Bridge monitoring data diagnosis method fusing improved graph neural network

Abstract

The application provides a bridge monitoring data diagnosis method fusing an improved graph neural network, and belongs to the technical field of data diagnosis. The method comprises the following steps: collecting segmented time series data corresponding to each sliding time window of a bridge; constructing a dynamic weighted graph topology based on dynamic correlation coefficients of monitoring data between monitoring points in the sliding time window and spatial connection relationships between components of the bridge, and determining node features of each graph node in the dynamic weighted graph topology; constructing an improved graph neural network with damage propagation path constraints, iteratively updating the dynamic weighted graph topology by using a multi-layer graph convolution unit of the improved graph neural network, and outputting deep spatial correlation features; constructing a spatio-temporal cross-attention fusion layer with damage propagation chain perception, and identifying a spatio-temporal fusion feature vector; and outputting a diagnosis result containing a damage propagation chain sequence based on the spatio-temporal fusion feature vector. The technical scheme provided by the application can realize efficient linkage diagnosis of bridge monitoring data and structural state.

Description

Technical Field

[0001] This application relates to the field of data diagnostics technology, and more specifically, to a bridge monitoring data diagnostics method that integrates an improved graph neural network. Background Technology

[0002] As a core hub of the transportation network, long-term health monitoring and accurate data diagnosis are crucial aspects of bridge operation and maintenance. Currently, bridge monitoring systems are equipped with various sensors, including those for strain, displacement, tilt, acceleration, temperature, and humidity, generating massive amounts of high-dimensional heterogeneous time-series data daily. Traditional diagnostic methods rely on single-index threshold judgments, shallow statistical analysis, or conventional machine learning models, which have many technical shortcomings that are difficult to overcome.

[0003] On the one hand, there is a strong spatial correlation between bridge beams, piers, bearings and other components, and there is a coupling effect between monitoring points and data of various indicators. Traditional methods cannot uncover deep structural correlation features and have extremely low sensitivity to identify early minor damage and hidden anomalies, making it easy to miss detections. On the other hand, monitoring data is easily affected by environmental factors such as temperature, humidity and traffic load, has high noise redundancy, weak generalization ability of conventional models, and is prone to false alarms, making it difficult to meet the needs of refined operation and maintenance.

[0004] Existing bridge monitoring patents integrating GCN (Graph Neural Network) technologies, such as Chinese patent publication number CN119961762A, relate to the fields of artificial intelligence and data processing technology. Specifically, it relates to a method and system for identifying abnormal data in bridge construction monitoring. The method involves: collecting bridge vibration signal data through sensors and equipment; cleaning, converting, and standardizing the collected data; constructing an abnormal data identification model for bridge construction monitoring based on a graph neural network model; and training the model with the processed data to identify abnormal situations. The system analyzes the abnormal identification results based on collected bridge structural information, hydrological and meteorological information, real-time monitoring information, and maintenance and traffic information and methods, and transmits the analysis results to the front end. This invention, by constructing and training a bridge construction monitoring abnormal data identification model based on a graph neural network model, can monitor complex vibration signal data during bridge construction and better identify abnormal situations. It is evident that this scheme mainly adopts a general graph convolutional network, without optimizing the topology modeling and network structure for the characteristics of bridge structures. It only performs conventional feature extraction and diagnosis, resulting in severe homogenization of technical solutions. It suffers from insufficient diagnostic accuracy and poor environmental adaptability, and cannot achieve efficient linkage diagnosis between monitoring data and structural status. Therefore, how to achieve efficient linkage diagnosis between bridge monitoring data and structural status has become an urgent technical problem to be solved in this field. Summary of the Invention

[0005] This application provides a bridge monitoring data diagnosis method that integrates an improved graph neural network, which can achieve efficient linkage diagnosis of bridge monitoring data and structural status.

[0006] In a first aspect, this application provides a bridge monitoring data diagnosis method that integrates an improved graph neural network, comprising the following steps: Collect multi-source time-series monitoring data of each monitoring point on the bridge within a preset time period, and divide the multi-source time-series monitoring data into segmented time-series data corresponding to each sliding time window; For segmented time series data within the sliding time window, the dynamic correlation coefficient between monitoring data at each monitoring point is calculated, and a dynamic weighted graph topology is constructed by combining the spatial connection relationship between various bridge components. The multi-source time series monitoring data within the sliding time window is mapped to the node characteristics of each graph node in the dynamic weighted graph topology. An improved graph neural network for damage propagation path constraints is constructed. The improved graph neural network for damage propagation path constraints consists of multiple layers of graph convolutional units cascaded in sequence. Each graph convolutional unit includes a graph convolutional layer and an attention weight adjustment layer for damage propagation path constraints. The dynamic weighted graph topology is iteratively updated by the multi-layer graph convolutional units of the improved graph neural network, and the deep spatial correlation features are output after the iterative update by the multi-layer graph convolutional units. A spatiotemporal cross-attention fusion layer for damage propagation chain perception is constructed. The spatiotemporal cross-attention fusion layer uses the deep spatial correlation features as query vectors and the temporal evolution features of segmented temporal data within a sliding time window as key vectors and value vectors, and performs cross-dimensional feature interaction to obtain spatiotemporal fusion feature vectors. The spatiotemporal fusion feature vector is input into the optimized classification and discrimination layer for data diagnosis, and the output includes the diagnosis result containing the damage propagation chain sequence.

[0007] Furthermore, dividing the multi-source time-series monitoring data into segmented time-series data corresponding to each sliding time window specifically includes: The multi-source time-series monitoring data is extracted at a preset time step to obtain the synchronous monitoring vector of each monitoring point at a single time step; Arrange the synchronous monitoring vectors according to the order of their acquisition times to form a multi-source monitoring time sequence matrix with the number of sampling points as its length. A windowed multi-source time series submatrix is ​​obtained by continuously sliding and truncating along the time dimension of the multi-source monitoring time series matrix using a sliding window with a preset window length and sliding step size. The synchronized monitoring vector sequence contained in each of the windowed multi-source time series sub-matrices constitutes the segmented time series data corresponding to each sliding time window.

[0008] Furthermore, for the segmented time-series data within the sliding time window, the dynamic correlation coefficient between the monitoring data at each monitoring point is calculated, and a dynamic weighted graph topology is constructed by combining the spatial connection relationships between the various components of the bridge. Specifically, this includes: Extract the monitoring data sequences of any two monitoring points within the sliding time window from the segmented time series data, and calculate the Pearson correlation coefficient between the monitoring data sequences of the two monitoring points. The Pearson correlation coefficient is used as the dynamic correlation coefficient between the two monitoring points within the sliding time window. The dynamic correlation matrix is ​​constructed by iterating through the dynamic correlation coefficients between all pairs of monitoring points. Obtain the spatial connection relationship adjacency matrix between each component of the bridge, and multiply the spatial connection relationship adjacency matrix element by element with the dynamic correlation matrix to obtain the dynamic weighted adjacency matrix; Using each monitoring point as a graph node and the aforementioned dynamic weighted adjacency matrix as edge weights, a dynamic weighted graph topology is constructed.

[0009] Furthermore, mapping the multi-source time-series monitoring data within the sliding time window to the node features of each graph node in the dynamically weighted graph topology specifically includes: Temporal features are extracted from the monitoring data sequences of each monitoring point in the segmented time series data within the sliding time window to obtain the local temporal feature vector of each monitoring point. Using the monitoring points corresponding to each graph node in the dynamic weighted graph topology as indexes, the local time-series feature vectors are assigned to the corresponding graph nodes to form the node features of each graph node.

[0010] Furthermore, an improved graph neural network for damage propagation path constraints is constructed. This improved graph neural network consists of multiple cascaded graph convolutional units. Each graph convolutional unit includes a graph convolutional layer and an attention weight adjustment layer for damage propagation path constraints. Specifically, it includes: A graph convolutional layer is constructed, which takes the adjacency matrix and node feature matrix of the dynamic weighted graph topology as input, and outputs the node embedding feature matrix through graph convolution operation; A damage propagation path constraint attention weight adjustment layer is cascaded after the graph convolutional layer. The damage propagation path constraint attention weight adjustment layer takes the node embedding feature matrix and the damage propagation path constraint matrix between each component of the bridge as input, calculates the constraint attention weight between each adjacent node through the path constraint attention mechanism, and performs weighted aggregation on the node embedding feature matrix according to the constraint attention weight, and outputs a node feature matrix with enhanced path constraints. The graph convolutional layer and the attention weight adjustment layer for the damage propagation path constraint are combined into a graph convolutional unit. Multiple graph convolutional units are cascaded in sequence, and the node feature matrix of the path constraint enhancement output by the previous graph convolutional unit is used as the input of the node feature matrix of the graph convolutional layer in the next graph convolutional unit, thus forming an improved graph neural network for damage propagation path constraint.

