Real-time anomaly prediction and fault diagnosis method and system for bridge component

Abstract

The invention relates to the technical field of bridge health monitoring, and discloses a real-time anomaly prediction and fault diagnosis method and system for a bridge component, and the method comprises the steps: collecting the real-time data and historical data of the bridge component, carrying out the modal parameter recognition of the real-time data, so as to obtain a modal parameter set of the bridge component, constructing a three-dimensional feature tensor of the bridge component based on the real-time data; and performing multi-head self-attention mechanism operation on the three-dimensional feature tensor to obtain a space-time correlation feature tensor of the three-dimensional feature tensor, performing correlation analysis on the space-time correlation feature tensor based on the historical data to obtain a real-time space-time feature tensor of the three-dimensional feature tensor, and storing the real-time space-time feature tensor of the three-dimensional feature tensor. And performing frequency domain spectral coherence analysis on the real-time spatial-temporal feature tensor and the modal parameter set to obtain a spectral coherence coefficient of the bridge component, and the method can perform accurate prediction and diagnosis on the abnormal event of the bridge component.

Description

Technical Field

[0001] This invention relates to the field of bridge health monitoring technology, and in particular to a method and system for real-time anomaly prediction and fault diagnosis of bridge components. Background Technology

[0002] In the field of bridge health monitoring, traditional bridge component anomaly prediction and fault diagnosis technologies often have fixed model structures, which cannot adaptively adjust according to changes in bridge sensor data characteristics. When faced with complex data features such as superimposed multimodal vibration signals, the lack of fine-grained multi-view feature extraction capabilities leads to low prediction accuracy. When data features are stable, unnecessary computational resource consumption cannot be reduced in a timely manner, resulting in wasted GPU memory and increased inference latency. Furthermore, the large number of training parameters leads to low computational efficiency.

[0003] Therefore, how to accurately predict and diagnose abnormal events in bridge components has become an urgent problem to be solved. Summary of the Invention

[0004] This invention provides a method and system for real-time anomaly prediction and fault diagnosis of bridge components. Its main purpose is to solve the problem of how to accurately predict and diagnose abnormal events in bridge components when anomalies occur.

[0005] To achieve the above objectives, the present invention provides a method for real-time anomaly prediction and fault diagnosis of bridge components, the method comprising: S1: Collect real-time and historical data of bridge components, perform modal parameter identification on the real-time data to obtain the modal parameter set of the bridge components, and construct the three-dimensional feature tensor of the bridge components based on the real-time data; S2: Perform multi-head self-attention mechanism operation on the three-dimensional feature tensor to obtain the spatiotemporal correlation feature tensor of the three-dimensional feature tensor, and perform correlation analysis on the spatiotemporal correlation feature tensor based on the historical data to obtain the real-time spatiotemporal feature tensor of the three-dimensional feature tensor. S3: Perform frequency domain spectral coherence analysis on the real-time spatiotemporal feature tensor and the modal parameter set to obtain the spectral coherence coefficient of the bridge component. Based on the spectral coherence coefficient, perform gated dynamic weighted calculation on the modal parameter set to obtain the spatiotemporal attention feature tensor of the modal parameter set. S4: Perform short-time mutation analysis on the spatiotemporal attention feature tensor to obtain the short-time mutation feature tensor of the bridge component; S5: Perform dynamic threshold comparison between the short-term mutation feature tensor and the reference feature tensor in the historical data to obtain the real-time anomaly index of the bridge component.

[0006] Optionally, the calculation formula for the spatiotemporal correlation feature tensor is: in: For the spatiotemporal correlation feature tensor, For a three-dimensional feature tensor, , , For each attention head Independent projection matrix, To project the 3D feature tensor into a query vector, To project the 3D feature tensor into a key vector, To project a 3D feature tensor into a value vector, For merging and feature reconstruction matrices of multi-head outputs, For each attention head Subspace dimension, For time step, For spatial location, For feature dimension, This is the matrix transpose operation. For weight aggregation and multi-head component merging, Self-attention weight matrix , This is a feature enhanced in the same dimension.

[0007] Optionally, the step of performing correlation analysis on the spatiotemporal correlation feature tensor based on the historical data to obtain the real-time spatiotemporal feature tensor of the three-dimensional feature tensor includes: Cosine similarity is calculated on the historical data and the spatiotemporal correlation feature tensor to obtain the historical similarity of the bridge components. Based on the historical similarity, a similarity threshold analysis is performed on the spatiotemporal correlation feature tensor to obtain the feature candidate set of the spatiotemporal correlation feature tensor. The weights corresponding to the feature candidate set are updated based on preset optimization steps to obtain the updated fusion weight matrix of the spatiotemporal correlation feature tensor. The parameters in the calculation formula of the spatiotemporal correlation feature tensor are updated using the updated fusion weight matrix to obtain the real-time spatiotemporal feature tensor of the three-dimensional feature tensor.

[0008] Optionally, the preset optimization steps include: D1: Based on the candidate weight set, perform attention fusion on the self-attention weight matrix to obtain the final fused weight matrix. The calculation formula for attention fusion is as follows: in: For the weighted candidate set, For adaptive mixing coefficients, Here is the self-attention weight matrix. The final fusion weight matrix, The median of the candidate weight set; D2: Update the final fusion weight matrix based on a preset update algorithm to obtain the updated fusion weight matrix corresponding to the final fusion weight matrix. The calculation formula of the update algorithm is as follows: in: Let be the objective function. The regularization coefficient is . Let KL divergence be the KL divergence. It is the L2 norm. According to The generated prediction output, To and The corresponding actual sensor measurement value, For the first The updated fusion weight matrix after the next iteration is also For the final fused weight matrix, The historical weights corresponding to the historical data. For the first The gradient of the objective function with respect to the set of hyperparameters in the next iteration. For the first The approximate Hessian inverse matrix of the nth iteration For the first The dynamic learning rate of each iteration. The number of iterations is obtained based on the historical data. For the first The set of hyperparameters for the next iteration For the first The set of hyperparameters for each iteration.

