Structure monitoring data anomaly ai diagnosis method and system

Abstract

The present application relates to the technical field of structure monitoring, and in particular to a structure monitoring data anomaly AI diagnosis method and system. A bidirectional correlation feature matrix is constructed by acquiring sensor node stress data sets and performing spatial topology correlation analysis and time series causal correlation analysis, an unlabeled training output anomaly probability distribution map and classification label are output based on a contrast learning algorithm, a causal explanation atlas is generated by reverse tracing the associated path of the sensor node corresponding to the abnormal area, the confidence is calculated by matching the abnormal source node with the historical maintenance record, and finally a maintenance decision scheme is generated by positioning the propagation link according to the abnormal source node and combining the confidence ranking. The method improves the accuracy and interpretability of anomaly diagnosis and optimizes the maintenance efficiency.

Description

Technical Field

[0001] This invention relates to the field of structural monitoring technology, and in particular to an AI-based diagnostic method and system for structural monitoring data anomalies. Background Technology

[0002] In the field of structural health monitoring, current conventional practices typically rely on threshold discrimination or time-series statistical modeling of single sensors for anomaly detection. Specifically, technicians pre-set safety thresholds for physical quantities such as stress and strain; when monitored data exceeds these thresholds, an alarm is triggered. Another approach uses time-series prediction models to predict single-node data and identifies anomalies through residual analysis. Some solutions introduce supervised learning methods, using labeled fault data to train a classifier, directly mapping monitoring data to normal or abnormal labels. These methods have been widely used in engineering, but all have significant limitations.

[0003] Conventional approaches fail to fully utilize the spatial topology information of structured sensor networks. The mechanical transmission relationships and damage propagation paths between adjacent sensor nodes are completely ignored, meaning that threshold exceedances of isolated nodes may be caused by anomalous coupling with surrounding nodes, making it impossible to distinguish between local faults and system-level responses. Furthermore, time-series analysis of a single sensor struggles to capture the causal relationships between multiple nodes, and the spatial diffusion patterns and temporal delay effects of anomalous signals are treated separately, resulting in high false alarm and false negative rates.

[0004] Another significant drawback lies in the lack of causal explanation for abnormal results. Current technologies only output abnormal alarms or probability values, failing to trace the root cause of the anomaly or reveal the propagation path and direction of the fault within the sensor network. Faced with a large volume of alarm information, maintenance personnel struggle to determine priorities and fault sources, leading to time-consuming and labor-intensive on-site troubleshooting, and potentially overlooking hidden, progressive damage. Furthermore, the lack of a correlation mapping mechanism between historical maintenance records and real-time monitoring data means that fault diagnosis results cannot be verified and calibrated using historical experience, reducing the reliability of the diagnostic results and their practical engineering value. Summary of the Invention

[0005] This invention provides an AI-based diagnostic method and system for abnormal structural monitoring data, which can solve the problems in the prior art.

[0006] A first aspect of the present invention provides an AI-based diagnostic method for anomalies in structural monitoring data, comprising:

[0007] Acquire the stress time series data of sensor nodes on the structure and output a multi-node stress dataset;

[0008] Spatial topological correlation analysis and temporal causal correlation analysis were performed on the multi-node stress dataset to construct a bidirectional correlation feature matrix that integrates spatial correlation and temporal dependence among sensor nodes.

[0009] Based on the bidirectional correlation feature matrix, a contrastive learning algorithm is used for unlabeled training to learn the feature boundaries of normal and abnormal patterns, and output an abnormal probability distribution map and abnormal pattern classification labels.

[0010] Based on the anomaly probability distribution map, the abnormal region is identified. The sensor node association path corresponding to the abnormal region is traced back along the temporal and spatial dimensions in the bidirectional association feature matrix. The propagation nodes on the association path are marked by the anomaly pattern classification label to generate a causal explanation map containing the anomaly source node and the propagation link.

[0011] Obtain historical maintenance records, match the anomaly source nodes in the causal explanation graph with the fault locations in the historical maintenance records, calculate the matching accuracy as the confidence level of the causal explanation graph, and output the verified causal explanation graph.

[0012] Based on the anomaly source nodes in the validated causal explanation graph, the propagation path is located, and the maintenance priority is determined by combining the confidence ranking, generating a decision scheme that includes maintenance location and execution order.

[0013] Spatial topological correlation analysis and temporal causal correlation analysis were performed on the multi-node stress dataset to construct a bidirectional correlation feature matrix that integrates the spatial correlation and temporal dependence among sensor nodes, including:

[0014] A spatial adjacency matrix is ​​constructed based on the geometric position and structural connection relationship of distributed sensor nodes. The peak value of the cross-correlation function of stress data between sensor node pairs is calculated as the spatial correlation coefficient, and a spatial correlation coefficient matrix is ​​constructed.

[0015] A threshold is set on the spatial correlation coefficient matrix to filter sensor node pairs with correlation coefficients greater than the threshold. A sliding time window is constructed to segment the multi-node stress dataset. The Granger causality test is used to calculate the temporal causal intensity value of the filtered sensor node pairs in each time window, and a temporal dependency matrix is ​​constructed.

[0016] The spatial correlation coefficient matrix is ​​arranged according to the spatial location of the sensor nodes to form a spatial correlation submatrix, and the temporal dependence matrix is ​​arranged according to the time order of the sliding time window to form a temporal correlation submatrix. The spatial correlation submatrix and the temporal correlation submatrix are concatenated column by column to form a bidirectional correlation feature matrix.

[0017] Based on the bidirectional correlation feature matrix, a contrastive learning algorithm is used for unlabeled training to learn the feature boundaries between normal and abnormal patterns. The output includes an anomaly probability distribution map and anomaly pattern classification labels:

[0018] Feature blocks are obtained by sliding window sampling of the bidirectional correlation feature matrix. Enhanced feature blocks are generated by time flip transformation and spatial rotation transformation of the feature blocks. Positive sample pairs are constructed by pairing the feature blocks with the enhanced feature blocks, and negative sample pairs are constructed by randomly pairing different feature blocks.

[0019] Positive and negative sample pairs are input into the encoder network to extract feature vectors. A contrastive loss function is constructed to maximize the similarity of feature vectors of positive sample pairs and minimize the similarity of feature vectors of negative sample pairs. The encoder network is trained iteratively. During the iterative training, the feature vectors are clustered to extract normal pattern cluster centers. When the displacement distance of the normal pattern cluster center between adjacent iterations is continuously lower than the preset convergence threshold, the normal pattern cluster center is solidified.

[0020] The bidirectional correlation feature matrix is ​​input into the trained encoder network to extract the sensor node feature vectors. The distance between the sensor node feature vectors and the fixed normal pattern cluster centers is calculated and normalized to obtain the anomaly probability value. The anomaly probability distribution map is generated by mapping the anomaly probability value according to the spatial location.

[0021] The neighborhood density is calculated for the feature vector of the sensor node. The peak point of the neighborhood density is searched as the center of the pattern cluster. The sensor nodes are assigned to the nearest pattern cluster center to form a pattern cluster. The abnormal pattern classification label is assigned according to the statistical features of the stress time series data of the sensor nodes in the pattern cluster.

[0022] Anomaly regions are identified based on the anomaly probability distribution map. The sensor node association paths corresponding to the anomaly regions are traced backward along the temporal and spatial dimensions in the bidirectional correlation feature matrix. Propagation nodes along the association paths are labeled with anomaly pattern classification tags, generating a causal explanation map containing the anomaly source node and its propagation links.

[0023] Set an anomaly threshold on the anomaly probability distribution map to extract sensor nodes with anomaly probability values ​​greater than the anomaly threshold and aggregate them to identify abnormal regions.

[0024] From the sensor nodes in the abnormal region, trace back along the spatial dimension of the spatial correlation submatrix of the bidirectional correlation feature matrix to find the spatial correlation path with increasing spatial correlation coefficient, and trace back along the temporal dimension of the temporal correlation submatrix to find the temporal correlation path with increasing temporal causal strength value. Merge the spatial correlation path and the temporal correlation path to form a sensor node correlation path network.

[0025] Traverse the sensor node association path network to calculate the gradient change rate of the anomaly probability value of the sensor node on each association path, and take the sensor node corresponding to the maximum gradient change rate as the anomaly source node.

[0026] The propagation path is extracted by forward traversal of the path associated with the anomaly source node in the sensor node network. The sensor nodes on the propagation path are labeled as propagation nodes by combining the anomaly pattern classification label, and a causal explanation graph containing the anomaly source node and the propagation path is generated.

[0027] Obtain historical maintenance records, match the anomaly source nodes in the causal explanation graph with the fault locations in the historical maintenance records, calculate the matching accuracy as the confidence level of the causal explanation graph, and output the validated causal explanation graph, including:

[0028] Retrieve historical maintenance records containing maintenance timestamps, fault location coordinates, and maintenance type labels;

[0029] Extract the timestamps and spatial coordinates of the anomaly source nodes in the causal explanation graph, retrieve the fault locations within the time window from the historical maintenance records as discrete fault locations, input the discrete fault locations into the bidirectional correlation feature matrix to calculate the bidirectional propagation probability between discrete fault locations, and connect the discrete fault locations according to the bidirectional propagation probability to construct the historical fault propagation link.