[0011] Furthermore, the improved graph neural network uses multi-layer graph convolutional units to iteratively update the dynamic weighted graph topology. The deep spatial correlation features output after the multi-layer graph convolutional units iteratively update specifically include: The adjacency matrix of the dynamically weighted graph topology is used as the input of the graph convolutional layer of the first graph convolutional unit of the improved graph neural network, and the node feature matrix is ​​used as the input of the graph convolutional layer of the first graph convolutional unit. The first graph convolution unit performs graph convolution operations on the adjacency matrix and the node feature matrix, and adjusts the attention weights of the damage propagation path constraints, outputting the node feature matrix with enhanced path constraints and the attention weight matrix of the damage propagation path constraints corresponding to the first graph convolution unit. The attention weight matrix output by the first graph convolutional unit is multiplied element-wise with the adjacency matrix of the dynamically weighted graph topology to obtain the iteratively updated adjacency matrix. The iteratively updated adjacency matrix and the path-constrained node feature matrix output by the first graph convolutional unit are used as the graph convolutional layer input of the next graph convolutional unit. The iterative update of multiple graph convolutional units is performed sequentially, and the path-constrained node feature matrix output by the last graph convolutional unit is used as the deep spatial association feature.

[0012] Furthermore, the spatiotemporal cross-attention fusion layer uses the deep spatial correlation features as the query vector and the temporal evolution features of segmented time-series data within the sliding time window as the key and value vectors, and performs cross-dimensional feature interaction to obtain the spatiotemporal fusion feature vector, specifically including: A linear transformation is performed on the deep spatial correlation features to generate a query matrix as a query vector. Linear transformations are performed on the temporal evolution characteristics of the segmented time series data to generate a key matrix as a key vector and a value matrix as a value vector. Calculate the dot product similarity matrix between the query vector and the key vector, and construct a spatiotemporal cross-attention weight matrix by combining the damage propagation chain position encoding vector with the dot product similarity matrix; The spatiotemporal cross-attention weight matrix and the value vector are weighted and summed to obtain the spatiotemporal fusion feature vector.

[0013] Secondly, this application provides a computer device including a memory and a processor, the memory storing code, and the processor being configured to acquire the code and execute the above-described bridge monitoring data diagnosis method incorporating an improved graph neural network.

[0014] Thirdly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described bridge monitoring data diagnosis method based on the fusion of improved graph neural networks.

[0015] The technical solutions provided by the embodiments disclosed in this application have the following beneficial effects: The bridge monitoring data diagnosis method integrating an improved graph neural network provided in this application first collects multi-source time-series monitoring data from various monitoring points of the bridge within a preset time period, and divides the multi-source time-series monitoring data into segmented time-series data corresponding to each sliding time window; secondly, for the segmented time-series data within the sliding time window, the dynamic correlation coefficient between the monitoring data of each monitoring point is calculated, and a dynamic weighted graph topology is constructed in combination with the spatial connection relationship between various bridge components, mapping the multi-source time-series monitoring data within the sliding time window to the node features of each graph node in the dynamic weighted graph topology; then, an improved graph neural network constraining damage propagation paths is constructed, which is composed of cascaded multilayer graph convolutional units. Each graph convolutional unit includes a graph convolutional layer and an attention weight adjustment layer for damage propagation path constraints. Further, the improved graph neural network's multi-layer graph convolutional units iteratively update the dynamically weighted graph topology, outputting deep spatial correlation features after the iterative update. Further still, a spatiotemporal cross-attention fusion layer for damage propagation chain awareness is constructed. This layer uses the deep spatial correlation features as the query vector and the temporal evolution features of segmented time-series data within a sliding time window as the key and value vectors, performing cross-dimensional feature interaction to obtain a spatiotemporal fusion feature vector. Finally, the spatiotemporal fusion feature vector is input into an optimized classification and discrimination layer for data diagnosis, outputting a diagnostic result containing the damage propagation chain sequence.

[0016] Therefore, this application can achieve efficient linkage diagnosis of bridge monitoring data and structural status. First, by dividing multi-source time-series monitoring data into segmented time-series data corresponding to sliding time windows, fine-grained time-domain segmentation of the continuous bridge monitoring data stream is achieved, enabling damage diagnosis to be performed independently for each local time segment, avoiding the defect of smoothing out local anomalies when performing global analysis of the entire data segment. Second, by calculating the dynamic correlation coefficient between monitoring data at each monitoring point and combining it with the spatial connection relationship between various bridge components to construct a dynamic weighted graph topology, physical topology priors and data-driven dynamic correlation are organically integrated to retain dynamic correlation between measuring point pairs with real mechanical force transmission paths. The correlation edge weights effectively filter out the interference of spurious statistical correlations between unconnected measurement points on subsequent graph learning. Furthermore, by constructing an improved graph neural network constraining damage propagation paths, consisting of cascaded multi-layer graph convolutional units, each graph convolutional unit alternately embeds graph convolutional layers and attention weight adjustment layers for damage propagation path constraints. This layer-by-layer combines the neighborhood smoothing aggregation capability of graph convolution with the directed attention selective aggregation capability of damage propagation paths, forcing node information to flow in a directional direction allowed by the transmission of damage mechanics failure, overcoming the limitation of traditional graph neural networks in distinguishing the directionality of connections between nodes. Then, the dynamic weighted graph topology is iteratively updated through multi-layer graph convolutional units. Each layer performs feature transformation on the graph topology updated by element-wise multiplication of the attention weight matrix and adjacency matrix of the previous layer. This allows the edge weights of the adjacency matrix to gradually strengthen adjacent edges that are consistent with the damage propagation direction and have high feature relevance during iteration. The final output deep spatial association features retain the hierarchical directional constraints of the damage propagation path within the progressively expanding receptive field. Then, by constructing a spatiotemporal cross-attention fusion layer that perceives the damage propagation chain, using deep spatial association features as query vectors and temporal evolution features as key and value vectors, the deep spatial context, which has been aggregated according to the damage propagation direction, is used as query conditions for cross-modal interactive matching with the local temporal evolution patterns in the temporal dimension. This approach achieves coupled modeling of the spatial transmission and temporal evolution patterns of damage propagation chains, avoiding the fragmentation of two-dimensional information caused by simply splicing together spatial correlation and temporal evolution after processing them independently. Finally, by inputting spatiotemporal fusion feature vectors into the optimized classification and discrimination layer and combining them with damage propagation chain location encoding vectors for damage propagation path tracing, a structured diagnostic result containing the damage propagation chain sequence is output. This allows the diagnostic conclusions to not only include where and to what extent the damage occurred, but also to further reveal the path of force transmission along the structure, providing bridge maintenance departments with a complete decision-making basis from damage identification to propagation source tracing. In summary, the technical solution provided in this application can achieve efficient linkage diagnosis of bridge monitoring data and structural status. Attached Figure Description

[0017] Figure 1This is an exemplary flowchart of a bridge monitoring data diagnosis method that integrates an improved graph neural network, according to some embodiments of this application. Figure 2 This is an exemplary flowchart illustrating the determination of a dynamic weighted graph topology according to some embodiments of this application; Figure 3 This is a schematic diagram of the structure of a computer device that implements a bridge monitoring data diagnosis method based on an improved graph neural network, according to some embodiments of this application. Detailed Implementation

[0018] To better understand the technical solution of this application, the technical solution of this application will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0019] refer to Figure 1 The figure is an exemplary flowchart of a bridge monitoring data diagnosis method based on the fusion of improved graph neural networks according to some embodiments of this application. The figure mainly includes the following steps: In step S101, multi-source time-series monitoring data of each monitoring point on the bridge are collected within a preset time period, and the multi-source time-series monitoring data is divided into segmented time-series data corresponding to each sliding time window.

