps: S1, synchronously collecting vibration acceleration data sets, displacement data sets, strain data sets and environmental parameter data sets of key positions of a bridge support, and performing data preprocessing; S2, inputting the four preprocessed data sets into a bridge support feature extraction module; a vibration deep feature extraction network based on a one-dimensional convolutional neural network and multi-head self-attention is used to extract features from the vibration acceleration data to obtain deep vibration features; a displacement-strain joint feature extraction network based on a space-time graph convolution is used to extract features from the displacement data and the strain data to obtain displacement-strain joint space-time features; an environmental-multimodal fusion feature extraction network based on a gated recurrent unit and cross-modal attention is used to extract features from the environmental parameter data to obtain environmental long-term dependence features; a cross-modal attention fusion network is used to perform multimodal feature fusion on the deep vibration features, the displacement-strain joint space-time features and the environmental long-term dependence features to obtain preliminarily fused multimodal features; S3, inputting the preliminarily fused multimodal features into a bridge support anomaly detection model, first performing multimodal deep fusion through a bridge support multimodal adversarial autoencoder to obtain latent space features; then inputting the latent space features into a hybrid anomaly detection model, combining a deep support vector data description module and a bridge support time series anomaly capturing convolutional network to calculate anomaly scores, fusing the anomaly scores and outputting whether the support state is abnormal and the abnormal type.

2. The bridge bearing abnormality detection method according to claim 1, characterized by: The vibration acceleration data set is vibration acceleration time series data of the support under different loads; The displacement data set is displacement and deformation time series data of the support in the vertical and transverse directions; The strain data set is strain distribution time series data of the support; The environmental parameter data set is temperature and humidity environmental time series data of the monitoring area.

3. The method of claim 1, wherein: The vibration deep feature extraction network specifically comprises: The pretreated vibration acceleration time series data is reshaped into a one-dimensional sequence feature input bridge support vibration feature extraction one-dimensional convolutional neural network, which comprises four convolutional blocks and three pooling layers, each convolutional block is composed of a one-dimensional convolutional layer, a batch normalization layer and an ELU activation function, the convolution kernel size is set to 3, 5, 7 and 9 respectively, and the receptive field is expanded layer by layer to capture local vibration features of different scales, and the pooling layer adopts a maximum pooling operation to realize feature dimension reduction and abstraction; an adaptive vibration frequency enhancement module is embedded after each convolutional block, the feature weight of different frequency intervals is dynamically adjusted by learning the frequency distribution characteristics of the vibration signal, and local vibration features are obtained extracting local vibration features The input multi-head self-attention mechanism module captures the global dependence relationship of the vibration signal in the time dimension through multiple parallel attention heads, and each attention head focuses on the vibration mode correlation of different time periods. The multi-head self-attention mechanism calculates the global vibration features through the following steps: First, each attention head calculates attention weights by dot product operation of query vectors, key vectors and value vectors, and then weights and sums the value vectors to obtain the output of a single attention head; then, the outputs of all attention heads are spliced and the multi-head information is fused through a linear projection layer to finally obtain the global vibration feature ; The local vibration features are concatenated with the global vibration features , integrated by a fully connected layer and a BatchNorm layer, and finally the deep vibration features are obtained .

4. The method of claim 1, wherein: The displacement-strain joint feature extraction network specifically comprises: The displacement- strain joint data is fused with the pretreated displacement amount, deformation time sequence data and strain distribution time sequence data to construct displacement- strain joint data And a bridge support space- time graph structure is constructed based on the sensor space deployment Wherein the node set Corresponding to the sensor deployment position on the support, each node contains displacement and strain data of the position; the edge set The edge weight between nodes is greater when the distance between nodes is closer and the mechanical transmission path is shorter; the feature matrix The displacement- strain joint data is fused with the pretreated displacement amount, deformation time sequence data and strain distribution time sequence data to construct displacement- strain joint data The feature matrix is composed of displacement- strain joint data Spatiotemporal graph structure The input bridge support spatiotemporal graph convolution network comprises four layers of spatiotemporal graph convolution blocks and two layers of time sequence pooling layers; each layer of the spatiotemporal graph convolution block is composed of a spatial graph convolution submodule and a time convolution submodule; the spatial graph convolution submodule adopts a graph convolution operation of a graph attention mechanism, and calculates a dynamically changing attention weight for each sensor node pair connected by an edge in the graph; the weight value is derived from the current feature vector of the node pair; the current feature vector of the node pair is first transformed and activated through a neural network layer, then the feature similarity between the two is calculated, and finally the normalized processing is performed; when the spatial information is aggregated, the features of each neighbor node are weighted and fused after being multiplied by the corresponding dynamic weight. The time convolution submodule adopts one-dimensional causal convolution operation to capture the time evolution law of the same node at different times, and the convolution kernel slides along the time axis to calculate: when calculating the output at any time, only the input data at the time and all historical times before the time are allowed to be used; by stacking multiple layers of such convolution layers and increasing the field of view of the convolution kernel on the time axis layer by layer; The spatiotemporal graph convolution block is embedded with a timing pooling layer, and an adaptive timing pooling operation is used to reduce the dimension of the timing feature and retain the feature information of the key timing node; the output of the bridge support spatiotemporal graph convolution network is aggregated through a global average pooling layer to obtain displacement-strain joint spatiotemporal features .