[0030] Extract the propagation links corresponding to the anomaly source nodes in the causal explanation graph, match the node sequences of the propagation links with the node sequences of historical fault propagation links using the longest common subsequence, and calculate the ratio of the length of the longest common subsequence to the length of the node sequence as the link matching degree between the anomaly source node and the fault location.

[0031] Extract the abnormal pattern classification label sequence of nodes in the propagation link and the maintenance type label sequence of nodes in the historical fault propagation link, and calculate the edit distance of the label sequence to correct the link matching degree.

[0032] The ratio of the number of abnormal source nodes corresponding to the corrected link matching degree to the total number of abnormal source nodes is used as the matching accuracy. The matching accuracy is used as the confidence level of the causal explanation graph. The confidence level of abnormal source nodes is marked in the causal explanation graph, and the verified causal explanation graph is output.

[0033] Based on the verified causal explanation graph, the propagation path of the anomaly source nodes is located. Maintenance priorities are determined by combining confidence ranking, and a decision scheme including maintenance location and execution order is generated, including:

[0034] Extract the propagation path and confidence level label corresponding to the anomaly source node from the verified causal explanation graph, and obtain the anomaly probability value sequence, anomaly pattern classification label sequence, spatial coordinate sequence and timestamp sequence of the propagation node on the propagation path;

[0035] The abnormal probability value sequence is divided into uncontrolled propagation segments by identifying the propagation acceleration inflection point through second-order difference. The frequency of abnormal pattern classification label transfer within the uncontrolled propagation segment is statistically analyzed to construct a transfer probability matrix. The maximum eigenvalue of the transfer probability matrix is ​​extracted as the pattern evolution rate. The dynamic risk index is obtained by exponentially operating the confidence label with the pattern evolution rate as the exponent.

[0036] The mutual information of the joint probability distribution of the spatial coordinate sequence and the timestamp sequence of the propagation link is calculated as the spatiotemporal coupling strength. Based on the spatiotemporal coupling strength, competitive groups and non-competitive groups are divided. For the competitive group, the maintenance time interval is converted according to the dynamic risk index difference to determine the maintenance priority. For the non-competitive group, the maintenance decision space is constructed by the dynamic risk index and the spatial expansion rate, and the maintenance priority is determined by Euclidean distance clustering and two-level sorting.

[0037] The maintenance priorities of the competing and non-competitive groups are merged to generate a complete maintenance priority sequence. The spatial coordinates of the anomaly source nodes are extracted as maintenance locations, and their arrangement order is used as the execution order. A decision scheme containing maintenance locations and execution order is then generated.

[0038] The abnormal probability value sequence is divided into out-of-control propagation segments by identifying propagation acceleration inflection points using second-order difference analysis. A transition probability matrix is ​​constructed by statistically analyzing the frequency of abnormal pattern classification label transitions within the out-of-control propagation segments. The largest eigenvalue of the transition probability matrix is ​​extracted as the pattern evolution rate. A dynamic risk index is obtained by exponentially multiplying the confidence level labels by the pattern evolution rate, including:

[0039] The propagation acceleration sequence is obtained by performing second-order difference on the abnormal probability value sequence. The position where the propagation acceleration changes from negative to zero is identified as the propagation acceleration inflection point. The propagation nodes before the propagation acceleration inflection point are divided into the runaway propagation segment.

[0040] Extract the classification labels of abnormal patterns within the out-of-control propagation segment and construct an abnormal pattern label chain in chronological order. Calculate the Shannon entropy of the abnormal pattern label chain as the pattern disorder degree. Use the pattern disorder degree as a weight to weight the abnormal probability values ​​of the propagation nodes within the out-of-control propagation segment to obtain a weighted abnormal probability value sequence.

[0041] A transition frequency matrix is ​​constructed by counting the transition frequencies of adjacent label pairs in the statistical abnormal pattern label chain. The transition frequency matrix is ​​then normalized by row to obtain the transition probability matrix. Eigenvalue decomposition is performed on the transition probability matrix to extract the eigenvalue with the largest modulus as the pattern evolution rate.

[0042] The variance of the weighted anomaly probability value sequence is calculated as the propagation fluctuation intensity. The model evolution rate and the propagation fluctuation intensity are weighted and fused together as the power exponent. The confidence level label is used as the base to perform power operation to obtain the dynamic risk index.

[0043] A second aspect of the present invention provides an AI diagnostic system for abnormal structure monitoring data, comprising:

[0044] The stress acquisition unit is used to acquire the stress time series data of the sensor nodes on the structure and output a multi-node stress dataset.

[0045] The correlation matrix unit is used to perform spatial topological correlation analysis and temporal causal correlation analysis on multi-node stress datasets, and to construct a bidirectional correlation feature matrix that integrates spatial correlation and temporal dependence among sensor nodes.

[0046] The anomaly diagnosis unit is used to perform unlabeled training based on the bidirectional correlation feature matrix using a contrastive learning algorithm, learn the feature boundaries between normal and abnormal patterns, and output an anomaly probability distribution map and anomaly pattern classification label.

[0047] The causal graph unit is used to identify abnormal regions based on the anomaly probability distribution map, trace the sensor node association path corresponding to the abnormal region in the bidirectional association feature matrix along the temporal and spatial dimensions, and combine the anomaly pattern classification label to mark the propagation nodes on the association path to generate a causal explanation graph containing the anomaly source node and the propagation link.

[0048] The verification and evaluation unit is used to obtain historical maintenance records, match the anomaly source nodes in the causal explanation graph with the fault locations in the historical maintenance records, calculate the matching accuracy as the confidence level of the causal explanation graph, and output the verified causal explanation graph.

[0049] The decision planning unit is used to locate the propagation link based on the abnormal source node in the verified causal explanation graph, determine the maintenance priority by combining the confidence ranking, and generate a decision scheme that includes the maintenance location and execution order.

[0050] A third aspect of the present invention provides an electronic device, comprising:

[0051] processor;

[0052] Memory used to store processor-executable instructions;

[0053] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0054] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0055] This invention enables accurate diagnosis of structural stress anomalies without requiring extensive labeled data, significantly reducing manual labeling costs. The anomaly probability distribution map visually presents the spatial distribution and severity of anomaly areas, while anomaly pattern classification labels precisely distinguish fault types. The fusion analysis of spatial topological correlation and temporal causal correlation fully uncovers hidden dependencies between data. The bidirectional correlation feature matrix simultaneously captures the dual correlations of sensor nodes in both geometric location and time series, effectively improving diagnostic accuracy and robustness. The causal interpretation map comprehensively presents the anomaly source nodes and their propagation links. This map clearly marks the source location and diffusion path of the anomaly signal, enabling maintenance personnel to quickly understand the causal relationship of the anomaly and avoid blind troubleshooting. The correspondence between anomaly pattern classification labels and propagation nodes further reveals the typical propagation characteristics of different anomaly types in space and time. Matching calculations between historical maintenance records and the causal interpretation map provide quantifiable confidence in the diagnostic results. Matching accuracy reflects the reliability of anomaly source location; the validated causal interpretation map eliminates the risk of false alarms and missed alarms. Based on high-confidence anomaly source nodes, the system automatically calculates the maintenance priority of each propagation link and generates a decision scheme that includes specific maintenance locations and execution order. This scheme prioritizes the allocation of limited maintenance resources to the most critical anomaly sources, significantly improving maintenance efficiency and reducing the long-term safety risks to the structure. Attached Figure Description

[0056] Figure 1 A flowchart illustrating the AI ​​diagnostic method for structural monitoring data anomalies;

[0057] Figure 2 This is a flowchart for dynamic risk assessment and maintenance decision-making based on causal explanation graphs for abnormal propagation. Detailed Implementation

[0058] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0059] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0060] Figure 1 This is a flowchart illustrating the AI ​​diagnostic method for structural monitoring data anomalies according to an embodiment of the present invention.

[0061] AI-based diagnostic methods for structural monitoring data anomalies include:

[0062] Acquire the stress time series data of sensor nodes on the structure and output a multi-node stress dataset;

[0063] Spatial topological correlation analysis and temporal causal correlation analysis were performed on the multi-node stress dataset to construct a bidirectional correlation feature matrix that integrates spatial correlation and temporal dependence among sensor nodes.

[0064] Based on the bidirectional correlation feature matrix, a contrastive learning algorithm is used for unlabeled training to learn the feature boundaries of normal and abnormal patterns, and output an abnormal probability distribution map and abnormal pattern classification labels.

[0065] Based on the anomaly probability distribution map, the abnormal region is identified. The sensor node association path corresponding to the abnormal region is traced back along the temporal and spatial dimensions in the bidirectional association feature matrix. The propagation nodes on the association path are marked by the anomaly pattern classification label to generate a causal explanation map containing the anomaly source node and the propagation link.

[0066] Obtain historical maintenance records, match the anomaly source nodes in the causal explanation graph with the fault locations in the historical maintenance records, calculate the matching accuracy as the confidence level of the causal explanation graph, and output the verified causal explanation graph.

[0067] Based on the anomaly source nodes in the validated causal explanation graph, the propagation path is located, and the maintenance priority is determined by combining the confidence ranking, generating a decision scheme that includes maintenance location and execution order.