[0020] Specifically, multi-source time-series monitoring data of various monitoring points on the bridge are collected within a preset time period using various types of sensors. This means that various types of sensors are pre-deployed on the key components of the bridge, including but not limited to the main beam, bridge tower, stay cables, and supports. The various types of sensors include strain gauges, accelerometers, displacement gauges, and temperature sensors for monitoring structural response, as well as anemometers for monitoring environmental loads. These sensors continuously collect physical quantities at a preset sampling frequency. After analog-to-digital conversion and wavelet threshold denoising by the data acquisition unit, a unified timestamp is added, and missing values ​​are filled in by linear interpolation to complete the standardization and normalization process, eliminating dimensional differences, thereby forming a multi-source time-series monitoring dataset covering various monitoring points on the bridge.

[0021] It should be noted that the multi-source time-series monitoring dataset in this application refers to a high-dimensional structured data set formed by synchronously collecting data from various types of sensors deployed at different spatial locations and aligning them according to a unified time axis during the bridge structural health monitoring process. Each data point in this dataset is accompanied by a time label and a spatial location label, thereby unifying multimodal physical quantities such as bridge structural response information, environmental action information, and external load information on a common time axis, providing a foundation for subsequent analysis of the dynamic correlation between different monitoring points at different time scales and the damage propagation path.

[0022] The multi-source time-series monitoring data is divided into segmented time-series data corresponding to each sliding time window using the following steps: The multi-source time-series monitoring data is extracted at a preset time step to obtain the synchronous monitoring vector of each monitoring point at a single time step; Arrange the synchronous monitoring vectors according to the order of their acquisition times to form a multi-source monitoring time sequence matrix with the number of sampling points as its length. A windowed multi-source time series submatrix is ​​obtained by continuously sliding and truncating along the time dimension of the multi-source monitoring time series matrix using a sliding window with a preset window length and sliding step size. The synchronized monitoring vector sequence contained in each of the windowed multi-source time series sub-matrices constitutes the segmented time series data corresponding to each sliding time window.

[0023] In specific implementation, firstly, the multi-source time-series monitoring data is extracted at equal intervals using a preset time step. The time step refers to the fixed time interval between two adjacent data acquisitions. For example, for multi-source signals such as acceleration, displacement, and strain of each monitoring point synchronously acquired at a sampling frequency of 10Hz, the synchronous readings of all monitoring channels at each independent moment are extracted from the multi-source time-series monitoring data with a time step of 0.1 seconds, resulting in a synchronous monitoring vector for each monitoring point at a single time step. This synchronous monitoring vector is a one-dimensional array composed of multi-source monitoring values ​​acquired from all monitoring points at the same sampling moment. Secondly, the synchronous monitoring vectors are stacked vertically according to their corresponding timestamps, forming a multi-source monitoring time-series matrix with the number of sampling points as the number of rows and the number of monitoring points as the number of columns. This multi-source monitoring time-series matrix is ​​a two-dimensional array that organizes the original multi-source time-series monitoring data into a structured table format, where each row corresponds to a sampling moment and each column corresponds to a monitoring channel. Then, a sliding window with a preset window length and sliding step is used to... The time dimension of the multi-source monitoring time series matrix is ​​continuously trunculated. In practice, the window length is set to 100 sampling points and the sliding step size is 50 sampling points. The window moves from front to back along the row direction of the multi-source monitoring time series matrix with the sliding step size as the increment. Each move truncates continuous rows of data within the window range, resulting in a series of windowed multi-source time series sub-matrices that partially overlap in time. The windowed multi-source time series sub-matrices refer to two-dimensional subarrays composed of local continuous rows trunculated from the complete multi-source monitoring time series matrix. Each sub-matrix contains data from all monitoring channels within a fixed time span. Finally, the synchronous monitoring vector sequence contained in each windowed multi-source time series sub-matrix directly constitutes the segmented time series data of the corresponding sliding time window. The synchronous monitoring vector sequence refers to the ordered synchronous monitoring vector formed by arranging each row in the sub-matrix according to the original time order. This segmented time series data serves as the input for subsequent dynamic weighted graph topology construction and node feature mapping, providing an independent spatiotemporal analysis basic unit for each sliding time window.

[0024] It should be noted that the mid-segment time-series data in this application refers to an ordered set of multi-channel synchronous monitoring records belonging to the same local time period, obtained by dividing the data according to a preset time window length and sliding step size. This data is used to measure the local dynamic behavior patterns of physical quantities such as vibration, strain, and displacement sensed by each monitoring point of the bridge over time within a specific sliding time window. This provides structured temporal input for subsequent extraction of time-series evolution feature vectors and participation in spatiotemporal cross-attention fusion, thereby enabling the damage diagnosis model to capture the spatiotemporal coupling characteristics of transient spatial correlation anomalies and damage propagation chains during the occurrence and development of damage within a fine-grained local time window.

[0025] In step S102, for the segmented time series data within the sliding time window, the dynamic correlation coefficient between the monitoring data of each monitoring point is calculated, and a dynamic weighted graph topology is constructed by combining the spatial connection relationship between the various components of the bridge. The multi-source time series monitoring data within the sliding time window is mapped to the node features of each graph node in the dynamic weighted graph topology.

[0026] refer to Figure 2 As shown, this figure is an exemplary flowchart for determining the dynamic weighted graph topology according to this application. In this application, for segmented time-series data within a sliding time window, the dynamic correlation coefficient between monitoring data at each monitoring point is calculated, and the dynamic weighted graph topology is constructed by combining the spatial connection relationships between various bridge components using the following steps: In step S1021, the monitoring data sequences of any two monitoring points in the segmented time series data within the sliding time window are extracted, and the Pearson correlation coefficient between the monitoring data sequences of the two monitoring points is calculated. In step S1022, the Pearson correlation coefficient is used as the dynamic correlation coefficient between the two monitoring points within the sliding time window. The dynamic correlation coefficients between all pairs of monitoring points are traversed to construct a dynamic correlation matrix. In step S1023, the spatial connection relationship adjacency matrix between each component of the bridge is obtained, and the spatial connection relationship adjacency matrix is ​​multiplied element by element with the dynamic correlation matrix to obtain the dynamic weighted adjacency matrix. In step S1024, a dynamic weighted graph topology is constructed using each monitoring point as a graph node and the dynamic weighted adjacency matrix as the edge weight.

[0027] In specific implementation, firstly, the monitoring data sequence of any two monitoring points within the sliding time window is extracted from the segmented time-series data. The monitoring data sequence refers to a one-dimensional vector arranged chronologically of the multi-source monitoring values ​​collected at each sampling time within the window for a single monitoring point. For example, from the segmented time-series data containing 100 sampling points, reading sequences of length 100 corresponding to accelerometer A and strain gauge B are extracted respectively. The Pearson correlation coefficient formula is used to measure the correlation between the normalized monitoring data sequences of the two monitoring points, calculating the ratio of the product of the covariance and the standard deviation of the two sequences to obtain the correlation coefficient between the two monitoring points. The Pearson correlation coefficient between the monitoring data sequences of the monitoring points is used to measure the strength of the linear correlation between two sequences. A positive value indicates that the monitoring data sequences of the two monitoring points show the same trend within the window, while a negative value indicates the opposite trend. Next, the calculated Pearson correlation coefficient is used as the dynamic correlation coefficient between the two monitoring points within the sliding time window. This dynamic correlation coefficient is distinct from the static fixed correlation and is a time-varying measure of the association between monitoring points as the sliding window progresses. All monitoring points are paired in this manner, and the dynamic correlation coefficient of each pair is filled in according to the row and column numbers to construct the... A symmetric matrix with the number of monitoring points as the dimension and the main diagonal elements all being 1 is used as the dynamic correlation matrix. This dynamic correlation matrix is ​​a quantitative expression representing the data coupling between monitoring points within a corresponding time window. Then, the spatial connection relationship adjacency matrix between each bridge component is obtained. This spatial connection relationship adjacency matrix is ​​a pre-determined adjacency matrix based on the bridge finite element model or structural design drawings, using binary elements of 0 or 1 to represent whether there is a direct physical connection between components. For example, if there is a continuous rigid connection between main beam segment i and segment i+1, the corresponding matrix element is 1; if there is no direct connection between main beam segment i and the bridge tower, the value is 0. This spatial connection... The adjacency matrix is ​​multiplied element-wise with the dynamic correlation matrix, that is, the elements at corresponding row and column positions are multiplied pairwise to obtain the dynamic weighted adjacency matrix. The dynamic weighted adjacency matrix refers to the weighted connection representation that only retains the dynamic correlation coefficient between measurement point pairs with physical connection relationships, while setting the dynamic correlation coefficient between measurement point pairs without physical connection relationships to zero. This operation filters out the true correlation strength between measurement points with structural force transmission path support through physical topological constraints. Finally, using each monitoring point as a graph node set and the value of each element in the dynamic weighted adjacency matrix as the weight value of the connection edge between the corresponding node pairs, an undirected weighted graph data structure is constructed as the dynamic weighted graph topology structure.

