Ground surface settlement anomaly detection method based on graph neural network

Abstract

The invention discloses a ground surface settlement anomaly detection method based on a graph neural network. The method comprises the steps of 1, collecting multi-source settlement monitoring data in a ground surface settlement monitoring area; 2, constructing a ground surface settlement monitoring map; 3, introducing a Time2Vec model and geological classification codes to construct a multi-modal node feature tensor of the ground surface settlement monitoring map; 4, performing space, time and multi-scale joint modeling on the ground surface settlement monitoring map and the multi-modal node feature tensor through an improved DCRNN model; 5, identifying abnormal nodes and outputting an abnormal monitoring point set; step 6, using an improved mean shift clustering algorithm to execute spatial clustering on the abnormal monitoring point set; and step 7, generating a ground surface settlement abnormity early warning response strategy and an early warning area label map. According to the invention, the space identification precision of ground surface settlement anomaly detection and the real-time performance of risk early warning response are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of geological disaster early warning and intelligent monitoring, and particularly relates to a surface subsidence anomaly detection method based on a graph neural network. BACKGROUND

[0002] With the increase of urban construction density and the complication of geological environment, the demand for surface subsidence monitoring and risk early warning is increasingly urgent. Traditional surface subsidence anomaly detection methods mainly rely on single-point displacement observation or change rate analysis based on empirical thresholds. In the face of large-scale regions, multi-source monitoring data fusion and complex geological backgrounds, the following problems generally exist:

[0003] The existing methods are insufficient in processing multi-source subsidence monitoring data, lack a unified standardized preprocessing and multi-dimensional change rate modeling process, resulting in problems such as non-uniform scales, non-aligned time sequences and difficulty in comprehensively identifying change patterns among monitoring information; anomaly detection methods based on rule setting or moving average cannot depict the spatial propagation characteristics and dynamic evolution rules of subsidence, ignoring the multi-scale diffusion and aggregation trends that subsidence anomalies may exhibit in graph structures; anomaly point detection is usually independent of regional risk clustering and early warning strategies, lacking spatial clustering mechanisms and hierarchical response recommendations oriented to engineering responses, and is difficult to support zoning disposal and risk control in geological disaster emergency management.

[0004] Therefore, how to provide a surface subsidence anomaly detection method based on a graph neural network is a problem that those skilled in the art need to solve. SUMMARY

[0005] One object of the present application is to provide a surface subsidence anomaly detection method based on a graph neural network. The present application makes full use of multi-source monitoring data fusion, graph structure modeling, spatio-temporal diffusion modeling and anomaly clustering analysis technology, and describes in detail the process from original monitoring data collection and processing, multi-modal feature construction, spatio-temporal graph convolution modeling, node anomaly score generation, cluster identification to anomaly region early warning response. The improved DCRNN model introduces multi-scale diffusion modeling and anomaly perception mechanism, realizes dynamic modeling of surface subsidence state and anomaly node identification; an improved weighted mean shift algorithm is used to spatially cluster the anomaly nodes, generate an anomaly region mask graph and a risk score sequence, and match a preset early warning strategy. The present application has the advantages of accurate anomaly identification, stable spatial clustering, clear early warning response and strong adaptability.

[0006] According to the surface subsidence anomaly detection method based on the graph neural network of the present application, the following steps are included:

[0007] Step 1: Collect multi-source subsidence monitoring data in the surface subsidence monitoring region to form a standardized subsidence data set;

[0008] Step 2: Construct a surface subsidence monitoring map using the surface subsidence monitoring points as graph nodes;

[0009] Step 3: Based on the standardized subsidence dataset, introduce the Time2Vec model and geological classification coding to construct a multimodal node feature tensor for the surface subsidence monitoring map;

[0010] Step 4: Input the surface subsidence monitoring map and multimodal node feature tensor into the improved DCRNN model. The network structure includes the introduction of a multi-scale diffusion structure and a structural residual fusion path, and performs joint modeling of space, time and multi-scale to generate subsidence evolution feature tensors.

[0011] Step 5: Based on the settlement evolution feature tensor, calculate the anomaly score value of each graph node at the current time step, identify abnormal nodes by threshold comparison, and output the set of abnormal monitoring points;

[0012] Step 6: Use the improved mean-shift clustering algorithm to perform spatial clustering on the set of anomaly monitoring points, generating anomaly region mask maps and region risk score sequences for each time step;

[0013] Step 7: Based on the anomaly area mask map and the regional risk score sequence, generate a land subsidence anomaly early warning response strategy and an early warning area label map.