[0068] Spatial topological correlation analysis and temporal causal correlation analysis were performed on the multi-node stress dataset to construct a bidirectional correlation feature matrix that integrates the spatial correlation and temporal dependence among sensor nodes, including:

[0069] A spatial adjacency matrix is ​​constructed based on the geometric position and structural connection relationship of distributed sensor nodes. The peak value of the cross-correlation function of stress data between sensor node pairs is calculated as the spatial correlation coefficient, and a spatial correlation coefficient matrix is ​​constructed.

[0070] A threshold is set on the spatial correlation coefficient matrix to filter sensor node pairs with correlation coefficients greater than the threshold. A sliding time window is constructed to segment the multi-node stress dataset. The Granger causality test is used to calculate the temporal causal intensity value of the filtered sensor node pairs in each time window, and a temporal dependency matrix is ​​constructed.

[0071] The spatial correlation coefficient matrix is ​​arranged according to the spatial location of the sensor nodes to form a spatial correlation submatrix, and the temporal dependence matrix is ​​arranged according to the time order of the sliding time window to form a temporal correlation submatrix. The spatial correlation submatrix and the temporal correlation submatrix are concatenated column by column to form a bidirectional correlation feature matrix.

[0072] After obtaining the multi-node stress dataset, it is necessary to extract the correlation features between sensor nodes from two dimensions: spatial topology and temporal causality. Finally, the two types of features are fused into a unified bidirectional correlation feature matrix for use by subsequent comparative learning algorithms.

[0073] For spatial topological correlation analysis, a spatial adjacency matrix is ​​first established based on the actual installation coordinates of each sensor node on the structure and their structural connection relationships. In this matrix, if two sensor nodes have a direct structural component connection (e.g., the two ends of the same beam segment, or the upper and lower sections of the same column), the corresponding matrix element is set to 1; otherwise, it is set to 0. This adjacency matrix reflects the direct connectivity at the physical topology level of the structure and serves as a priori constraint for subsequent spatial correlation analysis. After the adjacency matrix is ​​determined, a cross-correlation function is calculated for the stress time series data of each pair of sensor nodes in the multi-node stress dataset. The cross-correlation function measures the linear correlation between two time series signals at different time delays. The peak value of the cross-correlation function across all time delays is taken as the spatial correlation coefficient for that node pair. ,in and These represent the node numbers. The spatial correlation coefficients for all node pairs are also given. Fill in the corresponding positions to form a spatial correlation coefficient matrix. The matrix dimension is , This represents the total number of sensor nodes. The range of values ​​for the elements in the array is within Between the two nodes, the larger the absolute value, the stronger the spatial correlation of the stress response between the two nodes.

[0074] In the spatial correlation coefficient matrix After the construction is complete, set the correlation threshold. Filter the matrix and keep only those that meet the requirements. The node pairs participate in subsequent temporal causal analysis. Threshold The threshold setting needs to be determined comprehensively based on the structure type and sensor deployment density, and is typically between 0.5 and 0.8. A threshold that is too low will introduce a large number of weakly correlated node pairs, leading to computational redundancy, while a threshold that is too high may miss some physically meaningful correlation paths. The screening operation effectively reduces the computational load of subsequent Granger causality tests while ensuring the physical plausibility of the analyzed object.

[0075] A sliding time window is constructed to segment the multi-node stress dataset. The window length of the sliding time window is specified. With step size The window length is determined based on the sampling frequency and the dynamic characteristics of the structural response. It must be sufficient to cover the complete response period of the structure under typical loads, while the step size controls the overlap ratio between adjacent windows to ensure temporal continuity. For each selected node pair within a time window, a Granger causality test is used to evaluate the nodes. Does the stress sequence affect the nodes? The stress sequence has causal predictive power. The Granger causality test compares sequences containing nodes... Autoregressive models based on historical information and those containing only nodes The prediction residuals of the autoregressive model based on its own historical information are used to determine the significance of causality using the F-statistic. The F-statistic obtained from the test is then normalized and used as the time-series causality strength value. ,in Indicates the sequence number of the time window. The temporal causality strength value for all time windows and all filtered node pairs. Organize the data by node pair dimension and time window dimension to construct a time-series dependency matrix. . The rows correspond to different node pairs, the columns correspond to different time windows, and the matrix elements... Reflecting node pairs In the Temporal causality strength within a time window. For those that do not pass the threshold. The filtered node pairs have corresponding rows in Zero values ​​are filled in to maintain the consistency of the matrix dimensions.

[0076] After obtaining the spatial correlation coefficient matrix respectively With time-series dependency matrix Next, the two need to be integrated into a unified bidirectional correlation feature matrix. This involves the spatial correlation coefficient matrix. The sensor nodes are arranged according to their spatial location, that is, the nodes are numbered sequentially according to their physical coordinates on the structure (e.g., from left to right along the main beam, or from bottom to top vertically), so that the arrangement order of the matrix rows and columns is consistent with the spatial location of the nodes, forming a spatial correlation submatrix. This arrangement ensures that nodes in adjacent rows and columns of the matrix are also spatially adjacent, which helps subsequent convolutional feature extraction algorithms capture local spatial structure. (For temporal dependency matrices...) The columns are arranged in chronological order according to the sliding time window, so that the column directions of the matrix correspond to the positive evolution of the time axis, forming a time-series correlated submatrix. The row dimension of the temporal correlation submatrix is ​​consistent with that of the spatial correlation submatrix, both corresponding to node pair indices.

[0077] Spatial correlation submatrix Time-related submatrix Concatenate the matrices along the column direction to obtain the bidirectional correlation feature matrix. .like The number of columns is (Number of corresponding nodes) The number of columns is (corresponding to the number of time windows), then The number of columns is .matrix Each row corresponds to a sensor node pair, the first row... The column encodes the correlation distribution of the node pair at the spatial topology level, and then... The column encoding represents the dynamic changes in the temporal causal intensity of the node pair within different time windows. This column-by-column fusion method preserves the independent semantics of the two types of features while making them complementary within the same matrix framework, providing subsequent contrastive learning algorithms with input representations that simultaneously contain spatial structure information and temporal dynamic information.

[0078] Bidirectional correlation feature matrix Before inputting the data into the contrastive learning algorithm, normalization is required. This normalizes the spatial correlation submatrix and the temporal correlation submatrix to the same numerical range, preventing feature weight imbalance due to differences in units or numerical scales. The normalization uses a min-max normalization method, independently calculating the minimum and maximum values ​​for each element in each submatrix and linearly mapping all elements to... The normalized bidirectional correlation feature matrix can then be used as input for subsequent unlabeled contrastive learning training, supporting the learning process of feature boundaries between normal and abnormal modes.

[0079] Based on the bidirectional correlation feature matrix, a contrastive learning algorithm is used for unlabeled training to learn the feature boundaries between normal and abnormal patterns. The output includes an anomaly probability distribution map and anomaly pattern classification labels:

[0080] Feature blocks are obtained by sliding window sampling of the bidirectional correlation feature matrix. Enhanced feature blocks are generated by time flip transformation and spatial rotation transformation of the feature blocks. Positive sample pairs are constructed by pairing the feature blocks with the enhanced feature blocks, and negative sample pairs are constructed by randomly pairing different feature blocks.

[0081] Positive and negative sample pairs are input into the encoder network to extract feature vectors. A contrastive loss function is constructed to maximize the similarity of feature vectors of positive sample pairs and minimize the similarity of feature vectors of negative sample pairs. The encoder network is trained iteratively. During the iterative training, the feature vectors are clustered to extract normal pattern cluster centers. When the displacement distance of the normal pattern cluster center between adjacent iterations is continuously lower than the preset convergence threshold, the normal pattern cluster center is solidified.

[0082] The bidirectional correlation feature matrix is ​​input into the trained encoder network to extract the sensor node feature vectors. The distance between the sensor node feature vectors and the fixed normal pattern cluster centers is calculated and normalized to obtain the anomaly probability value. The anomaly probability distribution map is generated by mapping the anomaly probability value according to the spatial location.

[0083] The neighborhood density is calculated for the feature vector of the sensor node. The peak point of the neighborhood density is searched as the center of the pattern cluster. The sensor nodes are assigned to the nearest pattern cluster center to form a pattern cluster. The abnormal pattern classification label is assigned according to the statistical features of the stress time series data of the sensor nodes in the pattern cluster.

[0084] When sampling the bidirectional correlation feature matrix using a sliding window, a fixed window size is used to slide along the temporal and spatial dimensions of the matrix, obtaining a local feature block each time. The size of the feature block is determined by comprehensively considering the extension range of temporal correlation and the spatial coverage density of sensor nodes, ensuring that each feature block contains sufficient temporal variation information and covers the spatial correlation structure between adjacent nodes. Two types of data augmentation operations are performed on the sampled feature blocks: temporal flip transformation mirrors the feature blocks along the temporal axis, changing the original early-to-late temporal arrangement to late-to-earth, thus simulating the equivalent expression of temporal patterns in different directions; spatial rotation transformation rotates or permutates the feature blocks in the node spatial dimension, simulating the spatial perspective change caused by the different installation positions of sensor nodes. The augmented feature blocks obtained after the above transformations and the original feature blocks are semantically different expressions of the same structural state. Therefore, the original feature blocks are paired with their corresponding augmented feature blocks to form positive sample pairs, expressing the semantic equivalence relationship of "same state, different perspectives". Feature blocks from different sampling positions or different time periods are randomly paired to form negative sample pairs, expressing the semantic difference relationship of "different states". This method of constructing positive and negative sample pairs does not rely on manual annotation at all, and enables the preparation of contrastive learning data under unlabeled conditions.