[0028] It should be noted that the dynamic weighted graph topology in this application refers to the spatial distribution pattern and correlation strength of the dynamic coupling relationship between various monitoring areas of the bridge structure under the constraints of the static mechanical force transmission path within a local time period. Specifically, it is a dynamic graph representation in which nodes correspond to monitoring points and edge weights reflect the strength of the data correlation between physically connected measurement point pairs within the window. This structure simultaneously embeds the static spatial topology prior of the bridge and the data-driven dynamic correlation within the sliding window, providing a structure-aware graph input for the subsequent spatial feature extraction of the graph neural network. On the one hand, through the element-wise product constraint of the adjacency matrix of spatial connection relationship, it ensures that the information aggregation in the graph convolution operation strictly follows the real mechanical transmission path between the bridge components, avoiding the introduction of false statistical correlations between monitoring points that do not have physical connection relationships into the feature propagation process. On the other hand, by using the dynamic correlation coefficient as the continuous value of the edge weight, the graph topology can sensitively reflect the transient weakening or abrupt change in the correlation strength between measurement points on the same physical connection channel as damage occurs and expands, thereby enabling deep spatial correlation feature extraction to capture the spatial evolution law of damage propagation chain within a reasonable framework of structural constraints.

[0029] The following steps are used to map the multi-source time-series monitoring data within the sliding time window to the node features of each node in the dynamically weighted graph topology: Temporal features are extracted from the monitoring data sequences of each monitoring point in the segmented time series data within the sliding time window to obtain the local temporal feature vector of each monitoring point. Using the monitoring points corresponding to each graph node in the dynamic weighted graph topology as indexes, the local time-series feature vectors are assigned to the corresponding graph nodes to form the node features of each graph node.

[0030] In specific implementation, firstly, temporal features are extracted from the monitoring data sequences of each monitoring point within the sliding time window in the segmented time-series data. Taking the vibration data sequence of a certain accelerometer with 100 sampling points within the window as an example, the monitoring data sequence is input into a one-dimensional convolutional neural network for temporal convolution operations. Waveform morphology features within the local receptive field are extracted by sliding multi-scale convolution kernels along the time axis. After dimensionality reduction by pooling layers, a fixed-dimensional vector representation is output through a fully connected layer. Alternatively, a long short-term memory network is used to recursively encode the sequence step by step, and the last hidden state vector is taken as the overall temporal representation of the sequence, thereby obtaining the local temporal sequence of each monitoring point. The feature vector, specifically the local temporal feature vector, refers to a low-dimensional dense vector extracted from the monitoring data sequence of a single monitoring point within a sliding time window by a time encoder. This vector compresses and represents the local fluctuation patterns and trend changes in the sequence. Subsequently, using the monitoring point name corresponding to each graph node in the dynamic weighted graph topology as the index key, the extracted local temporal feature vectors are assigned to the corresponding graph nodes one by one according to the index matching relationship. This constructs a node feature matrix with the number of graph nodes as the number of rows and the dimension of the local temporal feature vector as the number of columns. During the assignment process, the one-to-one mapping relationship between the graph node number and the monitoring point number remains unchanged, thereby forming the node features of each graph node.

[0031] It should be noted that, in this application, node features refer to the initial feature representation of each graph node in the dynamic weighted graph topology within the sliding time window. Node features serve as the starting point and information source for the improved graph neural network constrained by damage propagation path to learn spatial correlation features. They encapsulate the local dynamic behavior patterns of the corresponding monitoring points extracted from the original monitoring data sequence by the time encoder within the sliding time window. This allows each graph node to carry the temporal characteristics of the structural response at the location of the measuring point from its initial state, providing feature conditions for the subsequent network to abstract deep spatial correlation features reflecting the damage propagation path layer by layer from the local temporal representation at the node level.

[0032] In step S103, an improved graph neural network for damage propagation path constraints is constructed. The improved graph neural network for damage propagation path constraints consists of multiple layers of graph convolutional units cascaded in sequence. Each graph convolutional unit includes a graph convolutional layer and an attention weight adjustment layer for damage propagation path constraints.

[0033] An improved graph neural network for damage propagation path constraints is constructed, which consists of multiple cascaded graph convolutional units. Each graph convolutional unit includes a graph convolutional layer and an attention weight adjustment layer for damage propagation path constraints. This is achieved through the following steps: A graph convolutional layer is constructed, which takes the adjacency matrix and node feature matrix of the dynamic weighted graph topology as input, and outputs the node embedding feature matrix through graph convolution operation; A damage propagation path constraint attention weight adjustment layer is cascaded after the graph convolutional layer. The damage propagation path constraint attention weight adjustment layer takes the node embedding feature matrix and the damage propagation path constraint matrix between each component of the bridge as input, calculates the constraint attention weight between each adjacent node through the path constraint attention mechanism, and performs weighted aggregation on the node embedding feature matrix according to the constraint attention weight, and outputs a node feature matrix with enhanced path constraints. The graph convolutional layer and the attention weight adjustment layer for the damage propagation path constraint are combined into a graph convolutional unit. Multiple graph convolutional units are cascaded in sequence, and the node feature matrix of the path constraint enhancement output by the previous graph convolutional unit is used as the input of the node feature matrix of the graph convolutional layer in the next graph convolutional unit, thus forming an improved graph neural network for damage propagation path constraint.

[0034] In specific implementation, firstly, a graph convolutional layer is constructed. This layer takes the adjacency matrix and node feature matrix of the dynamically weighted graph topology as input and performs graph convolution operations using the message-passing paradigm of graph convolutional networks. Specifically, after adding self-loops to the adjacency matrix, it performs symmetric normalization with the inverse square root of the degree matrix to obtain a normalized adjacency matrix. Then, it performs matrix multiplication with the node feature matrix. The product is multiplied by the learnable weight parameter matrix and then mapped through a non-linear activation function to output a node embedding feature matrix. This node embedding feature matrix refers to the hidden layer where each graph node aggregates first-order neighborhood information after a single-layer graph convolution. Feature representation; secondly, after the convolutional layer of this graph, an attention weight adjustment layer for damage propagation path constraints is cascaded. First, the damage propagation path constraint matrix between each component of the bridge is obtained. The damage propagation path constraint matrix refers to a directed constraint matrix pre-constructed based on the mechanical failure transmission mechanism of the bridge structure, used to indicate feasible paths for damage to propagate from one component to adjacent components along a specific direction. For example, for a continuous beam bridge, the damage propagation path constraint matrix stipulates that support damage only propagates unidirectionally to adjacent main beam segments, and main beam segment damage can propagate bidirectionally to adjacent segments and the support below. Elements with a value of 1 in this matrix indicate the existence of an allowed damage propagation direction. Elements with a value of 0 indicate path obstruction. Subsequently, the attention weight adjustment layer for the damage propagation path constraint takes the node embedding feature matrix and the damage propagation path constraint matrix as input and executes the path constraint attention mechanism. Specifically, the node embedding feature matrix is ​​parameterized and mapped through the first and second learnable weight matrices to generate a query matrix and a key matrix. The transpose of the query matrix and the key matrix is ​​then multiplied by a dot product to obtain the original attention score matrix. Finally, the damage propagation path constraint matrix is ​​multiplied element-wise by the original attention score matrix and a negative infinity bias is applied to the zero position of the constraint matrix. After normalization along the row direction using the Softmax function, a constrained attention weight matrix is ​​obtained. The constrained attention weight matrix refers to a normalized attention coefficient matrix that is non-zero only in the propagation direction allowed by the damage propagation path constraint matrix, and whose element values ​​reflect the feature similarity between neighboring nodes. The constrained attention weight matrix and the value matrix generated by the node embedding feature matrix through the third learnable weight matrix are multiplied to achieve weighted aggregation, and the path constraint enhanced node feature matrix is ​​output. The path constraint enhanced node feature matrix refers to the node-level feature representation after reassigning the contribution weight of neighboring node information through the attention mechanism of damage propagation path constraint.Finally, the graph convolutional layer and the attention weight adjustment layer for the damage propagation path constraint are combined into a single graph convolutional unit. Multiple graph convolutional units are stacked sequentially in this manner to form a multi-layer cascaded architecture. The input to each graph convolutional unit comes from the node feature matrix of the path constraint enhancement output of the previous graph convolutional unit. The input to the first graph convolutional unit is the node feature matrix of the dynamically weighted graph topology. This cascading structure constitutes an improved graph neural network for damage propagation path constraints.