[0014] Optionally, step one specifically includes:

[0015] Several surface subsidence monitoring points are set up within the target surface subsidence monitoring area, and multi-source subsidence monitoring data corresponding to the surface subsidence monitoring points are collected.

[0016] The multi-source subsidence monitoring data includes the surface elevation change sequence obtained based on synthetic aperture radar interferometry, the vertical displacement sequence of surface subsidence monitoring points obtained based on the global satellite navigation system, and the ground tilt angle change data obtained based on the tiltmeter equipment;

[0017] The multi-source settlement monitoring data were time-aligned, outliers were removed using the Z-Score method, smoothed using the sliding window averaging method, and normalized using the minimum-maximum normalization formula to obtain the normalized settlement monitoring data sequence.

[0018] Based on the normalized settlement monitoring data sequence, the settlement change rate sequence is calculated. Specifically, for each surface settlement monitoring point, a time-step difference operation is performed on the normalized elevation change sequence, vertical displacement sequence, and ground tilt change sequence to obtain the elevation change rate value, vertical displacement change rate value, and ground tilt change rate value for the corresponding time step. The elevation change rate value, vertical displacement change rate value, and ground tilt change rate value are then concatenated to obtain the settlement change rate vector for each surface settlement monitoring point at each time step.

[0019] Arrange the settlement rate vectors of all time steps in chronological order to form a settlement rate sequence:

[0020] The normalized settlement monitoring sequence and the settlement change rate sequence are concatenated to obtain the settlement monitoring sample matrix of each surface settlement monitoring point, and the settlement monitoring sample matrices of all surface settlement monitoring points are used to form a standardized settlement dataset.

[0021] Optionally, step two specifically includes:

[0022] All surface subsidence monitoring points are used as graph nodes to form a graph node set;

[0023] Set a spatial connectivity threshold, calculate the spatial distance between two graph nodes, and if the spatial distance is less than the spatial connectivity threshold, establish a graph edge between the two graph nodes to form a graph edge set.

[0024] An exponential decay function is used to weight the spatial distances between graph nodes to obtain the spatial distance decay weights between graph nodes, and all spatial distance decay weights are combined into a spatial decay weight matrix.

[0025] Set geological classification codes and obtain the geological classification code to which each surface subsidence monitoring point belongs;

[0026] Establish a geological structure similarity matrix. If the geological classification codes between graph nodes are the same, the corresponding element value in the geological structure similarity matrix is ​​equal to 1; otherwise, the corresponding element value in the geological structure similarity matrix is ​​0.

[0027] The spatial attenuation weight matrix and the geological structure similarity matrix are weighted and fused to generate the graph structure weight matrix;

[0028] The graph structure for monitoring land subsidence is composed of the graph node set, graph edge set, and graph structure weight matrix.

[0029] Optionally, step three specifically includes:

[0030] Based on the standardized settlement dataset, a settlement monitoring sample matrix is ​​extracted for each surface settlement monitoring point.

[0031] The Time2Vec model is used to embed time at each time step, generating a time embedding matrix.

[0032] The settlement monitoring sample matrix and the time-embedded vector sequence are concatenated along the feature dimension to form a graph node time-series feature matrix;

[0033] The geological classification code of each surface subsidence monitoring point is uniquely encoded to generate an auxiliary feature vector. The auxiliary feature vector is then repeatedly extended in the time dimension and concatenated with the time-series feature matrix of the graph node in the feature dimension to obtain the multimodal feature matrix of each graph node.

[0034] The multimodal feature matrices of all graph nodes are arranged in the order of the graph nodes to form the multimodal node feature tensor of the land subsidence monitoring map.

[0035] Optionally, step four specifically includes:

[0036] The surface subsidence monitoring map structure and multimodal node feature tensor are input into the improved DCRNN model. The improved DCRNN model introduces three parallel diffusion modeling branches and structural residual fusion paths. Each diffusion modeling branch consists of a set of dynamic diffusion convolution modules and anomaly perception gated recursive units. Each diffusion modeling branch independently executes the dynamic diffusion convolution and recursive modeling processes.

[0037] In each diffusion modeling branch, the multimodal node feature tensor is split at each time step to obtain node feature matrices for several time steps. The node feature matrices and graph structure weight matrices are then input into the dynamic diffusion convolution module. Spatial diffusion modeling is performed based on the dynamic diffusion graph convolution to obtain the spatial diffusion node feature matrix for each time step.