[0009] Optionally, the step of updating the parameters in the calculation formula of the spatiotemporal correlation feature tensor using the updated fusion weight matrix to obtain the real-time spatiotemporal feature tensor of the three-dimensional feature tensor includes: The self-attention weight matrix in the calculation formula of the spatiotemporal correlation feature tensor is updated using the updated fusion weight matrix to obtain the real-time spatiotemporal feature tensor of the three-dimensional feature tensor. The calculation formula of the real-time spatiotemporal feature tensor is as follows: in: For real-time spatiotemporal feature tensors, For a three-dimensional feature tensor, For each attention head after the update Independent projection matrix, For the updated value vector, For merging and feature reconstruction matrices of multi-head outputs, For time step, For spatial location, For feature dimension, For weight aggregation and multi-head component merging, For the first The updated fusion weight matrix after the next iteration is also For the final fused weight matrix, This is a feature enhanced in the same dimension.

[0010] Optionally, the step of performing frequency domain spectral coherence analysis on the real-time spatiotemporal feature tensor and the modal parameter set to obtain the spectral coherence coefficients of the bridge components includes: A time window alignment operation is performed on the real-time spatiotemporal feature tensor and the modal parameter set to obtain a synchronized data segment set of the bridge component. Based on the synchronized data segment set, the frequency domain amplitude spectrum of the real-time spatiotemporal feature tensor is extracted to obtain the dynamic energy distribution spectrum of the bridge component. Based on the natural frequencies and bandwidth parameters in the modal parameter set, narrowband pulse spectrum analysis is performed on the dynamic energy distribution spectrum to obtain the modal reference spectrum of the bridge component; The bridge component is subjected to spectral coherence analysis based on the dynamic energy distribution spectrum and the modal reference spectrum to obtain the spectral coherence coefficient of the bridge component.

[0011] Optionally, the step of performing spectral coherence analysis on the bridge component based on the dynamic energy distribution spectrum and the modal reference spectrum to obtain the spectral coherence coefficient of the bridge component includes: Based on the dynamic energy distribution spectrum and the modal reference spectrum, local energy correlation analysis is performed on the amplitude of frequency points in the bridge component to obtain the local coherence scaling value of the bridge component; The effective frequency bands in the historical data are weighted and fused based on the local coherence scaling value and the modal confidence weights corresponding to the modal parameter set to obtain the spectral coherence coefficient of the bridge component.

[0012] Optionally, short-time mutation analysis is performed on the spatiotemporal attention feature tensor to obtain the short-time mutation feature tensor of the bridge component, including: Based on a pre-trained BiLSTM model, forward and reverse time dependency learning is performed on the spatiotemporal attention feature tensor to obtain the temporal comprehensive feature sequence of the spatiotemporal attention feature tensor. Local feature extraction is performed on the time-series composite feature sequence to obtain the local gradient sharpness of the time-series composite feature sequence; Based on the local gradient sharpness, a mutation sharpness analysis is performed on the spatiotemporal attention feature tensor to obtain the mutation feature tensor of the spatiotemporal attention feature tensor.

[0013] Optionally, the step of performing a dynamic threshold comparison between the short-term mutation feature tensor and the reference feature tensor in the historical data to obtain the real-time anomaly index of the bridge component includes: The difference between the short-term abrupt change feature tensor and the reference feature tensor is calculated based on the dynamic Mahalanobis distance to obtain the real-time abnormal deviation of the bridge component; The real-time anomaly deviation is normalized using the Sigmoid function, and then the normalized real-time anomaly deviation is exponentially amplified using the adjustment coefficient in the historical data to obtain the real-time anomaly index of the bridge component.

[0014] A real-time anomaly prediction and fault diagnosis system for bridge components, the system comprising: Data processing module: used to collect real-time and historical data of bridge components, perform modal parameter identification on the real-time data to obtain the modal parameter set of the bridge components, and construct the three-dimensional feature tensor of the bridge components based on the real-time data; Transformer processing module: used to perform multi-head self-attention mechanism operation on the three-dimensional feature tensor to obtain the spatiotemporal correlation feature tensor of the three-dimensional feature tensor, and to perform correlation analysis on the spatiotemporal correlation feature tensor based on the historical data to obtain the real-time spatiotemporal feature tensor of the three-dimensional feature tensor; Feature processing module: used to perform frequency domain spectral coherence analysis on the real-time spatiotemporal feature tensor and the modal parameter set to obtain the spectral coherence coefficient of the bridge component, and to perform gated dynamic weighted calculation on the modal parameter set based on the spectral coherence coefficient to obtain the spatiotemporal attention feature tensor of the modal parameter set; BiLSTM processing module: used to perform short-time mutation analysis on the spatiotemporal attention feature tensor to obtain the short-time mutation feature tensor of the bridge component; Analysis module: used to perform dynamic threshold comparison between the short-term mutation feature tensor and the reference feature tensor in the historical data to obtain the real-time anomaly index of the bridge component. Beneficial effects

[0015] 1. Through online optimization of hyperparameter sets, adaptive matching between model structure and computing resources is achieved: When the characteristics of bridge sensor data are complex, such as the superposition of multimodal vibration signals, the hyperparameter set optimization automatically increases the number of attention heads to improve prediction accuracy through fine-grained multi-view feature extraction; when the data characteristics are stable, the number of attention heads is reduced, which can reduce memory usage and shorten inference latency. Secondly, the independent projection matrix is ​​dynamically reconstructed according to the number of attention heads to eliminate redundant parameters. For example, when the number of heads is reduced, the corresponding matrix block is deleted. Compared with the fixed structure model, the number of training parameters is reduced. Furthermore, the introduction of cross-cycle weight constraints and historical backtracking mechanisms significantly improves the diagnostic reliability under complex working conditions.

[0016] 2. By using the KL divergence constraint term to force the distribution of new weights to align with the historical weights corresponding to historical data, false alarms caused by instantaneous noise, such as vehicle impact vibration, are effectively suppressed. The measured abnormal recall rate is effectively improved and can maintain long-term stability. A phased optimization strategy of "generation → freeze → regeneration" is adopted. In each step of weight update, the fusion weight matrix is ​​fixedly updated as the constraint benchmark, and abnormal historical weights generated when the sensor fails are removed by Median similarity filtering, which ensures the performance of the model in continuous operation and improves the accuracy of monitoring. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating a method for real-time anomaly prediction and fault diagnosis of bridge components according to an embodiment of the present invention. Figure 2 A functional block diagram of a real-time anomaly prediction and fault diagnosis system for bridge components provided in an embodiment of the present invention; The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0018] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0019] This application provides a method for real-time anomaly prediction and fault diagnosis of bridge components. The execution subject of this method includes, but is not limited to, at least one of the following electronic devices that can be configured to execute the method provided in this application: a server, a terminal, etc. In other words, the method for real-time anomaly prediction and fault diagnosis of bridge components can be executed by software or hardware installed on a terminal device or a server device. The server includes, but is not limited to, a single server, a server cluster, a cloud server, or a cloud server cluster. The server can be an independent server or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms.