[0035] In this embodiment, the first, second, and third learnable weight matrices are three independent, identical-dimensional, trainable parameter matrices in the attention weight adjustment layer constrained by the damage propagation path. These matrices are jointly optimized with other model parameters via backpropagation and are used to map the input node embedding feature matrix into the query matrix, key matrix, and value matrix required by the attention mechanism, respectively. In a specific implementation, the dimension of the node embedding feature matrix is ​​set to N×D, where N is the number of graph nodes and D is the dimension of the node embedding feature vector. The first learnable weight matrix... The dimension is D× Its function is to linearly project the node embedding feature matrix from the original D-dimensional feature space to... A dimensional query subspace is used to generate a query matrix to represent the proactive query needs of each target node for neighborhood information; a second learnable weight matrix is ​​used. The dimension is D× Its function is to linearly project the node embedding feature matrix onto the same... A 3D key subspace is used to generate a key matrix to represent the feature identifiers of each neighboring node that can be matched, where Indicates the dimension of the query-key shared subspace; third learnable weight matrix. The dimension is D× Its function is to linearly project the node embedding feature matrix onto... A dimensional value subspace is used to generate a value matrix to carry the actual feature content of each neighboring node to be weighted and aggregated. This represents the dimension of the feature subspace. For example, for bridge monitoring data with high dimensionality, it can be set to... =64 to reduce the computational cost of dot product calculations, set =128 to retain richer aggregation feature information; each element in these three matrices is initially randomly assigned using Xavier uniform initialization or He Kaiming initialization. During the model training phase, the values ​​of each element are updated successively according to the derivative of the loss function with respect to the network parameters using gradient descent until the model converges and the final learnable weight parameters are solidified. This enables the attention mechanism to self-learn and adjust the matching relationship between the query vector and the key vector, as well as the contribution weight of the value vector to the aggregation output, within the feasible propagation direction limited by the damage propagation path constraint matrix.

[0036] It should be noted that the improved graph neural network for damage propagation path constraints in this application is a deep graph learning model that alternately embeds the neighborhood smoothing property of graph convolution and the attention-selective aggregation property of damage propagation path constraints layer by layer. It gradually expands the receptive field of nodes under the constraint of structural force transmission links through multi-layer iterations, and ensures that information flows in a directional manner along the feasible direction of damage propagation to capture the spatial step-by-step transmission features of the damage propagation chain. The graph convolution layer in each graph convolution unit of the improved graph neural network for damage propagation path constraints is responsible for smoothly aggregating the first-order neighborhood information of each graph node along the dynamically weighted edges to achieve local fusion of spatial associations. The attention weight adjustment layer for damage propagation path constraints that follows takes the damage propagation path constraint matrix between the bridge components as a priori constraint, and performs directed reweighting on the embedded features of the aggregated nodes through the attention mechanism of path constraints. It forces information to flow only along the feasible path direction allowed by the transmission of damage mechanical failure and adaptively allocates the information contribution weight of the neighborhood nodes. In this way, redundant associations unrelated to damage propagation are gradually filtered out in the layer-by-layer feature transformation, and the spatial dependency patterns that are indicative of the damage propagation chain are enhanced.

[0037] In step S104, the dynamic weighted graph topology is iteratively updated by the multi-layer graph convolutional unit of the improved graph neural network, and the deep spatial correlation features are output after the iterative update by the multi-layer graph convolutional unit.

[0038] The improved graph neural network iteratively updates the dynamic weighted graph topology using multi-layer graph convolutional units. The deep spatial correlation features output after the iterative update by the multi-layer graph convolutional units are achieved through the following steps: The adjacency matrix of the dynamically weighted graph topology is used as the input of the graph convolutional layer of the first graph convolutional unit of the improved graph neural network, and the node feature matrix is ​​used as the input of the graph convolutional layer of the first graph convolutional unit. The first graph convolution unit performs graph convolution operations on the adjacency matrix and the node feature matrix, and adjusts the attention weights of the damage propagation path constraints, outputting the node feature matrix with enhanced path constraints and the attention weight matrix of the damage propagation path constraints corresponding to the first graph convolution unit. The attention weight matrix output by the first graph convolutional unit is multiplied element-wise with the adjacency matrix of the dynamically weighted graph topology to obtain the iteratively updated adjacency matrix. The iteratively updated adjacency matrix and the path-constrained node feature matrix output by the first graph convolutional unit are used as the graph convolutional layer input of the next graph convolutional unit. The iterative update of multiple graph convolutional units is performed sequentially, and the path-constrained node feature matrix output by the last graph convolutional unit is used as the deep spatial association feature.

[0039] In specific implementation, firstly, the adjacency matrix and node feature matrix of the dynamically weighted graph topology are simultaneously used as inputs to the graph convolutional layer of the first graph convolutional unit of the improved graph neural network. The adjacency matrix refers to the weighted connection matrix obtained by filtering the spatial connection relationship adjacency matrix with dynamic correlation coefficients as edge weights. The node feature matrix refers to the feature matrix formed by arranging the local temporal feature vectors of each monitoring point according to the graph node index. Secondly, within the first graph convolutional unit, the graph convolutional layer performs graph convolution operations with the adjacency matrix and the node feature matrix as inputs. Specifically, through matrix multiplication of the normalized adjacency matrix and the node feature matrix and linear transformation of the learnable weight parameter matrix, the node embedding feature matrix is ​​output. Subsequently, the attention weight adjustment layer for damage propagation path constraints takes the node embedding feature matrix and the damage propagation path constraint matrix between the bridge components as inputs. Through the attention mechanism of path constraints, the node embedding feature matrix is ​​mapped to a query matrix, a key matrix, and a value matrix through the first learnable weight matrix, the second learnable weight matrix, and the third learnable weight matrix, respectively. The transpose of the query matrix and the key matrix is ​​then processed... The original attention score matrix is ​​obtained by row dot product operation. Then, the damage propagation path constraint matrix is ​​multiplied element-wise with the original attention score matrix, and a negative infinity bias is applied to the zero-value positions of the constraint matrix. After Softmax normalization, the attention weight matrix of damage propagation path constraint is output. The attention weight matrix of damage propagation path constraint refers to the neighboring node attention coefficient matrix that has non-zero normalized weights only in the feasible path direction of damage propagation. Finally, the attention weight matrix is ​​multiplied with the value matrix to achieve weighted aggregation, and the node feature matrix of path constraint enhancement corresponding to the first graph convolution unit is output. The node feature matrix of path constraint enhancement refers to the node-level feature representation after selective aggregation of neighborhood information by directed attention. Then, the attention weight matrix of damage propagation path constraint output by the first graph convolution unit is multiplied element-wise with the adjacency matrix of the dynamic weighted graph topology, that is, the elements at corresponding row and column positions are multiplied pairwise to obtain the iteratively updated adjacency matrix. The iteratively updated adjacency matrix refers to the new connection strength matrix after the edge weight values ​​of the original dynamic weighted adjacency matrix are adaptively adjusted by the attention mechanism.Subsequently, the iteratively updated adjacency matrix and the path-constrained enhanced node feature matrix output by the first graph convolutional unit are used as the input to the graph convolutional layer of the second graph convolutional unit. The second graph convolutional unit repeats the same graph convolution operation and attention weight adjustment operation for damage propagation path constraints as the first graph convolutional unit, and multiplies its own output attention weight matrix element-wise with the iteratively updated adjacency matrix to obtain the adjacency matrix for the next iteration. This iterative update of multiple graph convolutional units is performed sequentially. Each graph convolutional unit performs feature transformation and neighborhood aggregation based on the updated adjacency matrix and node features of the previous layer, until the last graph convolutional unit completes its calculation. Afterward, its output path-constrained enhanced node feature matrix is ​​used as the deep spatial association feature.