[0085] Positive and negative sample pairs are input into the encoder network. The encoder network employs a deep network structure with temporal modeling and spatial awareness capabilities to extract fixed-dimensional feature vectors from the input feature blocks. The design objective of the contrastive loss function is: for positive sample pairs, to maximize the cosine similarity between the two feature vectors, enabling the encoder to map different enhanced representations of the same state to similar positions in the feature space; for negative sample pairs, to minimize the cosine similarity between the two feature vectors, enabling the encoder to push the feature representations of different states to mutually distant positions in the feature space. Let the two feature vectors of a positive sample pair be... and The two feature vectors of the negative sample pair are and Comparison of loss functions Defined as: ,in This indicates the calculation of cosine similarity. This is a temperature coefficient used to adjust the concentration of the characteristic distribution. It is achieved by minimizing... The encoder network is trained iteratively through backpropagation, enabling the encoder to gradually learn the feature representations that distinguish between normal and abnormal states.

[0086] During iterative training, clustering operations are performed synchronously on the feature vectors of the current batch of sensor nodes to extract the cluster centers of normal patterns. The clustering operation employs a distance-based iterative clustering method, grouping densely distributed feature vectors in the feature space into normal pattern clusters and calculating the center vector of each cluster as the normal pattern cluster center. Let the th... In the next iteration, the normal pattern cluster centers are: , No. In the next iteration, the normal pattern cluster centers are: The displacement distance between adjacent iterations for: ||, when there are several consecutive iterations All are below the preset convergence threshold At this point, the cluster centers of the normal pattern are considered to have stabilized and are fixed, no longer updated during subsequent training. This fixing operation ensures that the reference benchmark for the normal pattern remains consistent in subsequent inference stages, avoiding the impact of minor fluctuations in the later stages of training on the stability of anomaly detection.

[0087] After training, the complete bidirectional correlation feature matrix is ​​input into the trained encoder network to extract feature vectors for each sensor node's corresponding feature block. The Euclidean distance between each sensor node's feature vector and the fixed normal pattern cluster centers is calculated. A larger distance indicates that the node's current state deviates further from the normal pattern, and the corresponding anomaly probability is higher. To ensure comparability of distance values ​​between different nodes, a minimum-maximum normalization process is performed on the distance values ​​of all nodes, mapping them to... The interval is used to obtain the anomaly probability value for each node. The normalization formula is: ,in Let be the Euclidean distance between the feature vector of a certain sensor node and the cluster center of the normal pattern. and These are the minimum and maximum distance values ​​among all nodes. Based on the actual spatial location of each sensor node on the structure, the anomaly probability values ​​are... Mapping to the corresponding spatial coordinates, the spatial coordinate grid is interpolated and smoothed to generate a continuous anomaly probability distribution map. The anomaly probability distribution map is presented in the form of a heatmap, with high anomaly probability areas marked with striking colors, providing an intuitive spatial reference for subsequent anomaly area identification and causal tracing.

[0088] Calculate the neighborhood density for the feature vectors of all sensor nodes, and statistically analyze the neighborhood density within a fixed radius, centered on the feature vector of each node. The more neighboring feature vectors a node has within its range, the denser its neighborhood is considered to be in the feature space. The search for neighborhood density peaks—nodes with the highest local density and distances exceeding a certain threshold from higher density points—is used to define pattern cluster centers. These centers represent typical states in the feature space, potentially corresponding to normal patterns or different types of anomalous patterns. Each sensor node is assigned to the pattern cluster center closest to its feature vector, resulting in the pattern cluster partitioning.

[0089] For each sensor node within a pattern cluster, its corresponding stress time-series data are aggregated, and statistical characteristics such as mean, standard deviation, peak factor, and waveform skewness are calculated. Based on the range and combination patterns of these statistical characteristics, the pattern cluster is matched with predefined anomaly types. For example, pattern clusters with significantly high mean and large waveform skewness correspond to sustained overload anomalies; pattern clusters with a sudden increase in standard deviation but no significant change in mean correspond to vibration and shock anomalies; and pattern clusters with abnormally high peak factors correspond to localized stress concentration anomalies. Anomaly pattern classification labels are assigned to each pattern cluster based on the matching results, and all sensor nodes within the same pattern cluster share the same classification label. These classification labels, along with the anomaly probability distribution map, constitute the input for subsequent causal interpretation mapping, enabling the anomaly diagnosis results to possess not only spatial localization capabilities but also semantic interpretation capabilities for anomaly types.

[0090] Anomaly regions are identified based on the anomaly probability distribution map. The sensor node association paths corresponding to the anomaly regions are traced backward along the temporal and spatial dimensions in the bidirectional correlation feature matrix. Propagation nodes along the association paths are labeled with anomaly pattern classification tags, generating a causal explanation map containing the anomaly source node and its propagation links.

[0091] Set an anomaly threshold on the anomaly probability distribution map to extract sensor nodes with anomaly probability values ​​greater than the anomaly threshold and aggregate them to identify abnormal regions.

[0092] From the sensor nodes in the abnormal region, trace back along the spatial dimension of the spatial correlation submatrix of the bidirectional correlation feature matrix to find the spatial correlation path with increasing spatial correlation coefficient, and trace back along the temporal dimension of the temporal correlation submatrix to find the temporal correlation path with increasing temporal causal strength value. Merge the spatial correlation path and the temporal correlation path to form a sensor node correlation path network.

[0093] Traverse the sensor node association path network to calculate the gradient change rate of the anomaly probability value of the sensor node on each association path, and take the sensor node corresponding to the maximum gradient change rate as the anomaly source node.

[0094] The propagation path is extracted by forward traversal of the path associated with the anomaly source node in the sensor node network. The sensor nodes on the propagation path are labeled as propagation nodes by combining the anomaly pattern classification label, and a causal explanation graph containing the anomaly source node and the propagation path is generated.

[0095] After obtaining the anomaly probability distribution map, it needs to be threshold-filtered to extract the sensor node regions that truly exhibit anomalies. Anomaly thresholds are set for the anomaly probability distribution map. Normalize all outlier probability values Greater than Sensor nodes are marked as candidate anomalous nodes. Anomalous threshold. The settings can be determined by combining the statistical distribution characteristics of historical data. For example, the mean of the anomaly probability distribution under normal operating conditions plus twice the standard deviation can be used as a reference benchmark, thereby controlling the false alarm rate while ensuring the recall rate. For the selected candidate anomaly nodes, they are aggregated according to their physical proximity in the spatial layout of the structure and their network topology connectivity. Candidate nodes that are spatially connected or have direct correlation edges in the correlation matrix are grouped into the same anomaly region, ultimately outputting several discrete or continuously distributed sets of anomaly regions. Each anomaly region represents a potential stress concentration or damage diffusion area on the structure, providing a spatial starting point for subsequent correlation path tracing.

[0096] After identifying the abnormal region, starting from each sensor node within the abnormal region, the spatial correlation submatrix of the bidirectional correlation feature matrix is ​​respectively... Time-related submatrix In reverse tracing, within the spatial correlation submatrix. In the middle, the spatial correlation coefficient with the current node is traced step by step along the spatial dimension using a reverse search strategy. Increasingly adjacent nodes, that is, prioritizing upstream nodes with stronger spatial correlation to the current node, form a spatial association path extending from the anomalous region node into the interior of the structure. This is reflected in the temporal correlation submatrix. In the middle, the temporal causal strength value is traced along the temporal dimension using a reverse causal search strategy. Incremental precursor nodes, that is, nodes with stronger causal driving force prior to the occurrence of the anomaly on the timeline, form a temporal correlation path. "Reverse tracing" refers to starting from the known result node of the anomaly and tracing upstream in the reverse direction of the causal path to reconstruct the source of the anomalous signal. Spatial correlation paths depict the diffusion trajectory of anomalies in physical space, while temporal correlation paths depict the causal transmission chain of anomalies on the timeline. Both describe the propagation mechanism of the same anomalous event from different dimensions.

[0097] Spatial and temporal correlation paths are fused to construct a sensor node correlation path network. The fusion strategy is as follows: for the same node pair... If a node has associated edges in both the spatial and temporal correlation paths, the comprehensive weight of that associated edge is the weighted sum of the corresponding weights of the two paths. The weight coefficient can be adjusted according to the specific structure type and monitoring target. If a node has an associated edge in only one path, that edge is retained and assigned a corresponding single-dimensional weight. The fused sensor node correlation path network is a directed weighted graph. Each directed edge in the graph represents an anomaly propagation channel from an upstream node to a downstream node. The edge weight comprehensively reflects the correlation strength in both spatial correlation and temporal causality. This network covers all possible propagation paths from potential anomaly sources to known anomaly areas, providing a complete set of candidate paths for subsequent anomaly source node localization.