[0020] Reference Figure 1 The diagram shown is a flowchart illustrating a real-time anomaly prediction and fault diagnosis method for bridge components according to the present invention. In this embodiment, the real-time anomaly prediction and fault diagnosis method for bridge components includes: S1: Collect real-time and historical data of bridge components, perform modal parameter identification on the real-time data to obtain the modal parameter set of the bridge components, and construct the three-dimensional feature tensor of the bridge components based on the real-time data; S2: Perform multi-head self-attention mechanism operation on the three-dimensional feature tensor to obtain the spatiotemporal correlation feature tensor of the three-dimensional feature tensor, and perform correlation analysis on the spatiotemporal correlation feature tensor based on the historical data to obtain the real-time spatiotemporal feature tensor of the three-dimensional feature tensor.

[0021] In this embodiment of the invention, the calculation formula for the spatiotemporal correlation feature tensor is: in: For the spatiotemporal correlation feature tensor, For a three-dimensional feature tensor, , , For each attention head Independent projection matrix, To project the 3D feature tensor into a query vector, To project the 3D feature tensor into a key vector, To project a 3D feature tensor into a value vector, For merging and feature reconstruction matrices of multi-head outputs, For each attention head Subspace dimension, For time step, For spatial location, For feature dimension, This is the matrix transpose operation. For weight aggregation and multi-head component merging, Self-attention weight matrix , This is a feature enhanced in the same dimension.

[0022] Specifically, The spatiotemporal correlation feature tensor is the output of the entire operation, integrating the correlation information of the three-dimensional feature tensor in the temporal and spatial dimensions. In the physical environment of bridge monitoring, it reflects the relationship between the characteristics of bridge components at different times and locations, such as the correlation of vibration and stress changes in different parts at different times, helping to analyze the overall state change trend of the bridge.

[0023] Specifically, This is a three-dimensional feature tensor constructed from real-time data of bridge components, containing information in three dimensions: time, space, and features. Taking bridge vibration monitoring as an example, the time dimension records monitoring data at different times, the spatial dimension corresponds to different monitoring point locations on the bridge, and the feature dimension covers physical characteristics such as vibration amplitude and frequency, comprehensively describing the multifaceted state of bridge components at a certain moment.

[0024] Specifically, , , For each attention head In multi-head self-attention mechanisms, these independent projection matrices project the 3D feature tensor into query vectors, key vectors, and value vectors, respectively. In bridge data processing scenarios, they act as filters, extracting and transforming features from different angles, enabling the model to focus on data relationships across various aspects.

[0025] Specifically, For each attention head The subspace dimension determines the dimensionality of the data processed by each attention head. In bridge monitoring, a larger subspace dimension... This means that the attention head can capture more detailed and richer feature information, but the computational cost will also increase; smaller attention heads can capture more detailed and richer feature information. This simplifies calculations and allows for focusing on key information. For example, appropriate settings can be configured based on the complexity of the bridge structure and the required monitoring accuracy. value.

[0026] Specifically, This refers to the time step, which corresponds to the time interval for collecting bridge data. For example, data is collected every 10 minutes. This represents a 10-minute time span, reflecting the degree of dispersion of the data in the time dimension, which determines the model's accuracy in capturing changes in the time series.

[0027] For spatial location, the location information of monitoring points on the bridge is clearly defined, such as the location of sensors installed in different parts of the bridge, such as piers and spans. These locations are used to identify the differences in data characteristics between different locations, thereby enabling the model to analyze the spatial correlation of changes in the state of different parts of the bridge.

[0028] As a feature dimension, it includes various physical features of bridge monitoring, such as stress, temperature, and displacement, reflecting the richness of features describing the state of bridge components. Different features reflect the working state of the bridge from different aspects.

[0029] Specifically, This is a merging and feature reconstruction matrix for multi-head outputs, which merges and reconstructs the features of multiple attention heads. In bridge diagnostics, it integrates the feature information mined from different angles by various attention heads, providing a more comprehensive feature tensor and a more complete data foundation for subsequent analysis.

[0030] Specifically, , , This is achieved by multiplying the three-dimensional feature tensor by the projection matrix. The data is transformed into query vectors, key vectors, and value vectors, respectively. This reorganization of the bridge data from different perspectives prepares the model for subsequent attention calculations, enabling it to more effectively uncover the relationships between data points.

[0031] Specifically, The key vector is transposed to match the dimensions of the query vector, satisfying the requirements of subsequent matrix multiplication operations and ensuring that the correlation between different vectors can be correctly measured when calculating attention weights.

[0032] Specifically, In the bridge data scenario, it assigns a weight to each data element based on the similarity between the query vector and the key vector. The higher the weight, the more important the element is in the current calculation, thus highlighting the impact of key data features.

[0033] By multiplying the self-attention weight matrix with the value vector and then summing the value vector according to the weights, the output can focus on important data features and strengthen the role of key information in feature representation.

[0034] It involves concatenating the computation results of multiple attention heads along a specific dimension, integrating the feature information mined by different attention heads, retaining data features from different perspectives, and enriching feature representation.

[0035] This involves reshaping the merged result to align its dimensions with the time step of the original data. Spatial location and feature dimensions By matching, a spatiotemporal correlation feature tensor that meets the needs of subsequent analysis is obtained. This facilitates further analysis of the spatiotemporal characteristics of bridge components.

[0036] In this embodiment of the invention, the step of performing correlation analysis on the spatiotemporal correlation feature tensor based on the historical data to obtain the real-time spatiotemporal feature tensor of the three-dimensional feature tensor includes: Cosine similarity is calculated on the historical data and the spatiotemporal correlation feature tensor to obtain the historical similarity of the bridge components. Based on the historical similarity, a similarity threshold analysis is performed on the spatiotemporal correlation feature tensor to obtain the feature candidate set of the spatiotemporal correlation feature tensor. The weights corresponding to the feature candidate set are updated based on preset optimization steps to obtain the updated fusion weight matrix of the spatiotemporal correlation feature tensor. The parameters in the calculation formula of the spatiotemporal correlation feature tensor are updated using the updated fusion weight matrix to obtain the real-time spatiotemporal feature tensor of the three-dimensional feature tensor.