[0040] It should be noted that the deep spatial correlation features in this application refer to the deep correlation features of the spatial dependence patterns and the step-by-step transmission of the damage propagation chain presented by each monitoring point of the bridge within the sliding time window after multi-level information interaction with neighboring nodes under the constraints of the structural force transmission path. That is, each node feature vector encodes spatial context information that is gradually aggregated from its multi-hop neighborhood along the feasible damage propagation direction, and the aggregation path of this information is strictly controlled by the direction of the damage propagation path constraint matrix. This enables the subsequent spatiotemporal cross-attention mechanism to accurately retrieve the time-varying patterns related to the damage propagation path from the temporal evolution features under the spatial semantic guidance of the clear damage propagation direction, and generate a spatiotemporal fusion feature vector that fully couples the spatial transmission law and temporal evolution law of the damage propagation chain.

[0041] In step S105, a spatiotemporal cross-attention fusion layer for damage propagation chain perception is constructed. The spatiotemporal cross-attention fusion layer uses the deep spatial correlation features as query vectors and the temporal evolution features of segmented time series data within the sliding time window as key vectors and value vectors to perform cross-dimensional feature interaction to obtain a spatiotemporal fusion feature vector.

[0042] The spatiotemporal cross-attention fusion layer for damage propagation chain awareness is constructed using the following steps: The monitoring data sequence of each monitoring point in the segmented time series data within the sliding time window is encoded with time domain features, the time evolution feature vector of each monitoring point is extracted, and the time evolution feature vectors of all monitoring points are concatenated into a time evolution feature matrix. A cross-attention mechanism module is constructed. The cross-attention mechanism module uses the query matrix generated by the linear transformation of the deep spatial correlation features as the query vector, and the key matrix and value matrix generated by the linear transformation of the temporal evolution feature matrix as the key vector and value vector, respectively. The cross-dimensional interaction weight is calculated by scaling dot product attention, and the value vector is weighted and aggregated to obtain the spatiotemporal fusion feature matrix. In the cross-attention mechanism module, a damage propagation chain position encoding vector is introduced. After the damage propagation chain position encoding vector is added element-wise to the query matrix and the key matrix, the scaling dot product attention calculation is performed to form a spatiotemporal cross-attention fusion layer for damage propagation chain perception.

[0043] In specific implementation, firstly, the monitoring data sequence of each monitoring point in the segmented time series data within the sliding time window is encoded with temporal features. Taking the strain data sequence of a strain gauge with a length of 100 sampling points within the window as an example, the monitoring data sequence is input into a time series encoder based on a gated recurrent unit or a one-dimensional convolutional network. Through recursive encoding along the time axis or convolution-pooling operations, the temporal evolution patterns such as trend changes, local fluctuation patterns, and abrupt change times in the sequence are extracted. The temporal evolution feature vector of the monitoring point is obtained from the encoder output. The temporal evolution feature vector refers to the feature extracted from the monitoring data sequence of a single monitoring point within the window. A dense vector representing the temporal dynamic behavior using fixed-dimensional compression is processed one by one at each monitoring point using the same encoder. The resulting temporal evolution feature vectors are stacked and concatenated along the row direction according to the index order of the monitoring points, forming a temporal evolution feature matrix with the number of monitoring points as the number of rows and the dimension of the temporal evolution feature vector as the number of columns. The temporal evolution feature matrix refers to the multi-point temporal representation of the temporal evolution features of all measurement points on the entire bridge within the same sliding time window, uniformly organized into a matrix form. Then, a cross-attention mechanism module is constructed. This cross-attention mechanism module first performs a linear transformation on the deep spatial correlation features, that is, the deep spatial correlation feature matrix and the first The learnable weight matrices are multiplied to generate a query matrix as a query vector. The query vector refers to the cross-modal interaction initiator carrying deep spatial correlation information. Simultaneously, the temporal evolution feature matrix undergoes two independent linear transformations: multiplying it by the second and third learnable weight matrices respectively to generate a key matrix as a key vector and a value matrix as a value vector. The key vector refers to the index features in the temporal dimension used for matching and retrieval by the query vector, and the value vector refers to the actual information content to be extracted in the temporal dimension. Then, the dot product between the query vector and the transpose of the key vector is calculated, and the result is divided by the dimension of the key vector. The similarity matrix is ​​obtained by scaling the square root. Softmax normalization is then performed on this similarity matrix in the row direction to obtain the cross-dimensional interaction weight matrix. The cross-dimensional interaction weight matrix refers to the distribution of attention of each query position in the spatial dimension to the information of each position in the temporal dimension. Finally, the cross-dimensional interaction weight matrix and the value vector are weighted and aggregated by matrix multiplication. The spatial dimension query vector is weighted and converged with the temporal dimension value vector to obtain the spatiotemporal fusion feature matrix. The spatiotemporal fusion feature matrix refers to the fusion representation matrix with the same size as the query vector, generated after the interaction of deep spatial correlation information and temporal evolution information through cross-dimensional attention.Based on the basic structure of the cross-attention mechanism module, a damage propagation chain position encoding vector is introduced. This vector, generated by a sine / cosine position encoding function or a learnable position embedding layer, is based on the relative order of each monitoring point on the bridge within a predefined damage propagation chain sequence. Its dimension matches the feature dimensions of the query and key matrices. Specifically, for each damage propagation chain, nodes are sequentially numbered 0, 1, 2, ... along the propagation direction from the damage initiation point. A sine function is used to calculate the position encoding value for odd-numbered indices, and a cosine function for even-numbered indices. The damage propagation chain position encoding vector is then added element-wise to the query matrix generated by linear transformation, and simultaneously added element-wise to the key matrix generated by linear transformation. This injects both the spatial query vector and the temporal key vector with prior information about the relative positions on the damage propagation chain. Finally, the scaled dot product attention calculation is performed based on the added query and key matrices, thus forming a spatiotemporal cross-attention fusion layer for damage propagation chain perception.

[0044] It should be noted that the spatiotemporal cross-attention fusion layer for damage propagation chain perception in this application refers to an attention fusion structure that incorporates positional information of the damage propagation chain during the cross-dimensional interaction of spatial and temporal features. This makes the allocation of attention weights not only dependent on the semantic similarity between spatial and temporal features, but also constrained by the sequential relationship of each measurement point in the damage propagation chain, thereby enhancing the ability to perceive the spatiotemporal coupling pattern of damage propagation along the propagation path.

[0045] The spatiotemporal cross-attention fusion layer uses the deep spatial correlation features as the query vector and the temporal evolution features of segmented temporal data within the sliding time window as the key and value vectors. It performs cross-dimensional feature interaction to obtain the spatiotemporal fusion feature vector using the following steps: A linear transformation is performed on the deep spatial correlation features to generate a query matrix as a query vector. Linear transformations are performed on the temporal evolution characteristics of the segmented time series data to generate a key matrix as a key vector and a value matrix as a value vector. Calculate the dot product similarity matrix between the query vector and the key vector, and construct a spatiotemporal cross-attention weight matrix by combining the damage propagation chain position encoding vector with the dot product similarity matrix; The spatiotemporal cross-attention weight matrix and the value vector are weighted and summed to obtain the spatiotemporal fusion feature vector.