[0098] After the sensor node association path network is constructed, each association path in the network is traversed, and the gradient rate of change of the anomaly probability value of each sensor node on the path is calculated. For two adjacent nodes on the path... and Its abnormal probability value gradient rate of change Defined as the ratio of the difference in normalized outlier probabilities between two nodes to the path step size, i.e. ,in and They are nodes With nodes The normalized outlier probability value, This represents the path step length (expressed as hops or physical distance) between two nodes in the associated path network. Gradient rate of change. This reflects the degree of local abrupt change in the anomaly probability value during spatial propagation: if the gradient change rate at a certain node is significantly higher than that at other nodes, it indicates that this node is the turning point where the anomaly signal suddenly rises from a low-probability region to a high-probability region, i.e., the initial triggering location of the anomaly. Iterate through all associated paths, calculate the gradient change rate of all node pairs on each path, and take the maximum gradient change rate. The corresponding sensor node is used as a candidate anomaly source node for that path. If multiple paths point to the same node, the anomaly source confidence of that node is further enhanced; if different paths point to different nodes, the candidate nodes are sorted according to the comprehensive weight of each path, and the candidate node with the highest comprehensive weight is taken as the final anomaly source node.

[0099] After identifying the anomaly source node, a forward traversal is performed from the source node through the sensor node association path network to extract all directed propagation links from the source node to each anomaly region node. The forward traversal uses a breadth-first search, prioritizing nodes with decreasing edge weights, sequentially visiting downstream nodes directly or indirectly connected to the source node, forming a propagation tree structure with the source node as the root node and each anomaly region node as its leaf node. Each path from the root to a leaf in the propagation tree represents a complete propagation link, signifying the specific diffusion path of the anomalous signal from its source through several intermediate nodes to finally reach the anomaly region.

[0100] After extracting the propagation path, semantic annotation is performed on each sensor node along the propagation path using the anomaly pattern classification labels output from the contrastive learning phase. The anomaly pattern classification labels categorize different types of anomalies into several patterns, such as fatigue crack propagation mode, local stress concentration mode, and connection loosening mode. For each node in the propagation path, a corresponding pattern label is assigned to the node based on the anomaly pattern category of its feature vector, indicating its role in the entire anomaly propagation process—that is, the anomaly pattern in which the node participates in the propagation. After annotation, the relationships between the anomaly source node, the propagation path, and the propagation nodes are organized in a graph format, generating a causal explanation graph containing the anomaly source node and the propagation path. Each node in the graph carries its spatial location information and anomaly probability value. The system includes anomaly pattern classification labels and their hierarchical positions in the propagation chain. Each directed edge carries a comprehensive correlation weight, which fully describes the causal structure of the entire process from the origin to the spread of the anomaly. This provides an interpretable and structured basis for subsequent historical maintenance record matching and maintenance decision generation.

[0101] Obtain historical maintenance records, match the anomaly source nodes in the causal explanation graph with the fault locations in the historical maintenance records, calculate the matching accuracy as the confidence level of the causal explanation graph, and output the validated causal explanation graph, including:

[0102] Retrieve historical maintenance records containing maintenance timestamps, fault location coordinates, and maintenance type labels;

[0103] Extract the timestamps and spatial coordinates of the anomaly source nodes in the causal explanation graph, retrieve the fault locations within the time window from the historical maintenance records as discrete fault locations, input the discrete fault locations into the bidirectional correlation feature matrix to calculate the bidirectional propagation probability between discrete fault locations, and connect the discrete fault locations according to the bidirectional propagation probability to construct the historical fault propagation link.

[0104] Extract the propagation links corresponding to the anomaly source nodes in the causal explanation graph, match the node sequences of the propagation links with the node sequences of historical fault propagation links using the longest common subsequence, and calculate the ratio of the length of the longest common subsequence to the length of the node sequence as the link matching degree between the anomaly source node and the fault location.

[0105] Extract the abnormal pattern classification label sequence of nodes in the propagation link and the maintenance type label sequence of nodes in the historical fault propagation link, and calculate the edit distance of the label sequence to correct the link matching degree.

[0106] The ratio of the number of abnormal source nodes corresponding to the corrected link matching degree to the total number of abnormal source nodes is used as the matching accuracy. The matching accuracy is used as the confidence level of the causal explanation graph. The confidence level of abnormal source nodes is marked in the causal explanation graph, and the verified causal explanation graph is output.

[0107] When acquiring historical maintenance records, it is essential to ensure that the records contain three key fields: maintenance timestamp, fault location coordinates, and maintenance type label. The maintenance timestamp is stored in a standardized format, accurate to the minute, for subsequent time window alignment. The fault location coordinates are represented using a spatial coordinate system consistent with the sensor node deployment coordinate system, ensuring that distance calculations can be performed within the same reference frame. The maintenance type label is annotated according to a predefined fault classification system, such as categories like cracks, corrosion, and fatigue damage, maintaining a correspondence with the anomaly pattern classification labels output during the comparative learning phase. Historical maintenance records typically originate from the structure operation and maintenance management platform. Before importing, they must undergo format standardization processing, removing records with missing fields or abnormal coordinates to ensure data quality for subsequent matching operations.

[0108] After extracting the timestamp and spatial coordinates of each anomaly source node in the causal explanation graph, a time retrieval window is formed by extending a preset time length forward and backward from the timestamp of the anomaly source node. In historical maintenance records, all records whose maintenance timestamps fall within this retrieval window are selected, and the corresponding fault location coordinates are used as a set of discrete fault locations. These discrete fault locations are spatially dispersed, and their propagation relationships have not yet been established; therefore, they need to be quantified using a bidirectional correlation feature matrix. The discrete fault locations are mapped to their nearest neighbors in the sensor node network. The correlation strength between any two discrete fault locations is then read using the constructed bidirectional correlation feature matrix, and this correlation strength is converted into a bidirectional propagation probability. Let the nodes... With nodes The correlation strength value read from the bidirectional correlation feature matrix The corresponding bidirectional propagation probability The result obtained after normalization is: The summation iterates through all the integers. Adjacent discrete fault location nodes. Discrete fault locations are connected in a directed manner according to bidirectional propagation probability, prioritizing the connection of node pairs with higher propagation probabilities, ultimately forming a historical fault propagation link. This link is stored as a node sequence, preserving the spatial propagation order information of historical faults within the structure.

[0109] The propagation path corresponding to the current anomaly source node is extracted from the causal interpretation graph, also represented as a node sequence. The node sequence of this propagation path is then matched with the node sequences of historical fault propagation paths using the longest common subsequence (LCS) algorithm. The LCS algorithm, through dynamic programming, finds the longest subsequence in which the same nodes appear in the same order in two node sequences, without requiring the subsequences to be consecutive. Let the length of the current propagation path node sequence be... The length of the historical fault propagation link node sequence is The length of the longest common subsequence is Then the link matching degree Defined as: Taking the larger of the two sequence lengths as the denominator can avoid the problem of artificially high matching degree due to a large difference in sequence length. The range of values ​​is The closer the value is to 1, the higher the node coverage of the current propagation link and the historical fault propagation link, and the stronger the similarity of the spatial propagation path.

[0110] Relying solely on the structural similarity of node sequences for matching may overlook the semantic differences attached to nodes. Therefore, we further extract the anomaly pattern classification label for each node in the propagation chain to form the current label sequence; simultaneously, we extract the maintenance type label for the corresponding node in the historical fault propagation chain to form the historical label sequence. The edit distance between the two label sequences is then calculated. This refers to the minimum number of single-step operations (including insertion, deletion, and replacement) required to convert the current tag sequence into a historical tag sequence. The link matching degree is corrected using edit distance; the corrected link matching degree... Expressed as: ,in The larger of the two label sequence lengths is used to normalize the edit distance to... The matching interval is defined as follows: When two label sequences are completely identical, the edit distance is 0, the correction term is 1, and the corrected matching degree is the same as the original matching degree. When the label sequences differ significantly, the correction term tends to be 0, and the matching degree is significantly suppressed, reflecting the semantic inconsistency between the current anomaly pattern and the historical fault type. This correction mechanism ensures that the matching degree takes into account both the structural similarity of the propagation path and the semantic consistency of the fault type, thus improving the reliability of the verification results.

[0111] Set link matching threshold Calculate the corrected link matching degree for all abnormal source nodes. Afterwards, the statistics satisfy Number of abnormal source nodes Total number of abnormal source nodes The ratio of these two values ​​is defined as the matching accuracy. ,Right now The matching accuracy reflects the overall degree of agreement between the current causal explanation map and historical maintenance experience. A higher value indicates a greater consistency between the location results of the abnormal source nodes in the map and the spatial distribution and propagation patterns of historical fault records, thus increasing the reliability of the map. As a global confidence index for the causal explanation graph, it also includes the corrected link matching degree corresponding to each anomaly source node. As the local confidence level of a node, it is attached to the corresponding node in the causal explanation graph using numerical labels or color coding, forming a verified causal explanation graph with confidence information. Confidence labeling enables subsequent decision-making stages to distinguish between high-confidence and low-confidence anomalies, providing a quantitative basis for maintaining priority ranking. The output verified causal explanation graph simultaneously retains the original propagation chain structure, anomaly pattern classification labels, and confidence values ​​at each level, providing a complete diagnostic basis for structural maintenance decisions.

[0112] like Figure 2 As shown, Figure 2 This embodiment demonstrates the flowchart for dynamic risk assessment and maintenance decision-making based on causal explanation graphs for abnormal propagation.