[0037] Specifically, historical data, within the physical environment of bridge components, refers to various data collected over a past period regarding the structural condition of the bridge, such as stress, vibration frequency, and displacement data at different points in time. This data records the bridge's past operational status and serves as a crucial reference for analyzing its current condition.

[0038] Specifically, cosine similarity calculation measures the degree of similarity between historical data and spatiotemporally related feature tensors. In the field of bridge engineering, it determines which past states are most similar to the current state characteristics of bridge components. The closer the value is to 1, the higher the similarity, which helps to find comparable situations from historical experience.

[0039] Specifically, the similarity threshold is a pre-set numerical standard. In bridge monitoring, it is determined based on factors such as bridge structural characteristics and safety requirements, and is used to filter out data that is sufficiently similar to historical conditions.

[0040] In this embodiment of the invention, the preset optimization steps include: D1: Based on the candidate weight set, perform attention fusion on the self-attention weight matrix to obtain the final fused weight matrix. The calculation formula for attention fusion is as follows: in: For the weighted candidate set, For adaptive mixing coefficients, Here is the self-attention weight matrix. The final fusion weight matrix, The median of the candidate weight set; D2: Update the final fusion weight matrix based on a preset update algorithm to obtain the updated fusion weight matrix corresponding to the final fusion weight matrix. The calculation formula of the update algorithm is as follows: in: Let be the objective function. The regularization coefficient is . Let KL divergence be the KL divergence. It is the L2 norm. According to The generated prediction output, To and The corresponding actual sensor measurement value, For the first The updated fusion weight matrix after the next iteration is also For the final fused weight matrix, The historical weights corresponding to the historical data. For the first The gradient of the objective function with respect to the set of hyperparameters in the next iteration. For the first The approximate Hessian inverse matrix of the nth iteration For the first The dynamic learning rate of each iteration. The number of iterations is obtained based on the historical data. For the first The set of hyperparameters for the next iteration For the first The set of hyperparameters for each iteration.

[0041] Specifically, during the optimization step, in the generated final fusion weight matrix Then, input it into the target function. Update the fusion weight matrix (that is:) For the final fusion weight matrix The first update of the fusion weight matrix at this point This remains unchanged in this round of optimization, serving only as a parameter constraint based on the objective function. The set of hyperparameters for the first iteration Optimize to obtain the first hyperparameter set for the next iteration ; Then, based on the first hyperparameter set for the next iteration Regenerate the final fusion weight matrix , (here used) (Indicates) will Input to the target function In the middle, as the second update of the fusion weight matrix (that is:) for At this point, the fusion weight matrix is ​​updated for the second time. This remains unchanged in this round of optimization, serving only as a parameter constraint based on the objective function. No. hyperparameter set for the next iteration Optimize to obtain the first hyperparameter set for the next iteration Then repeat the above operation until the preset number of optimizations is reached.

[0042] Furthermore, during each optimization process, it is necessary to remove the old... Stored in the historical database, which is the historical weight corresponding to the historical data. .

[0043] Furthermore, For the first The hyperparameter set of the next iteration not only adjusts the model complexity, but also achieves adaptive allocation of computing resources and online adjustment of model capacity by dynamically reconfiguring the independent projection matrix.

[0044] Specifically, The weight candidate set, within the physical environment of bridge components, is a set of candidate weight values ​​selected from numerous calculated weights. These weights may be derived from different calculation methods, different time periods, or different monitoring point data, and serve as the foundational material for subsequent fusion operations.

[0045] Specifically, For adaptive mixing coefficients, It is a coefficient between 0 and 1 used to control the proportion of the self-attention weight matrix and the number of candidate weights in the fusion process. In bridge monitoring scenarios, it can be adaptively adjusted according to factors such as the dynamic changes of the bridge structure and data fluctuations. For example, when the bridge is in a special construction stage or affected by extreme environments (such as strong winds or after an earthquake). The weights can be appropriately increased to make the self-attention weight matrix dominate the fusion process, focusing more on the current real-time data characteristics; while during periods of relatively stable bridge operation, This can be reduced; refer to more information on candidate weight sets.

[0046] Specifically, This is a self-attention weight matrix, calculated through a self-attention mechanism. It reflects the correlation and importance weights between different elements in bridge data, such as monitoring data features at different times and locations. For example, when analyzing bridge vibration data, it can highlight the data weights related to vibrations in critical structural safety areas.

[0047] Specifically, The median of the candidate weight set is the value in the middle after sorting the weight values ​​in the candidate set by size. It represents the central tendency of the candidate weight set and plays a balancing and reference role in the fusion operation, avoiding excessive influence from a few extreme weight values.

[0048] Specifically, The calculation formula is to combine the self-attention weight matrix with the median of the candidate weight set according to the adaptive mixing coefficients. This involves weighted combination. Essentially, it combines the importance of current data features mined based on a self-attention mechanism. and the overall trend information represented by the candidate weight set This generates a more comprehensive final fusion weight matrix that better adapts to the actual condition of the bridge, providing a reasonable weight allocation for subsequent analysis of bridge component data.

[0049] Specifically, The objective function is used to measure the difference between the model's predicted output and the actual sensor measurements, while also considering the similarity constraints between the model parameters and historical parameters. In bridge monitoring, it is a quantitative indicator for evaluating the accuracy of the model's predictions of the bridge's structural condition; the goal is to minimize this value by adjusting the model parameters.

[0050] Specifically, This is the regularization coefficient, used to control model complexity, balance the model's fit to the training data, and prevent overfitting. In bridge data processing, if... If the size is too large, the model may be too conservative and insensitive to newly emerging changes in the bridge's state; if If the value is too small, the model may overfit the training data and have poor generalization ability when faced with new actual monitoring data.

[0051] Specifically, KL divergence measures the difference between two probability distributions; here, it measures the degree of difference between the currently updated weight matrix and the historical weight matrix. In bridge applications, it ensures that the model does not deviate too far from historical experience when updating weights, avoiding unreasonable parameter changes due to excessive pursuit of current data fit.

[0052] Specifically, L2 norm is used to calculate the magnitude of a vector and measures the error between the predicted output and the actual value in the objective function. In bridge monitoring data processing, it quantifies the degree of deviation between the model's predicted bridge condition (such as stress, displacement, etc.) and the actual sensor measurements.