[0046] In specific implementation, firstly, a linear transformation is performed on the deep spatial correlation features, that is, the deep spatial correlation feature matrix is ​​transformed into a matrix of dimension D×. The first learnable weight matrix Multiply the matrix and project it from the D-dimensional feature space to the d_k-dimensional query subspace to generate a query matrix as a query vector. This query vector is used as the vector representation for initiating cross-modal information retrieval in the time-series dimension. Next, perform two independent linear transformations on the temporal evolution features of the segmented time-series data, that is, multiply the temporal evolution feature matrix with the matrix of dimension D_k. × The second learnable weight matrix And the dimension is D× The third learnable weight matrix Multiplication, the former transforms the temporal evolution characteristics from the original Dimension mapping to The 3D key subspace uses the generated key matrix as the key vector, which in turn transforms the temporal evolution features from... Dimension mapping to The value subspace uses generated value matrices as value vectors. The key vectors refer to the feature indexes in the temporal dimension used for retrieval of query vectors through similarity matching. The value vectors refer to the temporal information content in the temporal dimension to be extracted based on matching weights. Then, the dot product similarity matrix between the query vector and the key vector is calculated, i.e., matrix multiplication is performed on the transpose of the query matrix and the key matrix, and the product result is divided by the key vector dimension. The similarity matrix is ​​obtained by scaling the square root of the matrix to stabilize the gradient, resulting in a scaled dot product similarity matrix. This matrix is ​​a two-dimensional matrix that measures the feature similarity between each spatial query location and each temporal key location. Subsequently, a damage propagation chain position encoding vector is introduced. This vector is generated using a sine-cosine position encoding function based on the relative order of each monitoring point in a predefined damage propagation chain sequence. This vector is then element-wise added to the scaled dot product similarity matrix, incorporating the damage propagation chain position along the damage path into the similarity matrix. The order of damage propagation direction information is used to normalize the summed matrix along the row direction using the Softmax function, so that the sum of each row element is 1, resulting in a spatiotemporal cross-attention weight matrix. The spatiotemporal cross-attention weight matrix refers to the normalized attention distribution of each query position in the spatial dimension to each key position in the temporal dimension under the constraint of the damage propagation chain order. Finally, the spatiotemporal cross-attention weight matrix and the value vector are multiplied by matrix and weighted, that is, each row vector in the value matrix is ​​weighted and accumulated according to the corresponding spatiotemporal cross-attention weight coefficient to obtain the spatiotemporal fusion feature vector.

[0047] It should be noted that, in this application, the spatiotemporal fusion feature vector refers to the fusion representation vector generated by the cross-dimensional attention interaction between the deep spatial correlation query information of the spatial dimension and the temporal evolution features of the temporal dimension through the damage propagation chain position encoding constraint. This spatiotemporal fusion feature vector serves as the direct input feature for the subsequent optimized classification and discrimination layer, and can provide a unified discrimination basis for the identification of the damage status of each monitoring point of the bridge and the inference of the damage propagation chain sequence, which simultaneously encodes the spatial damage propagation path directional constraint and the temporal evolution dynamic coupling information.

[0048] In step S106, the spatiotemporal fusion feature vector is input into the optimized classification and discrimination layer for data diagnosis, and the diagnostic results containing the damage propagation chain sequence are output.

[0049] The spatiotemporal fusion feature vector is input into the optimized classification and discrimination layer for data diagnosis, and the diagnostic results containing the damage propagation chain sequence are output using the following steps: The spatiotemporal fusion feature vector is input into an optimized classification and discrimination layer consisting of a fully connected layer and a Softmax activation function to calculate the posterior probability of each monitoring point belonging to each damage category. The posterior probability of each monitoring point is classified point by point to determine the damage status label of each monitoring point, and the damage spatial distribution within the sliding time window is generated based on the damage status label of each monitoring point. The damage spatial distribution is aligned and mapped with the damage propagation chain position encoding vector. The damage status labels within the sliding time window are concatenated along the propagation direction indicated by the damage propagation chain position encoding vector using a damage propagation path tracing algorithm to form a damage propagation chain sequence as the diagnostic result output.

[0050] In specific implementation, firstly, the spatiotemporal fusion feature vector is input into an optimized classification and discriminant layer composed of a fully connected layer and a softmax activation function. The fully connected layer is a linear classification layer that multiplies the spatiotemporal fusion feature vector with a learnable weight matrix, superimposes a bias vector, and then maps it through a non-linear activation function. Essentially, it linearly transforms the dimension of the spatiotemporal fusion feature vector from the original feature space to the dimension of the number of damage categories. For example, if the number of damage categories is set to four (no damage, minor damage, moderate damage, and severe damage), the fully connected layer outputs a raw classification score vector of length four. This raw classification score vector refers to the unnormalized discriminant function value for each damage category. Then, the raw classification score is mapped to... The input is a Softmax activation function. For each monitoring point, the score vector is calculated by taking the exponent of each element with the natural constant as the base and dividing by the sum of the exponents of each element. The posterior probability of each monitoring point belonging to each damage category is calculated. The posterior probability is the probability estimate vector of the model for the current monitoring point belonging to each damage state. The sum of all its elements is 1 and the value of each element is in the range of 0 to 1. Next, a point-by-point category determination is performed on the posterior probability of each monitoring point. The maximum posterior probability criterion is used to select the damage category corresponding to the element with the largest value in the posterior probability vector as the damage state label of the monitoring point. For example, the posterior probability vector of a certain support measuring point is [0.05, 0.12, 0.73, 0.12 ...

[10] Then, moderate damage is taken as the damage status label of the measuring point. After traversing all monitoring points and completing point-by-point discrimination, the damage status labels of all monitoring points are organized into the damage spatial distribution within the sliding time window according to their corresponding spatial coordinate positions. The damage spatial distribution refers to the discrete state mapping with the monitoring point number or spatial coordinate as the index and the damage status label of each point as the value, which intuitively indicates the spatial location of damage on each component of the bridge within the window. Finally, the damage spatial distribution is aligned and mapped with the damage propagation chain position encoding vector. The alignment and mapping operation is to rearrange the damage status labels of each monitoring point in the damage spatial distribution according to the chain number order indicated by the damage propagation chain position encoding vector, so that the spatially dispersed damage labels are organized into a label sequence arranged according to the propagation direction. The damage propagation path tracking algorithm is used to track the sliding time window. Within a time window, damage status labels are concatenated along the propagation direction indicated by the damage propagation chain position encoding vector. The damage propagation path tracing algorithm is a graph search-based sequence generation algorithm. Specifically, starting from the initial node of the damage propagation chain, nodes on the chain are visited sequentially along the propagation direction specified by the damage propagation chain position encoding vector. If the damage status label of the currently visited node is not of the "no damage" category, it is retained in the sequence. If a node's label is "no damage" but its subsequent nodes still have damage labels, it is marked as a propagation breakpoint, and the search continues. Continuously appearing "no damage" labels are concatenated along the propagation direction to form an ordered label chain. For example, for the damage propagation chain of a three-span continuous beam bridge from support to main beam to pier, the damage propagation chain sequence "minor damage to support → moderate damage to main beam → minor damage to pier" is obtained, and this sequence is output as the diagnostic result.

[0051] It should be noted that the damage propagation chain sequence in this application refers to an ordered list of labels describing the cascading state of damage propagation from the initiation point to subsequent components along the structural force transmission path. This sequence, together with the damage status labels and spatial distribution of each monitoring point, is output as a diagnostic result, providing complete structured diagnostic information including damage location, damage degree, and damage propagation path for bridge operation and maintenance decisions.

[0052] In addition, this application also provides a computer device, the computer device including a memory and a processor, the memory storing code, the processor being configured to acquire the code and execute the above-described bridge monitoring data diagnosis method fused with improved graph neural networks.

[0053] In some embodiments, reference Figure 3 The figure is a schematic diagram of the structure of a computer device implementing a bridge monitoring data diagnosis method based on fused improved graph neural networks, according to some embodiments of this application. The bridge monitoring data diagnosis method based on fused improved graph neural networks in the above embodiments can... Figure 3The computer device shown is used to implement this, and the computer device includes at least one processor 301, a communication bus 302, a memory 303, and at least one communication interface 304.

[0054] The processor 301 may be a general-purpose central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more devices used to control the execution of the bridge monitoring data diagnostic method that incorporates the improved graph neural network in this application.

[0055] The communication bus 302 can be used to transmit information between the aforementioned components.

[0056] The memory 303 may be a read-only memory (ROM) or other type of static storage device capable of storing static information and instructions, random access memory (RAM) or other type of dynamic storage device capable of storing information and instructions, or electrically erasable programmable read-only memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital universal optical discs, Blu-ray discs, etc.), magnetic disks or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but not limited thereto. The memory 303 may exist independently and be connected to the processor 301 via the communication bus 302. The memory 303 may also be integrated with the processor 301.

[0057] The memory 303 stores program code for executing the scheme of this application, and its execution is controlled by the processor 301. The processor 301 executes the program code stored in the memory 303. The program code may include one or more software modules. In the above embodiments, the determination of the bridge monitoring data diagnosis method integrating improved graph neural networks can be implemented by the processor 301 and one or more software modules in the program code in the memory 303.

[0058] Communication interface 304 uses any transceiver-like device for communicating with other devices or communication networks, such as Ethernet, radio access network (RAN), wireless local area networks (WLAN), etc.