5. The method of claim 1, wherein: The environmental-multimodal fusion feature extraction network specifically comprises: The normalized environment time series data is input into a bridge support environment feature extraction gated recurrent unit, which uses its internal update gate and reset gate mechanism to regulate the forgetting of historical information and the fusion of current information; through the gating mechanism, the long-term time dependence of the environmental parameters is adaptively captured, and finally, the final hidden state of the bridge support environment feature extraction gated recurrent unit converges the context information of the entire input sequence, and the final hidden state is the long-term dependence feature of the environment .

6. The method of claim 1, wherein: The cross-modal attention fusion network specifically comprises: First, each modal feature is mapped to the same dimension through two fully connected networks, then the importance score is obtained by dot product with the shared projection vector, and finally the deep vibration feature attention weight coefficient is obtained by normalizing through the softmax function , displacement-strain joint space-time feature attention weight coefficient , environmental long-term dependence feature attention weight coefficient ; The preliminary fused multi-modal features are obtained based on the attention weight coefficients .

7. The method of claim 1, wherein: The bridge support multimodal adversarial autoencoder is composed of an encoder, a decoder and a discriminator; The encoder adopts a multi-layer perception structure, introduces a modal adaptive normalization layer, dynamically adjusts the distribution of different modal features, and calculates the latent space features ; The decoder employs a multi-layer perceptron structure symmetric to the encoder to map the latent space features back to the original multi-modal feature space for computing the reconstruction loss; The discriminator adopts a convolutional neural network structure and is used for distinguishing latent space features The encoder is forced to generate latent feature distribution approximating the prior distribution through adversarial training with samples of the preset prior distribution.

8. The method of claim 7, wherein: The hybrid anomaly detection model comprises a deep support vector data description module and a time series anomaly scoring network module; The deep support vector data description module learns a minimum hypersphere that tightly encloses the fused features of the normal state within the sphere, while the features of the abnormal state fall outside the sphere. First, the center of the hypersphere is learned using samples labeled as normal states in the training set. With radius , center of the ball Initialized to the mean of normal sample features, then iteratively optimized through backpropagation, radius Determined based on the distance distribution from normal sample features to the center of the sphere; for input multimodal deep fusion features Calculate its distance to the center of the ball. The distance is used as part of the outlier score; the final outlier score is obtained. ; The time series anomaly scoring network module adopts a bridge support time series anomaly capturing convolutional network (BBTAC-TCN) containing four causal convolution blocks, each causal convolution block being composed of a one-dimensional dilated convolution layer, a batch normalization layer, a ReLU activation function and a time series attention submodule, and the dilation coefficient being increased layer by layer to capture the dependence relationship at different time scales; The temporal attention submodule calculates the importance weights of features at different time steps to obtain the attention-weighted first-order feature. Features at each time step BBTAC-TCN models the weighted time-series features and calculates the anomaly probability at each time step. Thus, abnormal scores are obtained. .

9. The method of claim 8, wherein: The fused anomaly scores output whether the support state is abnormal and the abnormal type, and the specific process is as follows: The hybrid anomaly detection model is optimized through a joint loss function, and the total loss function includes cross-entropy losses of the deep support vector data description module and the time series anomaly scoring network module; After optimizing the model parameters by minimizing the total loss function through multi-task learning, the final anomaly score is obtained by fusing the anomaly scores of the two modules ; Setting an abnormal score threshold When , the state of the support is determined to be abnormal, and is further classified as displacement overrun, abnormal vibration frequency or strain mutation in combination with the characteristic response mode; when , the state of the support is determined to be normal.

Images