[0038] The update process of the graph structure weight matrix is ​​as follows: the graph structure weight matrix of the current time step is concatenated with the node feature matrices of the corresponding two nodes, and then input into the MLP structure for mapping. Normalization is performed through the Softmax function to obtain the graph structure weight matrix of the next time step.

[0039] The spatial diffusion feature matrix of each time step is input into the anomaly perception gated recursive unit to perform temporal modeling and obtain the node hidden state matrix. The anomaly perception gated recursive unit introduces the node-level anomaly scoring vector as an additional modulation term in the update gate.

[0040] The node-level anomaly scoring vector is obtained by calculating and normalizing the prediction residual of the previous time step. Specifically, the difference between the node-level model prediction vector and the actual observation vector of each graph node in the previous time step is taken as the node residual vector, and the maximum value of all node residual vectors is normalized to form the node-level anomaly scoring vector.

[0041] The hidden state matrix of each node generated by each diffusion modeling branch is used to generate a scale-aware feature matrix through an MLP structure. The scale-aware feature matrix is ​​then normalized using the Softmax function to obtain the scale adjustment weight matrix of each diffusion modeling branch.

[0042] Perform element-wise multiplication of the hidden state matrices of all nodes at all scales with the corresponding scale adjustment weight matrices and then perform a weighted summation to obtain the node fusion feature matrix;

[0043] In the structural residual fusion path, the node fusion feature matrix is ​​residually fused with the spatial diffusion node feature matrix of the first-order diffusion in the first diffusion modeling branch to obtain the settlement evolution feature matrix at each time step.

[0044] The settlement evolution feature matrices of all time steps are concatenated in chronological order to form a settlement evolution feature tensor.

[0045] Optionally, step five specifically includes:

[0046] Based on the settlement evolution feature tensor, the settlement evolution feature vector of each graph node at each time step is obtained;

[0047] The settlement evolution feature vector is input into the MLP structure to generate the model prediction vector for each graph node at each time step;

[0048] Perform vector difference operation between the model prediction vector and the actual monitoring data at the corresponding time step in the standardized settlement dataset to obtain the prediction residual vector of each graph node at each time step;

[0049] For each graph node, the predicted residual vector is subjected to vector norm calculation to obtain a residual score. The residual score is then normalized to its maximum value to obtain the anomaly score for each graph node at each time step.

[0050] Set a residual judgment threshold. If the abnormal score value is greater than or equal to the residual judgment threshold, the current graph node is judged as an abnormal node at the current time step.

[0051] Organize the time steps, node numbers, and anomaly scores of all abnormal nodes into an anomaly monitoring point set.

[0052] Optionally, step six specifically includes:

[0053] Based on the set of anomaly monitoring points, the time step, graph node number, and anomaly score of each anomaly node are obtained, and the spatial coordinates are obtained according to the graph node number.

[0054] At each time step, the bandwidth parameter for each abnormal node is set, and the abnormal score value is transformed into a weight coefficient through linear mapping;

[0055] Define a set of spatial centroid locations, and calculate the weighted kernel density estimate of the spatial centroid locations at the current time step using a Gaussian kernel function based on bandwidth parameters, weight coefficients, and spatial coordinates.

[0056] In the current clustering iteration, a weighted mean drift update is performed on the spatial centroid position at the current time step to obtain the spatial centroid position for the next round of clustering iteration;

[0057] If the Euclidean distance between the spatial centroid position of the current clustering iteration and the spatial centroid position of the next clustering iteration is less than or equal to the set convergence threshold, it indicates that the centroid has converged, and the current spatial centroid position is output as the cluster center.

[0058] All cluster centers corresponding to the set of spatial centroid locations are combined into a cluster center set at the current time step, and a cluster set is set according to the cluster center set, with each cluster center corresponding to a cluster.

[0059] Calculate the Euclidean distance between each outlier node and all cluster centers at the current time step, and assign each outlier node to the cluster containing the cluster center with the smallest Euclidean distance;

[0060] The spatial coordinates of all abnormal nodes in each cluster are used to construct the cluster space boundary through the convex hull algorithm, and the abnormal region mask unit of each cluster at the current time step is obtained. All abnormal region mask units are combined into an abnormal region mask map.

[0061] Calculate the mean of all outlier scores in each cluster as the regional risk level score for each cluster, and combine all regional risk level scores at the current time step into a regional risk score sequence;

[0062] Output the anomaly region mask map and region risk score sequence for all time steps.