[0053] Specifically, According to The generated prediction output is based on the input data. For example, monitoring data from different parts of a bridge at different times, and predictions about the structural state of the bridge obtained through model calculations, such as predicting the vibration amplitude of a certain part of the bridge over a future period of time.

[0054] Specifically, To and The corresponding actual sensor measurements are the bridge structural status data actually collected by various sensors installed on the bridge, such as stress sensors and displacement sensors, which truly reflect the current state of the bridge.

[0055] Specifically, For the first The updated fusion weight matrix after the nth iteration is, in the nth iteration... The weight matrix obtained after each iteration update is the result of the model continuously adjusting its parameters to adapt to the real-time status data of the bridge.

[0056] Specifically, The historical weights corresponding to the historical data are weight matrices calculated based on historical monitoring data. They reflect the distribution of the characteristic importance of the bridge's past state data and serve as a reference for the current weight update, preventing the model parameters from changing too drastically.

[0057] Specifically, The iteration count, derived from the historical data, represents the number of times model parameters were updated based on that historical data. In bridge data analysis, as new data is continuously collected, the iteration count increases, and the model continuously optimizes its weight matrix to better adapt to changes in bridge condition.

[0058] Specifically, This represents the rate of change of the objective function value with respect to hyperparameters, reflecting the direction in which adjusting the hyperparameters will lead to a faster decrease in the objective function value. In bridge model training, it guides the direction of hyperparameter updates, enabling the model to continuously improve its predictive performance.

[0059] Specifically, For the first The approximate Hessian inverse matrix of the next iteration is used in optimization algorithms to approximate second-order derivative information, helping to adjust hyperparameters more accurately and accelerate model convergence. In bridge data model optimization, it assists in determining the step size and direction of hyperparameter adjustment, making the weight matrix update more reasonable.

[0060] Specifically, For the first The gradient of the objective function with respect to the hyperparameter set in each iteration controls the step size of hyperparameter updates in each iteration. During bridge model training, if the learning rate is too large, the model may oscillate around the optimal solution and fail to converge; if the learning rate is too small, the model will converge very slowly. A dynamic learning rate can adaptively adjust based on the model's performance during iterations, improving training efficiency.

[0061] Specifically, For the first The hyperparameter set for the next iteration, through the hyperparameter set Online optimization of factors such as the number of attention heads and the dimension of the subspace enables adaptive matching between model structure and computing resources.

[0062] Specifically, The above formula comprehensively considers the error between the predicted and actual values, measured by the L2 norm, and the difference between the current weight matrix and the historical weight matrix, measured by the KL divergence. Then, based on the preset update algorithm, the formula is used... The hyperparameter set is updated. This operation involves adjusting the hyperparameter set based on the gradient information of the objective function with respect to the hyperparameter set, combined with the approximate Hessian inverse matrix and dynamic learning rate. By continuously iterating through this process, the final fused weight matrix is ​​gradually optimized, enabling the model to better fit the actual monitoring data of the bridge and improve the accuracy of predicting the state of bridge components and diagnosing faults.

[0063] Furthermore, through online optimization of the hyperparameter set, adaptive matching between model structure and computing resources is achieved: when the bridge sensor data features are complex, such as the superposition of multimodal vibration signals, the hyperparameter set optimization automatically increases the number of attention heads to improve prediction accuracy through fine-grained multi-view feature extraction; when the data features are stable, the number of attention heads is reduced, which can reduce memory usage and shorten inference latency. Secondly, the independent projection matrix is ​​dynamically reconstructed according to the number of attention heads to eliminate redundant parameters. For example, when the number of heads is reduced, the corresponding matrix block is deleted, which reduces the number of training parameters compared to a fixed structure model. In addition, cross-cycle weight constraints and historical backtracking mechanisms are introduced to significantly improve the diagnostic reliability under complex working conditions.

[0064] In this embodiment of the invention, the step of updating the parameters in the calculation formula of the spatiotemporal correlation feature tensor using the updated fusion weight matrix to obtain the real-time spatiotemporal feature tensor of the three-dimensional feature tensor includes: The self-attention weight matrix in the calculation formula of the spatiotemporal correlation feature tensor is updated using the updated fusion weight matrix to obtain the real-time spatiotemporal feature tensor of the three-dimensional feature tensor. The calculation formula of the real-time spatiotemporal feature tensor is as follows: in: For real-time spatiotemporal feature tensors, For a three-dimensional feature tensor, For each attention head after the update Independent projection matrix, For the updated value vector, For merging and feature reconstruction matrices of multi-head outputs, For time step, For spatial location, For feature dimension, For weight aggregation and multi-head component merging, For the first The updated fusion weight matrix after the next iteration is also For the final fused weight matrix, This is a feature enhanced in the same dimension.

[0065] S3: Perform frequency domain spectral coherence analysis on the real-time spatiotemporal feature tensor and the modal parameter set to obtain the spectral coherence coefficient of the bridge component. Based on the spectral coherence coefficient, perform gated dynamic weighted calculation on the modal parameter set to obtain the spatiotemporal attention feature tensor of the modal parameter set.

[0066] Specifically, This matrix comprehensively considers information such as self-attention mechanisms and candidate weight sets. In bridge monitoring scenarios, it can dynamically adjust the importance of different features based on real-time bridge status data. For example, when a bridge is subjected to different loads, the weights of different monitoring data features will be adjusted through this matrix.

[0067] Specifically, For the first The updated fusion weight matrix after the next iteration is also For the final fusion weight matrix Specifically, The real-time spatiotemporal feature tensor is a tensor that reflects the current spatiotemporal characteristics of bridge components after parameter updates. It integrates various feature information of the bridge at different times and spatial locations, and is a key data structure for subsequent analysis of the bridge's condition. For example, it can be used to analyze the comprehensive state of the bridge at different locations at a certain moment, including stress and displacement characteristics.

[0068] In this embodiment of the invention, the step of performing frequency domain spectral coherence analysis on the real-time spatiotemporal feature tensor and the modal parameter set to obtain the spectral coherence coefficients of the bridge components includes: A time window alignment operation is performed on the real-time spatiotemporal feature tensor and the modal parameter set to obtain a synchronized data segment set of the bridge component. Based on the synchronized data segment set, the frequency domain amplitude spectrum of the real-time spatiotemporal feature tensor is extracted to obtain the dynamic energy distribution spectrum of the bridge component. Based on the natural frequencies and bandwidth parameters in the modal parameter set, narrowband pulse spectrum analysis is performed on the dynamic energy distribution spectrum to obtain the modal reference spectrum of the bridge component; The bridge component is subjected to spectral coherence analysis based on the dynamic energy distribution spectrum and the modal reference spectrum to obtain the spectral coherence coefficient of the bridge component.