[0113] Based on the verified causal explanation graph, the propagation path of the anomaly source nodes is located. Maintenance priorities are determined by combining confidence ranking, and a decision scheme including maintenance location and execution order is generated, including:

[0114] Extract the propagation path and confidence level label corresponding to the anomaly source node from the verified causal explanation graph, and obtain the anomaly probability value sequence, anomaly pattern classification label sequence, spatial coordinate sequence and timestamp sequence of the propagation node on the propagation path;

[0115] The abnormal probability value sequence is divided into uncontrolled propagation segments by identifying the propagation acceleration inflection point through second-order difference. The frequency of abnormal pattern classification label transfer within the uncontrolled propagation segment is statistically analyzed to construct a transfer probability matrix. The maximum eigenvalue of the transfer probability matrix is ​​extracted as the pattern evolution rate. The dynamic risk index is obtained by exponentially operating the confidence label with the pattern evolution rate as the exponent.

[0116] The mutual information of the joint probability distribution of the spatial coordinate sequence and the timestamp sequence of the propagation link is calculated as the spatiotemporal coupling strength. Based on the spatiotemporal coupling strength, competitive groups and non-competitive groups are divided. For the competitive group, the maintenance time interval is converted according to the dynamic risk index difference to determine the maintenance priority. For the non-competitive group, the maintenance decision space is constructed by the dynamic risk index and the spatial expansion rate, and the maintenance priority is determined by Euclidean distance clustering and two-level sorting.

[0117] The maintenance priorities of the competing and non-competitive groups are merged to generate a complete maintenance priority sequence. The spatial coordinates of the anomaly source nodes are extracted as maintenance locations, and their arrangement order is used as the execution order. A decision scheme containing maintenance locations and execution order is then generated.

[0118] The propagation path and its confidence level labeling information corresponding to each anomaly source node are extracted from the validated causal explanation graph. Each propagation node on the propagation path carries four types of attribute sequences: anomaly probability value sequence, anomaly pattern classification label sequence, spatial coordinate sequence, and timestamp sequence. The anomaly probability value sequence reflects the intensity change of the anomaly along the propagation path over time; the spatial coordinate sequence records the physical location of each propagation node in the three-dimensional space of the structure; and the timestamp sequence marks the time when the anomaly signal was detected at each node. These four sequences together constitute the raw input for subsequent dynamic risk assessment and maintenance priority ranking.

[0119] Performing a second-order difference operation on the sequence of anomalous probability values ​​can identify the inflection points of propagation acceleration. Let the nth... The anomaly probability value of each node is Then the first difference is divided into The second difference is divided into .when The value changes from negative to positive and its absolute value exceeds the preset acceleration threshold. At this point, the location is marked as a turning point in propagation acceleration, and the corresponding link segment is designated as the out-of-control propagation segment. The out-of-control propagation segment means that the anomaly is spreading at an accelerated pace within this range, and it is an area that needs to be focused on in risk assessment.

[0120] After identifying the out-of-control propagation segment, the frequency of transitions between abnormal pattern classification labels between adjacent nodes within that segment is counted, and a transition probability matrix is ​​constructed. The first in the matrix Line 1 Column elements This indicates that the abnormal pattern within the uncontrolled propagation segment is classified from category. Transfer to Category The probability of [the transition probability matrix]. Maximum eigenvalue As the rate of pattern evolution, A larger value indicates a more dramatic and unstable evolution of the anomalous pattern during its propagation. The evolution rate of the pattern... The index is labeled with the confidence level corresponding to this propagation path. Perform exponentiation to obtain the dynamic risk index. ,Right now The dynamic risk index integrates information from two dimensions: map confidence and anomalous evolution rate. High confidence and high evolution rate jointly drive a higher dynamic risk index, thus making the assessment of maintenance priorities more meaningful in practical physics.

[0121] The mutual information of the joint probability distribution of the spatial coordinate sequence and the timestamp sequence in the propagation link is calculated to quantify the spatiotemporal coupling strength. Let the random variable after discretization of the spatial coordinates be... The random variable after timestamp discretization is Then the spatiotemporal coupling strength Depend on Mutual information formula The calculation yielded, where For joint probability distribution, and These are the marginal probability distributions. The larger the value, the stronger the coupling between the anomaly and the anomaly in the propagation path in both spatial and temporal dimensions, indicating that the anomaly propagation exhibits significant spatiotemporal synchronicity. This is based on a preset spatiotemporal coupling strength threshold. Divide all propagation links into competing and non-competitive groups: The fact that the links are grouped into the competition group indicates that these links have a competitive diffusion relationship in space and time and need to be coordinated. Links that are assigned to non-competitive groups can have their maintenance order arranged relatively independently.

[0122] For the propagation links within the competing group, calculate the difference between the dynamic risk indices corresponding to any two links. ,in and They are respectively the first in the competition group Article and No. The dynamic risk index of the link. The difference... Transformed into maintenance time intervals via linear mapping The larger the difference, the shorter the corresponding maintenance interval, indicating a clear risk gradient between the two links, and the link with higher risk should be prioritized. Based on the dynamic risk index arranged from high to low, and combined with the maintenance interval constraint, the maintenance priority sequence of each link in the competition group is determined.

[0123] For propagation links within non-competitive groups, the dynamic risk index of each link is used. With spatial expansion rate A two-dimensional maintenance decision space is constructed using two evaluation dimensions. Space expansion rate. Defined as the mean of the Euclidean distances between adjacent nodes in the spatial coordinate sequence of the propagation link, it reflects the spatial diffusion speed of anomalies. In this two-dimensional decision space, all non-competitive links are clustered according to Euclidean distance, and links with similar risk characteristics are grouped into the same maintenance batch. Based on the clustering results, a two-level sorting is performed: the outer layer arranges each maintenance batch from high to low according to the comprehensive risk level of the cluster centers, and the inner layer arranges the links within the same batch from high to low according to the dynamic risk index, ultimately forming a maintenance priority sequence for non-competitive groups.

[0124] The maintenance priority sequences of competing and non-competitive groups are merged to generate a complete maintenance priority sequence. The merging strategy is as follows: the highest priority links in the competing group are placed before the highest priority links in the non-competitive group, because links in the competing group have stronger spatiotemporal coupling, and delayed processing may cause synchronous deterioration of multiple links. The spatial coordinates of the anomaly source node corresponding to each propagation link are extracted as the maintenance location, and the order of arrangement in the maintenance priority sequence is used as the execution order, ultimately generating a decision scheme that includes the maintenance location and execution order. This decision scheme is output in the form of a structured table, with each row corresponding to a maintenance task, including fields such as anomaly source node number, physical spatial coordinates, the identifier of the propagation link, dynamic risk index, spatiotemporal coupling strength, maintenance priority number, and suggested execution time window, for on-site maintenance personnel to directly refer to and execute, thereby realizing a closed loop of the entire process from anomaly detection to maintenance decision-making.

[0125] The abnormal probability value sequence is divided into out-of-control propagation segments by identifying propagation acceleration inflection points using second-order difference analysis. A transition probability matrix is ​​constructed by statistically analyzing the frequency of abnormal pattern classification label transitions within the out-of-control propagation segments. The largest eigenvalue of the transition probability matrix is ​​extracted as the pattern evolution rate. A dynamic risk index is obtained by exponentially multiplying the confidence level labels by the pattern evolution rate, including:

[0126] The propagation acceleration sequence is obtained by performing second-order difference on the abnormal probability value sequence. The position where the propagation acceleration changes from negative to zero is identified as the propagation acceleration inflection point. The propagation nodes before the propagation acceleration inflection point are divided into the runaway propagation segment.

[0127] Extract the classification labels of abnormal patterns within the out-of-control propagation segment and construct an abnormal pattern label chain in chronological order. Calculate the Shannon entropy of the abnormal pattern label chain as the pattern disorder degree. Use the pattern disorder degree as a weight to weight the abnormal probability values ​​of the propagation nodes within the out-of-control propagation segment to obtain a weighted abnormal probability value sequence.

[0128] A transition frequency matrix is ​​constructed by counting the transition frequencies of adjacent label pairs in the statistical abnormal pattern label chain. The transition frequency matrix is ​​then normalized by row to obtain the transition probability matrix. Eigenvalue decomposition is performed on the transition probability matrix to extract the eigenvalue with the largest modulus as the pattern evolution rate.

[0129] The variance of the weighted anomaly probability value sequence is calculated as the propagation fluctuation intensity. The model evolution rate and the propagation fluctuation intensity are weighted and fused together as the power exponent. The confidence level label is used as the base to perform power operation to obtain the dynamic risk index.

[0130] After obtaining the sequence of anomaly probability values ​​for each node on the propagation path, a second-order difference operation is performed on this sequence to obtain the propagation acceleration sequence. Specifically, let the sequence be the sequence of anomaly probability values ​​for each node on the propagation path. The anomaly probability value of each node is Then the first difference The second-order difference reflects the rate of change of the anomaly probability between adjacent nodes. This reflects the acceleration or deceleration trend of the rate of change itself, i.e., propagation acceleration. In actual structural monitoring scenarios, when an anomaly spreads outward from the source node, if the propagation acceleration remains positive, it indicates that the anomaly propagation is intensifying; when the propagation acceleration changes from negative to zero or close to zero, it means that the accelerating trend of anomaly propagation is about to stop, and this turning point is the propagation acceleration inflection point. All locations that meet the "from negative to zero" condition are identified, and all propagation nodes before the inflection point are designated as the uncontrolled propagation segment. The nodes within this segment represent the area where the anomaly propagation is most intense and the risk is most concentrated, and are the focus of subsequent analysis. The determination of the propagation acceleration inflection point uses a preset threshold. Apply constraints when The absolute value is less than Furthermore, if the preceding value is negative, the position is considered to meet the turning point condition, thereby avoiding misjudgment caused by numerical noise.