[0063] Optionally, step seven specifically includes:

[0064] The risk level scores of clusters are matched with the set abnormal warning intervals to generate warning level labels. The warning level labels include Level 1 warning, Level 2 warning, Level 3 warning and Level 4 warning; Level 1 warning is the highest warning and Level 4 warning is the lowest warning.

[0065] The specific response strategy for abnormal surface subsidence early warning is as follows: Level 1 warning involves emergency investigation, suspension of construction, and site closure; Level 2 warning involves increased monitoring, on-site inspections, and risk notification; Level 3 warning involves risk alerts, remote alarms, and surrounding signage; and Level 4 warning involves continued observation, weekly reporting, and no intervention.

[0066] Map all warning level labels at the current time step to the abnormal area mask map to form the warning area label map at the current time step, and visualize the warning area label map using a pseudo-color encoding method.

[0067] The beneficial effects of this invention are:

[0068] This invention proposes a surface subsidence anomaly detection method based on graph neural networks, which can effectively improve the processing efficiency of multi-source subsidence monitoring data and the accuracy of spatial anomaly identification. First, by constructing a surface subsidence monitoring map and integrating geological classification coding and Time2Vec temporal embedding, a multimodal node feature tensor is generated, which fully explores the spatial structure information, temporal dependencies, and geological background differences in the subsidence evolution process, effectively enhancing the expressive power of graph structure modeling.

[0069] Secondly, the proposed improved DCRNN model employs a multi-scale diffusion structure and structural residual fusion path, supporting joint modeling of surface subsidence propagation characteristics at different spatial diffusion scales. Furthermore, it introduces node-level anomaly scoring vectors for modulation within the anomaly perception-gated recursive unit, effectively enhancing the model's dynamic response to abrupt subsidence anomalies. Further, it constructs prediction residuals based on the subsidence evolution feature vectors of each graph node, forming node-level anomaly scoring values. Spatial clustering is then performed using an improved mean-shift clustering algorithm, achieving accurate conversion from the set of anomaly monitoring points to an anomaly region mask map. Simultaneously, it generates regional risk scoring sequences to quantify risk levels.

[0070] Furthermore, this method establishes a four-level early warning mechanism based on regional risk level scoring, clarifies the corresponding abnormal early warning response strategies, and combines pseudo-color coding mapping to generate early warning area label maps, thereby realizing intelligent visualization and hierarchical management of land subsidence risk areas.

[0071] In summary, this invention combines multimodal modeling capabilities, spatiotemporal anomaly perception capabilities, and regional risk classification capabilities, significantly improving the accuracy of land subsidence monitoring systems in identifying anomalous areas under complex evolution processes and the pertinence of early warning responses. It has good practical application prospects and promotional value. Attached Figure Description

[0072] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0073] Figure 1 This is a schematic diagram of a surface subsidence anomaly detection method based on graph neural networks proposed in this invention;

[0074] Figure 2 This is a structural diagram of the improved DCRNN model in the surface subsidence anomaly detection method based on graph neural networks proposed in this invention;

[0075] Figure 3 This is a flowchart of anomaly clustering and early warning response in a surface subsidence anomaly detection method based on graph neural networks proposed in this invention. Detailed Implementation

[0076] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0077] refer to Figures 1-3 A method for detecting surface subsidence anomalies based on graph neural networks includes the following steps:

[0078] Step 1: Collect multi-source subsidence monitoring data within the surface subsidence monitoring area to form a standardized subsidence dataset;

[0079] Step 2: Construct a surface subsidence monitoring map using the surface subsidence monitoring points as graph nodes;

[0080] Step 3: Based on the standardized subsidence dataset, introduce the Time2Vec model and geological classification coding to construct a multimodal node feature tensor for the surface subsidence monitoring map;

[0081] Step 4: Input the surface subsidence monitoring map and multimodal node feature tensor into the improved DCRNN model. The network structure includes the introduction of a multi-scale diffusion structure and a structural residual fusion path, and performs joint modeling of space, time and multi-scale to generate subsidence evolution feature tensors.

[0082] Step 5: Based on the settlement evolution feature tensor, calculate the anomaly score value of each graph node at the current time step, identify abnormal nodes by threshold comparison, and output the set of abnormal monitoring points;

[0083] Step 6: Use the improved mean-shift clustering algorithm to perform spatial clustering on the set of anomaly monitoring points, generating anomaly region mask maps and region risk score sequences for each time step;

[0084] Step 7: Based on the anomaly area mask map and the regional risk score sequence, generate a land subsidence anomaly early warning response strategy and an early warning area label map.