[0059] In a specific implementation, as one example, a computer device may include multiple processors, each of which may be a single-core (single-CPU) processor or a multi-core (multi-CPU) processor. Here, a processor may refer to one or more devices, circuits, and / or processing cores used to process data (e.g., computer program instructions).

[0060] The aforementioned computer device can be a general-purpose computer device or a special-purpose computer device. In specific implementations, the computer device can be a desktop computer, a portable computer, a network server, a handheld digital assistant (PDA), a mobile phone, a tablet computer, a wireless terminal device, a communication device, or an embedded device. This application does not limit the type of computer device.

[0061] In addition, this application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described bridge monitoring data diagnosis method incorporating an improved graph neural network.

[0062] Although preferred embodiments of this application have been described, those skilled in the art, once they have learned the basic inventive concept, can make other changes and modifications to these embodiments.

[0063] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application.

Claims

1. A bridge monitoring data diagnosis method integrating an improved graph neural network, characterized in that, Includes the following steps: Collect multi-source time-series monitoring data of each monitoring point on the bridge within a preset time period, and divide the multi-source time-series monitoring data into segmented time-series data corresponding to each sliding time window; For segmented time series data within the sliding time window, the dynamic correlation coefficient between monitoring data at each monitoring point is calculated, and a dynamic weighted graph topology is constructed by combining the spatial connection relationship between various bridge components. The multi-source time series monitoring data within the sliding time window is mapped to the node characteristics of each graph node in the dynamic weighted graph topology. An improved graph neural network for damage propagation path constraints is constructed. The improved graph neural network for damage propagation path constraints consists of multiple layers of graph convolutional units cascaded in sequence. Each graph convolutional unit includes a graph convolutional layer and an attention weight adjustment layer for damage propagation path constraints. The dynamic weighted graph topology is iteratively updated by the multi-layer graph convolutional units of the improved graph neural network, and the deep spatial correlation features are output after the iterative update by the multi-layer graph convolutional units. A spatiotemporal cross-attention fusion layer for damage propagation chain perception is constructed. The spatiotemporal cross-attention fusion layer uses the deep spatial correlation features as query vectors and the temporal evolution features of segmented temporal data within a sliding time window as key vectors and value vectors, and performs cross-dimensional feature interaction to obtain spatiotemporal fusion feature vectors. The spatiotemporal fusion feature vector is input into the optimized classification and discrimination layer for data diagnosis, and the output includes the diagnosis result containing the damage propagation chain sequence.

2. The method as described in claim 1, characterized in that, The division of the multi-source time-series monitoring data into segmented time-series data corresponding to each sliding time window specifically includes: The multi-source time-series monitoring data is extracted at a preset time step to obtain the synchronous monitoring vector of each monitoring point at a single time step; Arrange the synchronous monitoring vectors according to the order of their acquisition times to form a multi-source monitoring time sequence matrix with the number of sampling points as its length. A windowed multi-source time series submatrix is ​​obtained by continuously sliding and truncating along the time dimension of the multi-source monitoring time series matrix using a sliding window with a preset window length and sliding step size. The synchronized monitoring vector sequence contained in each of the windowed multi-source time series sub-matrices constitutes the segmented time series data corresponding to each sliding time window.

3. The method as described in claim 1, characterized in that, For segmented time-series data within a sliding time window, the dynamic correlation coefficient between monitoring data at each monitoring point is calculated, and a dynamic weighted graph topology is constructed by combining the spatial connection relationships between various bridge components. Specifically, this includes: Extract the monitoring data sequences of any two monitoring points within the sliding time window from the segmented time series data, and calculate the Pearson correlation coefficient between the monitoring data sequences of the two monitoring points. The Pearson correlation coefficient is used as the dynamic correlation coefficient between the two monitoring points within the sliding time window. The dynamic correlation matrix is ​​constructed by iterating through the dynamic correlation coefficients between all pairs of monitoring points. Obtain the spatial connection relationship adjacency matrix between each component of the bridge, and multiply the spatial connection relationship adjacency matrix element by element with the dynamic correlation matrix to obtain the dynamic weighted adjacency matrix; Using each monitoring point as a graph node and the aforementioned dynamic weighted adjacency matrix as edge weights, a dynamic weighted graph topology is constructed.

4. The method as described in claim 1, characterized in that, Mapping multi-source time-series monitoring data within a sliding time window to the node features of each node in the dynamically weighted graph topology specifically includes: Temporal features are extracted from the monitoring data sequences of each monitoring point in the segmented time series data within the sliding time window to obtain the local temporal feature vector of each monitoring point. Using the monitoring points corresponding to each graph node in the dynamic weighted graph topology as indexes, the local time-series feature vectors are assigned to the corresponding graph nodes to form the node features of each graph node.

5. The method as described in claim 1, characterized in that, An improved graph neural network for damage propagation path constraints is constructed, which consists of multiple cascaded graph convolutional units. Each graph convolutional unit includes a graph convolutional layer and an attention weight adjustment layer for damage propagation path constraints. Specifically, it includes: A graph convolutional layer is constructed, which takes the adjacency matrix and node feature matrix of the dynamic weighted graph topology as input, and outputs the node embedding feature matrix through graph convolution operation; A damage propagation path constraint attention weight adjustment layer is cascaded after the graph convolutional layer. The damage propagation path constraint attention weight adjustment layer takes the node embedding feature matrix and the damage propagation path constraint matrix between each component of the bridge as input, calculates the constraint attention weight between each adjacent node through the path constraint attention mechanism, and performs weighted aggregation on the node embedding feature matrix according to the constraint attention weight, and outputs a node feature matrix with enhanced path constraints. The graph convolutional layer and the attention weight adjustment layer for the damage propagation path constraint are combined into a graph convolutional unit. Multiple graph convolutional units are cascaded in sequence, and the node feature matrix of the path constraint enhancement output by the previous graph convolutional unit is used as the input of the node feature matrix of the graph convolutional layer in the next graph convolutional unit, thus forming an improved graph neural network for damage propagation path constraint.

6. The method as described in claim 1, characterized in that, The improved graph neural network iteratively updates the dynamic weighted graph topology using multi-layer graph convolutional units. The deep spatial correlation features output after this iterative update by the multi-layer graph convolutional units specifically include: The adjacency matrix of the dynamically weighted graph topology is used as the input of the graph convolutional layer of the first graph convolutional unit of the improved graph neural network, and the node feature matrix is ​​used as the input of the graph convolutional layer of the first graph convolutional unit. The first graph convolution unit performs graph convolution operations on the adjacency matrix and the node feature matrix, and adjusts the attention weights of the damage propagation path constraints, outputting the node feature matrix with enhanced path constraints and the attention weight matrix of the damage propagation path constraints corresponding to the first graph convolution unit. The attention weight matrix output by the first graph convolutional unit is multiplied element-wise with the adjacency matrix of the dynamically weighted graph topology to obtain the iteratively updated adjacency matrix. The iteratively updated adjacency matrix and the path-constrained node feature matrix output by the first graph convolutional unit are used as the graph convolutional layer input of the next graph convolutional unit. The iterative update of multiple graph convolutional units is performed sequentially, and the path-constrained node feature matrix output by the last graph convolutional unit is used as the deep spatial association feature.

7. The method as described in claim 1, characterized in that, The spatiotemporal cross-attention fusion layer uses the deep spatial correlation features as the query vector and the temporal evolution features of segmented temporal data within the sliding time window as the key and value vectors, and performs cross-dimensional feature interaction to obtain the spatiotemporal fusion feature vector, specifically including: A linear transformation is performed on the deep spatial correlation features to generate a query matrix as a query vector. Linear transformations are performed on the temporal evolution characteristics of the segmented time series data to generate a key matrix as a key vector and a value matrix as a value vector. Calculate the dot product similarity matrix between the query vector and the key vector, and construct a spatiotemporal cross-attention weight matrix by combining the damage propagation chain position encoding vector with the dot product similarity matrix; The spatiotemporal cross-attention weight matrix and the value vector are weighted and summed to obtain the spatiotemporal fusion feature vector.

8. A computer device, characterized in that, The computer device includes a memory and a processor, the memory storing code, and the processor being configured to retrieve the code and execute the bridge monitoring data diagnostic method fused with an improved graph neural network as described in any one of claims 1 to 7.

9. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the bridge monitoring data diagnosis method fused with an improved graph neural network as described in any one of claims 1 to 7.

Images