[0069] Specifically, the spectral coherence coefficient is a coefficient obtained by performing spectral coherence analysis on the real-time spatiotemporal feature tensor and modal parameter set, used to measure the degree of correlation between the two in the spectrum. In bridge monitoring, it can reflect the close correlation between the current state characteristics of the bridge and the vibration modal characteristics, helping to determine whether the bridge structure is in a normal vibration state or whether there is potential damage.

[0070] Specifically, for the real-time spatiotemporal feature tensor Spectral coherence analysis is performed on the modal parameter set to calculate the spectral coherence coefficient, which measures the correlation between the two in the spectral dimension. Based on the obtained spectral coherence coefficient, a gated dynamic weighted calculation is performed on the modal parameter set. The gating mechanism is similar to a filter, adjusting the weights of different parameters in the modal parameter set according to the magnitude of the spectral coherence coefficient; parameters with high correlation have increased weights, while parameters with low correlation have decreased weights. Through this weighted calculation, a spatiotemporal attention feature tensor is obtained, which makes the modal feature information closely related to the current state of the bridge more prominent, improving the accuracy and effectiveness of the bridge structural state analysis.

[0071] In this embodiment of the invention, the step of performing spectral coherence analysis on the bridge component based on the dynamic energy distribution spectrum and the modal reference spectrum to obtain the spectral coherence coefficient of the bridge component includes: Based on the dynamic energy distribution spectrum and the modal reference spectrum, local energy correlation analysis is performed on the amplitude of frequency points in the bridge component to obtain the local coherence scaling value of the bridge component; The effective frequency bands in the historical data are weighted and fused based on the local coherence scaling value and the modal confidence weights corresponding to the modal parameter set to obtain the spectral coherence coefficient of the bridge component.

[0072] Specifically, the dynamic energy distribution spectrum is a spectrum obtained in the physical environment of a bridge by collecting energy information about the vibrations of the bridge structure during dynamic processes such as vehicle traffic and wind loads, and then organizing it according to frequency distribution. This spectrum reflects the energy distribution of the bridge at different frequencies. For example, higher vibration energy at certain frequencies may correspond to weak points or resonant frequencies in the structure.

[0073] Specifically, the modal reference spectrum is a reference spectrum related to the vibration modes of a bridge structure. It contains frequency information about the bridge's inherent vibration characteristics, such as its natural frequencies. It is obtained based on theoretical calculations or preliminary experimental tests of the bridge's structural design and material properties, and serves as an important reference for determining whether the actual vibration state of the bridge is normal.

[0074] Specifically, the amplitude at a frequency point is the vibration amplitude value corresponding to each frequency point in the frequency spectrum. In bridge vibration monitoring, the magnitude of the amplitude reflects the severity of the bridge vibration at that frequency. The larger the amplitude, the stronger the vibration at that frequency, and the greater the potential impact on the bridge structure.

[0075] Specifically, the local coherence scaling value is obtained by performing local energy correlation analysis on the amplitude of frequency points in the dynamic energy distribution spectrum and the modal reference spectrum. It measures the degree of correlation between the actual vibration energy distribution of the bridge and the theoretical modal vibration characteristics within a specific frequency range, and is a quantitative indicator for judging whether the vibration state of the bridge structure meets expectations.

[0076] S4: Perform short-time mutation analysis on the spatiotemporal attention feature tensor to obtain the short-time mutation feature tensor of the bridge component.

[0077] In this embodiment of the invention, short-time mutation analysis is performed on the spatiotemporal attention feature tensor to obtain the short-time mutation feature tensor of the bridge component, including: Based on a pre-trained BiLSTM model, forward and reverse time dependency learning is performed on the spatiotemporal attention feature tensor to obtain the temporal comprehensive feature sequence of the spatiotemporal attention feature tensor. Local feature extraction is performed on the time-series composite feature sequence to obtain the local gradient sharpness of the time-series composite feature sequence; Based on the local gradient sharpness, a mutation sharpness analysis is performed on the spatiotemporal attention feature tensor to obtain the mutation feature tensor of the spatiotemporal attention feature tensor.

[0078] Specifically, in the context of bridge structural condition monitoring, the bidirectional long short-term memory network model can simultaneously learn the positive and negative dependencies of bridge monitoring data over time. For example, it can not only learn the impact of data from before a certain moment on the current state, but also consider the correlation of data after that moment, thus more comprehensively capturing the changing patterns of the bridge's condition over time.

[0079] Specifically, the spatiotemporal attention feature tensor is a tensor that comprehensively considers the spatiotemporal characteristics of the bridge and its correlation with modal parameters. It contains various feature data of the bridge with specific weights at different times and spatial locations.

[0080] Specifically, the temporal integrated feature sequence is the result obtained by learning the spatiotemporal attention feature tensor through the BiLSTM model. It processes the spatiotemporal attention feature tensor in the time dimension, integrates the dependencies between feature information at different time points, and forms a feature sequence that reflects the changes in the bridge state over time, similar to a more temporally related description of the bridge state arranged in chronological order.

[0081] Specifically, local feature extraction involves performing local feature extraction operations on the temporal comprehensive feature sequence. Mathematical methods can be used, such as calculating the difference or slope of feature values ​​at adjacent time points, to quantify the changes in features within a local range. This method yields local gradient sharpness, which highlights significantly changing local areas in the feature sequence, helping to identify locations where rapid changes or abnormal fluctuations in the bridge's condition may occur within a short period.

[0082] S5: Perform dynamic threshold comparison between the short-term mutation feature tensor and the reference feature tensor in the historical data to obtain the real-time anomaly index of the bridge component.

[0083] In this embodiment of the invention, the step of performing a dynamic threshold comparison between the short-term mutation feature tensor and the reference feature tensor in the historical data to obtain the real-time anomaly index of the bridge component includes: The difference between the short-term abrupt change feature tensor and the reference feature tensor is calculated based on the dynamic Mahalanobis distance to obtain the real-time abnormal deviation of the bridge component; The real-time anomaly deviation is normalized using the Sigmoid function, and then the normalized real-time anomaly deviation is exponentially amplified using the adjustment coefficient in the historical data to obtain the real-time anomaly index of the bridge component.