[0085] In this embodiment, step one specifically includes:

[0086] Several surface subsidence monitoring points are set up within the target surface subsidence monitoring area, and multi-source subsidence monitoring data corresponding to the surface subsidence monitoring points are collected.

[0087] The multi-source subsidence monitoring data includes the surface elevation change sequence obtained based on synthetic aperture radar interferometry, the vertical displacement sequence of surface subsidence monitoring points obtained based on the global satellite navigation system, and the ground tilt angle change data obtained based on the tiltmeter equipment;

[0088] The multi-source settlement monitoring data were time-aligned, outliers were removed using the Z-Score method, smoothed using the sliding window averaging method, and normalized using the minimum-maximum normalization formula to obtain the normalized settlement monitoring data sequence.

[0089] Based on the normalized settlement monitoring data sequence, the settlement change rate sequence is calculated. Specifically, for each surface settlement monitoring point, a time-step difference operation is performed on the normalized elevation change sequence, vertical displacement sequence, and ground tilt change sequence to obtain the elevation change rate value, vertical displacement change rate value, and ground tilt change rate value for the corresponding time step. The elevation change rate value, vertical displacement change rate value, and ground tilt change rate value are then concatenated to obtain the settlement change rate vector for each surface settlement monitoring point at each time step.

[0090] Arrange the settlement rate vectors of all time steps in chronological order to form a settlement rate sequence:

[0091] The normalized settlement monitoring sequence and the settlement change rate sequence are concatenated to obtain the settlement monitoring sample matrix of each surface settlement monitoring point, and the settlement monitoring sample matrices of all surface settlement monitoring points are used to form a standardized settlement dataset.

[0092] In this embodiment, step two specifically includes:

[0093] All surface subsidence monitoring points are used as graph nodes to form a graph node set;

[0094] Set a spatial connectivity threshold, calculate the spatial distance between two graph nodes, and if the spatial distance is less than the spatial connectivity threshold, establish a graph edge between the two graph nodes to form a graph edge set.

[0095] An exponential decay function is used to weight the spatial distances between graph nodes to obtain the spatial distance decay weights between graph nodes, and all spatial distance decay weights are combined into a spatial decay weight matrix.

[0096] Set geological classification codes and obtain the geological classification code to which each surface subsidence monitoring point belongs;

[0097] Establish a geological structure similarity matrix. If the geological classification codes between graph nodes are the same, the corresponding element value in the geological structure similarity matrix is ​​equal to 1; otherwise, the corresponding element value in the geological structure similarity matrix is ​​0. The spatial attenuation weight matrix and the geological structure similarity matrix are weighted and fused to generate a graph structure weight matrix.

[0098] The graph structure for monitoring land subsidence is composed of the graph node set, graph edge set, and graph structure weight matrix.

[0099] In this embodiment, step three specifically includes:

[0100] Based on the standardized settlement dataset, a settlement monitoring sample matrix is ​​extracted for each surface settlement monitoring point.

[0101] The Time2Vec model is used to embed time at each time step, generating a time embedding matrix.

[0102] The settlement monitoring sample matrix and the time-embedded vector sequence are concatenated along the feature dimension to form a graph node time-series feature matrix;

[0103] The geological classification code of each surface subsidence monitoring point is uniquely encoded to generate an auxiliary feature vector. The auxiliary feature vector is then repeatedly extended in the time dimension and concatenated with the time-series feature matrix of the graph node in the feature dimension to obtain the multimodal feature matrix of each graph node.

[0104] The multimodal feature matrices of all graph nodes are arranged in the order of the graph nodes to form the multimodal node feature tensor of the land subsidence monitoring map.

[0105] In this embodiment, step four specifically includes:

[0106] The surface subsidence monitoring map structure and multimodal node feature tensors are input into an improved DCRNN model. The improved DCRNN model introduces three parallel diffusion modeling branches and a structural residual fusion path. Each diffusion modeling branch consists of a set of dynamic diffusion convolution modules and anomaly perception gated recursive units. Each diffusion modeling branch independently executes the dynamic diffusion convolution and recursive modeling processes to model the surface subsidence propagation characteristics at different spatial diffusion scales. The subsidence evolution feature tensor is output through attention fusion. The dimensions of the multimodal node feature tensor include the number of time steps, the number of graph nodes, and the feature dimension of each node.