[0131] Anomaly pattern classification labels are extracted from each propagation node within the runaway propagation segment and arranged sequentially according to their temporal order along the propagation path, forming anomaly pattern label chain. This label chain records the complete sequence information of the evolution of the anomaly pattern along the propagation path within the runaway propagation segment. Shannon entropy is calculated for the anomaly pattern label chain as a quantitative indicator of pattern disorder. Assume that a total of [number missing] anomaly patterns appear in the label chain. Different exception pattern categories, the first The frequency of occurrence of each category is Then Shannon entropy The calculation is as follows The more evenly the categories are distributed in the label chain, the greater the Shannon entropy, indicating frequent switching of abnormal patterns and chaotic propagation behavior within the out-of-control propagation segment; conversely, if a certain category dominates, the Shannon entropy is smaller, indicating a relatively simple and stable propagation pattern. The calculated Shannon entropy... As a weighting coefficient, it represents the anomaly probability value of each propagation node within the out-of-control propagation segment. By performing weighting, a weighted anomaly probability value sequence is obtained, where the th... The weighted outlier probability value of each node is The significance of weighting is that the more chaotic the propagation segment, the higher the probability value of node anomalies should be assigned in subsequent fusion calculations, so that the dynamic risk index can truly reflect the complexity of the propagation process.

[0132] Calculate the transition frequency of all adjacent tag pairs in the anomaly pattern tag chain and construct a transition frequency matrix. Assume the tag chain involves... There are several abnormal pattern categories, and the transition frequency matrix is ​​as follows: The square array, in which the first Line 1 The elements of the column record the pattern category. Transfer to Category The frequency of occurrences is determined by normalizing the transition frequency matrix row by row, so that the sum of the elements in each row is 1, thus obtaining the transition probability matrix. , of which elements Indicates the pattern category Transfer to Category The probability of transition. Essentially, it's a Markov transition matrix that describes the dynamic evolution of anomalous patterns within the runaway propagation phase. Perform eigenvalue decomposition, calculate all eigenvalues, and extract the eigenvalue with the largest modulus as the mode evolution rate. In Markov chain theory, the magnitude of the largest eigenvalue reflects the dominant influence strength of the system's state transitions: when... When the value is close to 1, it indicates that the transition of abnormal patterns between categories tends to stabilize; when... A value significantly greater than 1 indicates a dramatic evolution of the abnormal pattern and a continuously amplifying risk of transmission. This is achieved through extraction... As a model evolution rate, it can compress the dynamic characteristics of abnormal patterns within the runaway propagation phase into a single scalar, which facilitates subsequent fusion with other risk indicators.

[0133] Calculate the weighted anomaly probability value sequence The variance of the propagation wave intensity is used as the variance of the propagation wave intensity. Its calculation is ,in This represents the total number of nodes within the uncontrolled propagation segment. This represents the mean of the weighted sequence of anomaly probability values. (Propagation wave intensity) The dispersion of weighted anomaly probability values ​​for each node within the runaway propagation segment was quantified: if the weighted anomaly probability values ​​for each node are concentrated, the propagation fluctuation intensity is small, indicating a relatively uniform propagation process; if the distribution is dispersed, the propagation fluctuation intensity is large, indicating the existence of nodes with concentrated local sudden risks during the propagation process. The model evolution rate was also analyzed. With propagation wave intensity By performing weighted fusion, a comprehensive power index is obtained. ,in and These are the fusion weights for the mode evolution rate and the intensity of propagation fluctuations, respectively, with a sum of 1. These weights can be adjusted according to the relative importance of mode evolution and fluctuation levels in actual engineering scenarios. Labeled with confidence levels. As the base, the comprehensive power exponent As an index, perform exponentiation to calculate the dynamic risk index. .

[0134] Dynamic Risk Index The physical meaning is that when the evolution rate of the propagation process mode is high and the fluctuation intensity is large, the comprehensive power exponent... If the confidence level is relatively high, If it is relatively high (close to 1), then The fact that it remains at a high level indicates that this propagation link presents a serious risk of diffusion under high confidence conditions; if the confidence level... If it is lower, then even Larger The risk index will also be appropriately suppressed due to its small base, preventing overestimation of risk under low confidence conditions. This power operation design with confidence level as the base allows the dynamic risk index to simultaneously reflect the dynamic characteristics of anomaly propagation and the credibility of the causal explanation map, providing a more differentiated quantitative basis for subsequent priority ranking. In scenarios where multiple propagation links compete for ranking, the risk index can be compared with each link. The value can directly determine which link should be prioritized, thereby achieving optimal risk control decisions with limited maintenance resources.

[0135] A second aspect of the present invention provides an AI diagnostic system for abnormal structure monitoring data, comprising:

[0136] The stress acquisition unit is used to acquire the stress time series data of the sensor nodes on the structure and output a multi-node stress dataset.

[0137] The correlation matrix unit is used to perform spatial topological correlation analysis and temporal causal correlation analysis on multi-node stress datasets, and to construct a bidirectional correlation feature matrix that integrates spatial correlation and temporal dependence among sensor nodes.

[0138] The anomaly diagnosis unit is used to perform unlabeled training based on the bidirectional correlation feature matrix using a contrastive learning algorithm, learn the feature boundaries between normal and abnormal patterns, and output an anomaly probability distribution map and anomaly pattern classification label.

[0139] The causal graph unit is used to identify abnormal regions based on the anomaly probability distribution map, trace the sensor node association path corresponding to the abnormal region in the bidirectional association feature matrix along the temporal and spatial dimensions, and combine the anomaly pattern classification label to mark the propagation nodes on the association path to generate a causal explanation graph containing the anomaly source node and the propagation link.

[0140] The verification and evaluation unit is used to obtain historical maintenance records, match the anomaly source nodes in the causal explanation graph with the fault locations in the historical maintenance records, calculate the matching accuracy as the confidence level of the causal explanation graph, and output the verified causal explanation graph.

[0141] The decision planning unit is used to locate the propagation link based on the abnormal source node in the verified causal explanation graph, determine the maintenance priority by combining the confidence ranking, and generate a decision scheme that includes the maintenance location and execution order.

[0142] A third aspect of the present invention provides an electronic device, comprising:

[0143] processor;

[0144] Memory used to store processor-executable instructions;

[0145] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0146] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0147] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0148] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An AI-based diagnostic method for anomalies in structural monitoring data, characterized in that, include: Acquire the stress time series data of sensor nodes on the structure and output a multi-node stress dataset; Spatial topological correlation analysis and temporal causal correlation analysis were performed on the multi-node stress dataset to construct a bidirectional correlation feature matrix that integrates spatial correlation and temporal dependence among sensor nodes. Based on the bidirectional correlation feature matrix, a contrastive learning algorithm is used for unlabeled training to learn the feature boundaries of normal and abnormal patterns, and output an abnormal probability distribution map and abnormal pattern classification labels. Based on the anomaly probability distribution map, the abnormal region is identified. The sensor node association path corresponding to the abnormal region is traced back along the temporal and spatial dimensions in the bidirectional association feature matrix. The propagation nodes on the association path are marked by the anomaly pattern classification label to generate a causal explanation map containing the anomaly source node and the propagation link. Obtain historical maintenance records, match the anomaly source nodes in the causal explanation graph with the fault locations in the historical maintenance records, calculate the matching accuracy as the confidence level of the causal explanation graph, and output the verified causal explanation graph. Based on the anomaly source nodes in the validated causal explanation graph, the propagation path is located, and the maintenance priority is determined by combining the confidence ranking, generating a decision scheme that includes maintenance location and execution order.

2. The method according to claim 1, characterized in that, Spatial topological correlation analysis and temporal causal correlation analysis were performed on the multi-node stress dataset to construct a bidirectional correlation feature matrix that integrates the spatial correlation and temporal dependence among sensor nodes, including: A spatial adjacency matrix is ​​constructed based on the geometric position and structural connection relationship of distributed sensor nodes. The peak value of the cross-correlation function of stress data between sensor node pairs is calculated as the spatial correlation coefficient, and a spatial correlation coefficient matrix is ​​constructed. A threshold is set on the spatial correlation coefficient matrix to filter sensor node pairs with correlation coefficients greater than the threshold. A sliding time window is constructed to segment the multi-node stress dataset. The Granger causality test is used to calculate the temporal causal intensity value of the filtered sensor node pairs in each time window, and a temporal dependency matrix is ​​constructed. The spatial correlation coefficient matrix is ​​arranged according to the spatial location of the sensor nodes to form a spatial correlation submatrix, and the temporal dependence matrix is ​​arranged according to the time order of the sliding time window to form a temporal correlation submatrix. The spatial correlation submatrix and the temporal correlation submatrix are concatenated column by column to form a bidirectional correlation feature matrix.