[0084] Specifically, regarding the accuracy and effectiveness of anomaly prediction and diagnosis, frequency domain spectral coherence analysis is performed on real-time spatiotemporal feature tensors and modal parameter sets. Combined with short-time abrupt change analysis and dynamic threshold comparison with historical data, anomalies in bridge components can be accurately identified. Real-time anomaly indices are calculated based on methods such as dynamic Mahalanobis distance and the Sigmoid function, making the diagnostic results more consistent with reality. This enables timely detection of potential faults, avoiding serious consequences caused by undetected faults, and significantly improving the safety and reliability of bridge operation.

[0085] Specifically, real-time anomaly indices are compared and analyzed based on historical data to obtain the type and factors of the fault.

[0086] like Figure 2 The diagram shown is a functional block diagram of a real-time anomaly prediction and fault diagnosis system for bridge components provided in an embodiment of the present invention.

[0087] The real-time anomaly prediction and fault diagnosis system 100 for bridge components described in this invention can be installed in an electronic device. Depending on the functions implemented, the real-time anomaly prediction and fault diagnosis system 100 for bridge components may include a data processing module 101, a Transformer processing module 102, a feature processing module 103, a BiLSTM processing module 104, and an analysis module 105. The modules described in this invention can also be referred to as units, which are a series of computer program segments that can be executed by the processor of an electronic device and can perform a fixed function, and are stored in the memory of the electronic device.

[0088] In this embodiment, the functions of each module / unit are as follows: Data processing module 101: used to collect real-time data and historical data of bridge components, perform modal parameter identification on the real-time data to obtain the modal parameter set of the bridge components, and construct the three-dimensional feature tensor of the bridge components based on the real-time data; Transformer processing module 102: is used to perform multi-head self-attention mechanism operation on the three-dimensional feature tensor to obtain the spatiotemporal correlation feature tensor of the three-dimensional feature tensor, and to perform correlation analysis on the spatiotemporal correlation feature tensor based on the historical data to obtain the real-time spatiotemporal feature tensor of the three-dimensional feature tensor. Feature processing module 103: used to perform frequency domain spectral coherence analysis on the real-time spatiotemporal feature tensor and the modal parameter set to obtain the spectral coherence coefficient of the bridge component, and to perform gated dynamic weighted calculation on the modal parameter set based on the spectral coherence coefficient to obtain the spatiotemporal attention feature tensor of the modal parameter set; BiLSTM processing module 104: used to perform short-time mutation analysis on the spatiotemporal attention feature tensor to obtain the short-time mutation feature tensor of the bridge component; Analysis module 105: Performs dynamic threshold comparison between the short-term mutation feature tensor and the reference feature tensor in the historical data to obtain the real-time anomaly index of the bridge component. In the several embodiments provided by this invention, it should be understood that the disclosed methods and systems can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for instance, the module division is only a logical functional division, and other division methods may exist in actual implementation.

[0089] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0090] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or in the form of hardware plus software functional modules.

[0091] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0092] The embodiments of this application can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence is the theory, method, technology, and application system that uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results.

[0093] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for real-time anomaly prediction and fault diagnosis of bridge components, characterized in that, The method includes: S1: Collect real-time and historical data of bridge components, perform modal parameter identification on the real-time data to obtain the modal parameter set of the bridge components, and construct the three-dimensional feature tensor of the bridge components based on the real-time data; S2: Perform multi-head self-attention mechanism operation on the three-dimensional feature tensor to obtain the spatiotemporal correlation feature tensor of the three-dimensional feature tensor, and perform correlation analysis on the spatiotemporal correlation feature tensor based on the historical data to obtain the real-time spatiotemporal feature tensor of the three-dimensional feature tensor. S3: Perform frequency domain spectral coherence analysis on the real-time spatiotemporal feature tensor and the modal parameter set to obtain the spectral coherence coefficient of the bridge component. Based on the spectral coherence coefficient, perform gated dynamic weighted calculation on the modal parameter set to obtain the spatiotemporal attention feature tensor of the modal parameter set. S4: Perform short-time mutation analysis on the spatiotemporal attention feature tensor to obtain the short-time mutation feature tensor of the bridge component; S5: Perform dynamic threshold comparison between the short-term mutation feature tensor and the reference feature tensor in the historical data to obtain the real-time anomaly index of the bridge component.

2. The method for real-time anomaly prediction and fault diagnosis of bridge components as described in claim 1, characterized in that, The formula for calculating the spatiotemporal correlation feature tensor is: in: For the spatiotemporal correlation feature tensor, For a three-dimensional feature tensor, , , For each attention head Independent projection matrix, To project the 3D feature tensor into a query vector, To project the 3D feature tensor into a key vector, To project a 3D feature tensor into a value vector, For merging and feature reconstruction matrices of multi-head outputs, For each attention head The subspace dimension, For time step, For spatial location, For feature dimension, This is the matrix transpose operation. For weight aggregation and multi-head component merging, Self-attention weight matrix , This is a feature enhanced in the same dimension.

3. The method for real-time anomaly prediction and fault diagnosis of bridge components as described in claim 1, characterized in that, The step of performing correlation analysis on the spatiotemporal correlation feature tensor based on the historical data to obtain the real-time spatiotemporal feature tensor of the three-dimensional feature tensor includes: Cosine similarity is calculated on the historical data and the spatiotemporal correlation feature tensor to obtain the historical similarity of the bridge components. Based on the historical similarity, a similarity threshold analysis is performed on the spatiotemporal correlation feature tensor to obtain the feature candidate set of the spatiotemporal correlation feature tensor. The weights corresponding to the feature candidate set are updated based on preset optimization steps to obtain the updated fusion weight matrix of the spatiotemporal correlation feature tensor. The parameters in the calculation formula of the spatiotemporal correlation feature tensor are updated using the updated fusion weight matrix to obtain the real-time spatiotemporal feature tensor of the three-dimensional feature tensor.

4. The method for real-time anomaly prediction and fault diagnosis of bridge components as described in claim 2, characterized in that, The preset optimization steps include: D1: Based on the candidate weight set, perform attention fusion on the self-attention weight matrix to obtain the final fused weight matrix. The calculation formula for attention fusion is as follows: in: For the weighted candidate set, For adaptive mixing coefficients, Here is the self-attention weight matrix. The final fusion weight matrix, The median of the candidate weight set; D2: Update the final fusion weight matrix based on a preset update algorithm to obtain the updated fusion weight matrix corresponding to the final fusion weight matrix. The calculation formula of the update algorithm is as follows: in: Let be the objective function. The regularization coefficient is . Let KL divergence be the KL divergence. It is the L2 norm. According to The generated prediction output, To and The corresponding actual sensor measurement value, For the first The updated fusion weight matrix after the next iteration is also For the final fused weight matrix, The historical weights corresponding to the historical data. For the first The gradient of the objective function with respect to the set of hyperparameters in the next iteration. For the first The approximate Hessian inverse matrix of the nth iteration For the first The dynamic learning rate of each iteration. The number of iterations is obtained based on the historical data. For the first The set of hyperparameters for the next iteration For the first The set of hyperparameters for each iteration.

5. The method for real-time anomaly prediction and fault diagnosis of bridge components as described in claim 4, characterized in that, The step of updating the parameters in the calculation formula of the spatiotemporal correlation feature tensor using the updated fusion weight matrix to obtain the real-time spatiotemporal feature tensor of the three-dimensional feature tensor includes: The self-attention weight matrix in the calculation formula of the spatiotemporal correlation feature tensor is updated using the updated fusion weight matrix to obtain the real-time spatiotemporal feature tensor of the three-dimensional feature tensor. The calculation formula of the real-time spatiotemporal feature tensor is as follows: in: For real-time spatiotemporal feature tensors, For a three-dimensional feature tensor, For each attention head after the update Independent projection matrix, For the updated value vector, For merging and feature reconstruction matrices of multi-head outputs, For time step, For spatial location, For feature dimension, For weight aggregation and multi-head component merging, For the first The updated fusion weight matrix after the next iteration is also For the final fused weight matrix, This is a feature enhanced in the same dimension.

6. The method for real-time anomaly prediction and fault diagnosis of bridge components as described in claim 1, characterized in that, The step of performing frequency domain spectral coherence analysis on the real-time spatiotemporal feature tensor and the modal parameter set to obtain the spectral coherence coefficients of the bridge components includes: A time window alignment operation is performed on the real-time spatiotemporal feature tensor and the modal parameter set to obtain a synchronized data segment set of the bridge component. Based on the synchronized data segment set, the frequency domain amplitude spectrum of the real-time spatiotemporal feature tensor is extracted to obtain the dynamic energy distribution spectrum of the bridge component. Based on the natural frequencies and bandwidth parameters in the modal parameter set, narrowband pulse spectrum analysis is performed on the dynamic energy distribution spectrum to obtain the modal reference spectrum of the bridge component; The bridge component is subjected to spectral coherence analysis based on the dynamic energy distribution spectrum and the modal reference spectrum to obtain the spectral coherence coefficient of the bridge component.

7. The method for real-time anomaly prediction and fault diagnosis of bridge components as described in claim 6, characterized in that, The step of performing spectral coherence analysis on the bridge component based on the dynamic energy distribution spectrum and the modal reference spectrum to obtain the spectral coherence coefficient of the bridge component includes: Based on the dynamic energy distribution spectrum and the modal reference spectrum, local energy correlation analysis is performed on the amplitude of frequency points in the bridge component to obtain the local coherence scaling value of the bridge component; The effective frequency bands in the historical data are weighted and fused based on the local coherence scaling value and the modal confidence weights corresponding to the modal parameter set to obtain the spectral coherence coefficient of the bridge component.

8. The method for real-time anomaly prediction and fault diagnosis of bridge components as described in claim 1, characterized in that, Short-time abrupt change analysis is performed on the spatiotemporal attention feature tensor to obtain the short-time abrupt change feature tensor of the bridge component, including: Based on a pre-trained BiLSTM model, forward and reverse time dependency learning is performed on the spatiotemporal attention feature tensor to obtain the temporal comprehensive feature sequence of the spatiotemporal attention feature tensor. Local feature extraction is performed on the time-series composite feature sequence to obtain the local gradient sharpness of the time-series composite feature sequence; Based on the local gradient sharpness, a mutation sharpness analysis is performed on the spatiotemporal attention feature tensor to obtain the mutation feature tensor of the spatiotemporal attention feature tensor.

9. The method for real-time anomaly prediction and fault diagnosis of bridge components as described in claim 1, characterized in that, The step of performing a dynamic threshold comparison between the short-term mutation feature tensor and the reference feature tensor in the historical data to obtain the real-time anomaly index of the bridge component includes: The difference between the short-term abrupt change feature tensor and the reference feature tensor is calculated based on the dynamic Mahalanobis distance to obtain the real-time abnormal deviation of the bridge component; The real-time anomaly deviation is normalized using the Sigmoid function, and then the normalized real-time anomaly deviation is exponentially amplified using the adjustment coefficient in the historical data to obtain the real-time anomaly index of the bridge component.

10. A real-time anomaly prediction and fault diagnosis system for bridge components, characterized in that, The system includes: Data processing module: used to collect real-time and historical data of bridge components, perform modal parameter identification on the real-time data to obtain the modal parameter set of the bridge components, and construct the three-dimensional feature tensor of the bridge components based on the real-time data; Transformer processing module: used to perform multi-head self-attention mechanism operation on the three-dimensional feature tensor to obtain the spatiotemporal correlation feature tensor of the three-dimensional feature tensor, and to perform correlation analysis on the spatiotemporal correlation feature tensor based on the historical data to obtain the real-time spatiotemporal feature tensor of the three-dimensional feature tensor; Feature processing module: used to perform frequency domain spectral coherence analysis on the real-time spatiotemporal feature tensor and the modal parameter set to obtain the spectral coherence coefficient of the bridge component, and to perform gated dynamic weighted calculation on the modal parameter set based on the spectral coherence coefficient to obtain the spatiotemporal attention feature tensor of the modal parameter set; BiLSTM processing module: used to perform short-time mutation analysis on the spatiotemporal attention feature tensor to obtain the short-time mutation feature tensor of the bridge component; Analysis module: used to perform dynamic threshold comparison between the short-term mutation feature tensor and the reference feature tensor in the historical data to obtain the real-time anomaly index of the bridge component.

Images