3. The method according to claim 1, characterized in that, Based on the bidirectional correlation feature matrix, a contrastive learning algorithm is used for unlabeled training to learn the feature boundaries between normal and abnormal patterns. The output includes an anomaly probability distribution map and anomaly pattern classification labels: Feature blocks are obtained by sliding window sampling of the bidirectional correlation feature matrix. Enhanced feature blocks are generated by time flip transformation and spatial rotation transformation of the feature blocks. Positive sample pairs are constructed by pairing the feature blocks with the enhanced feature blocks, and negative sample pairs are constructed by randomly pairing different feature blocks. Positive and negative sample pairs are input into the encoder network to extract feature vectors. A contrastive loss function is constructed to maximize the similarity of feature vectors of positive sample pairs and minimize the similarity of feature vectors of negative sample pairs. The encoder network is trained iteratively. During the iterative training, the feature vectors are clustered to extract normal pattern cluster centers. When the displacement distance of the normal pattern cluster center between adjacent iterations is continuously lower than the preset convergence threshold, the normal pattern cluster center is solidified. The bidirectional correlation feature matrix is ​​input into the trained encoder network to extract the sensor node feature vectors. The distance between the sensor node feature vectors and the fixed normal pattern cluster centers is calculated and normalized to obtain the anomaly probability value. The anomaly probability distribution map is generated by mapping the anomaly probability value according to the spatial location. The neighborhood density is calculated for the feature vector of the sensor node. The peak point of the neighborhood density is searched as the center of the pattern cluster. The sensor nodes are assigned to the nearest pattern cluster center to form a pattern cluster. The abnormal pattern classification label is assigned according to the statistical features of the stress time series data of the sensor nodes in the pattern cluster.

4. The method according to claim 1, characterized in that, Anomaly regions are identified based on the anomaly probability distribution map. The sensor node association paths corresponding to the anomaly regions are traced backward along the temporal and spatial dimensions in the bidirectional correlation feature matrix. Propagation nodes along the association paths are labeled with anomaly pattern classification tags, generating a causal explanation map containing the anomaly source node and its propagation links. Set an anomaly threshold on the anomaly probability distribution map to extract sensor nodes with anomaly probability values ​​greater than the anomaly threshold and aggregate them to identify abnormal regions. From the sensor nodes in the abnormal region, trace back along the spatial dimension of the spatial correlation submatrix of the bidirectional correlation feature matrix to find the spatial correlation path with increasing spatial correlation coefficient, and trace back along the temporal dimension of the temporal correlation submatrix to find the temporal correlation path with increasing temporal causal strength value. Merge the spatial correlation path and the temporal correlation path to form a sensor node correlation path network. Traverse the sensor node association path network to calculate the gradient change rate of the anomaly probability value of the sensor node on each association path, and take the sensor node corresponding to the maximum gradient change rate as the anomaly source node. The propagation path is extracted by forward traversal of the path associated with the anomaly source node in the sensor node network. The sensor nodes on the propagation path are labeled as propagation nodes by combining the anomaly pattern classification label, and a causal explanation graph containing the anomaly source node and the propagation path is generated.

5. The method according to claim 1, characterized in that, Obtain historical maintenance records, match the anomaly source nodes in the causal explanation graph with the fault locations in the historical maintenance records, calculate the matching accuracy as the confidence level of the causal explanation graph, and output the validated causal explanation graph, including: Retrieve historical maintenance records containing maintenance timestamps, fault location coordinates, and maintenance type labels; Extract the timestamps and spatial coordinates of the anomaly source nodes in the causal explanation graph, retrieve the fault locations within the time window from the historical maintenance records as discrete fault locations, input the discrete fault locations into the bidirectional correlation feature matrix to calculate the bidirectional propagation probability between discrete fault locations, and connect the discrete fault locations according to the bidirectional propagation probability to construct the historical fault propagation link. Extract the propagation links corresponding to the anomaly source nodes in the causal explanation graph, match the node sequences of the propagation links with the node sequences of historical fault propagation links using the longest common subsequence, and calculate the ratio of the length of the longest common subsequence to the length of the node sequence as the link matching degree between the anomaly source node and the fault location. Extract the abnormal pattern classification label sequence of nodes in the propagation link and the maintenance type label sequence of nodes in the historical fault propagation link, and calculate the edit distance of the label sequence to correct the link matching degree. The ratio of the number of abnormal source nodes corresponding to the corrected link matching degree to the total number of abnormal source nodes is used as the matching accuracy. The matching accuracy is used as the confidence level of the causal explanation graph. The confidence level of abnormal source nodes is marked in the causal explanation graph, and the verified causal explanation graph is output.

6. The method according to claim 1, characterized in that, Based on the verified causal explanation graph, the propagation path of the anomaly source nodes is located. Maintenance priorities are determined by combining confidence ranking, and a decision scheme including maintenance location and execution order is generated, including: Extract the propagation path and confidence level label corresponding to the anomaly source node from the verified causal explanation graph, and obtain the anomaly probability value sequence, anomaly pattern classification label sequence, spatial coordinate sequence and timestamp sequence of the propagation node on the propagation path; The abnormal probability value sequence is divided into uncontrolled propagation segments by identifying the propagation acceleration inflection point through second-order difference. The frequency of abnormal pattern classification label transfer within the uncontrolled propagation segment is statistically analyzed to construct a transfer probability matrix. The maximum eigenvalue of the transfer probability matrix is ​​extracted as the pattern evolution rate. The dynamic risk index is obtained by exponentially operating the confidence label with the pattern evolution rate as the exponent. The mutual information of the joint probability distribution of the spatial coordinate sequence and the timestamp sequence of the propagation link is calculated as the spatiotemporal coupling strength. Based on the spatiotemporal coupling strength, competitive groups and non-competitive groups are divided. For the competitive group, the maintenance time interval is converted according to the dynamic risk index difference to determine the maintenance priority. For the non-competitive group, the maintenance decision space is constructed by the dynamic risk index and the spatial expansion rate, and the maintenance priority is determined by Euclidean distance clustering and two-level sorting. The maintenance priorities of the competing and non-competitive groups are merged to generate a complete maintenance priority sequence. The spatial coordinates of the anomaly source nodes are extracted as maintenance locations, and their arrangement order is used as the execution order. A decision scheme containing maintenance locations and execution order is then generated.

7. The method according to claim 6, characterized in that, The abnormal probability value sequence is divided into out-of-control propagation segments by identifying propagation acceleration inflection points using second-order difference analysis. A transition probability matrix is ​​constructed by statistically analyzing the frequency of abnormal pattern classification label transitions within the out-of-control propagation segments. The largest eigenvalue of the transition probability matrix is ​​extracted as the pattern evolution rate. A dynamic risk index is obtained by exponentially multiplying the confidence level labels by the pattern evolution rate, including: The propagation acceleration sequence is obtained by performing second-order difference on the abnormal probability value sequence. The position where the propagation acceleration changes from negative to zero is identified as the propagation acceleration inflection point. The propagation nodes before the propagation acceleration inflection point are divided into the runaway propagation segment. Extract the classification labels of abnormal patterns within the out-of-control propagation segment and construct an abnormal pattern label chain in chronological order. Calculate the Shannon entropy of the abnormal pattern label chain as the pattern disorder degree. Use the pattern disorder degree as a weight to weight the abnormal probability values ​​of the propagation nodes within the out-of-control propagation segment to obtain a weighted abnormal probability value sequence. A transition frequency matrix is ​​constructed by counting the transition frequencies of adjacent label pairs in the statistical abnormal pattern label chain. The transition frequency matrix is ​​then normalized by row to obtain the transition probability matrix. Eigenvalue decomposition is performed on the transition probability matrix to extract the eigenvalue with the largest modulus as the pattern evolution rate. The variance of the weighted anomaly probability value sequence is calculated as the propagation fluctuation intensity. The model evolution rate and the propagation fluctuation intensity are weighted and fused together as the power exponent. The confidence level label is used as the base to perform power operation to obtain the dynamic risk index.

8. An AI diagnostic system for structural monitoring data anomalies, used to implement the method as described in any one of claims 1-7, characterized in that, include: The stress acquisition unit is used to acquire the stress time series data of the sensor nodes on the structure and output a multi-node stress dataset. The correlation matrix unit is used to perform spatial topological correlation analysis and temporal causal correlation analysis on multi-node stress datasets, and to construct a bidirectional correlation feature matrix that integrates spatial correlation and temporal dependence among sensor nodes. The anomaly diagnosis unit is used to perform unlabeled training based on the bidirectional correlation feature matrix using a contrastive learning algorithm, learn the feature boundaries between normal and abnormal patterns, and output an anomaly probability distribution map and anomaly pattern classification label. The causal graph unit is used to identify abnormal regions based on the anomaly probability distribution map, trace the sensor node association path corresponding to the abnormal region in the bidirectional association feature matrix along the temporal and spatial dimensions, and combine the anomaly pattern classification label to mark the propagation nodes on the association path to generate a causal explanation graph containing the anomaly source node and the propagation link. The verification and evaluation unit is used to obtain historical maintenance records, match the anomaly source nodes in the causal explanation graph with the fault locations in the historical maintenance records, calculate the matching accuracy as the confidence level of the causal explanation graph, and output the verified causal explanation graph. The decision planning unit is used to locate the propagation link based on the abnormal source node in the verified causal explanation graph, determine the maintenance priority by combining the confidence ranking, and generate a decision scheme that includes the maintenance location and execution order.

9. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 7.

Images