[0107] In each diffusion modeling branch, the multimodal node feature tensor is split at each time step to obtain node feature matrices for several time steps. These node feature matrices and graph structure weight matrices are then input into the dynamic diffusion convolution module. Spatial diffusion modeling is performed based on the dynamic diffusion graph convolution to obtain the spatial diffusion node feature matrix for each time step.

[0108] ;

[0109] in, This represents the feature matrix of the spatial diffusion nodes at time step t. This represents the node feature matrix at time step t. Indicates the first The weights of the forward diffusion path of order 1. Indicates the first The weights of the second-order backdiffusion path, Indicates the maximum diffusion order. Indicates the first step at time t The graph structure weight matrix of order diffusion, This indicates the time step t. The transpose of the graph structure weight matrix of order diffusion;

[0110] The update process of the graph structure weight matrix is ​​as follows: the graph structure weight matrix of the current time step is concatenated with the node feature matrices of the corresponding two nodes, and then input into the MLP structure for mapping. Normalization is performed through the Softmax function to obtain the graph structure weight matrix of the next time step.

[0111] The spatial diffusion feature matrix of each time step is input into the anomaly perception gated recursive unit to perform temporal modeling and obtain the node hidden state matrix. The anomaly perception gated recursive unit introduces the node-level anomaly scoring vector as an additional modulation term in the update gate.

[0112] The node-level anomaly scoring vector is obtained by calculating and normalizing the prediction residuals of the previous time step. Specifically, the difference between the node-level model prediction vector and the actual observation vector of each graph node in the previous time step is used as the node residual vector, and the maximum value of all node residual vectors is normalized to form the node-level anomaly scoring vector. The node-level model prediction vector is obtained by mapping the node hidden state matrix through an MLP structure. The actual observation vector refers to the actual monitoring data of each surface subsidence monitoring point in the standardized subsidence dataset in the previous time step.

[0113] The node hidden state matrix generated by each diffusion modeling branch is used to generate a scale-aware feature matrix through an MLP structure. The scale-aware feature matrix is ​​then normalized using the Softmax function to obtain the scale-adjusted weight matrix for each diffusion modeling branch. The scale-adjusted weight matrix is ​​used to measure the contribution of node representation to the final settlement evolution state modeling result at different diffusion scales.

[0114] Perform element-wise multiplication of the hidden state matrices of all nodes at all scales with the corresponding scale adjustment weight matrices and then perform a weighted summation to obtain the node fusion feature matrix;

[0115] In the structural residual fusion path, the node fusion feature matrix is ​​residually fused with the spatial diffusion node feature matrix of the first-order diffusion in the first diffusion modeling branch to obtain the settlement evolution feature matrix at each time step.

[0116] The settlement evolution feature matrices of all time steps are concatenated in chronological order to form a settlement evolution feature tensor.

[0117] In this invention, by introducing three parallel modeling branches with different diffusion scales, the improved DCRNN model's ability to express the characteristics of land subsidence across different spatial propagation ranges is enhanced. Combined with dynamic diffusion convolution and anomaly-aware gated recursive structures, joint modeling of spatial and temporal dependencies and anomaly-driven changes is effectively achieved, improving the accuracy of perceiving subsidence evolution dynamics. Adaptive weighting of multi-scale information through scale-adjusted weights ensures a reasonable allocation of contributions from different diffusion layers to the overall modeling results, enhancing the model's multi-scale adaptability and stability. Furthermore, selecting the first-order diffusion feature of the first diffusion modeling branch as the residual baseline in the structural residual fusion path helps retain the most local and original spatial diffusion information. This serves as a structural compensation benchmark for deep multi-scale fusion features, significantly alleviating feature drift and gradient degradation problems caused by excessively deep network structures, and improving the expressive integrity and physical consistency of the subsidence evolution feature tensor.

[0118] In this embodiment, step five specifically includes:

[0119] Based on the settlement evolution feature tensor, the settlement evolution feature vector of each graph node at each time step is obtained;

[0120] The settlement evolution feature vector is input into the MLP structure to generate the model prediction vector for each graph node at each time step;

[0121] Perform vector difference operation between the model prediction vector and the actual monitoring data at the corresponding time step in the standardized settlement dataset to obtain the prediction residual vector of each graph node at each time step;

[0122] For each graph node, the predicted residual vector is subjected to vector norm calculation to obtain a residual score. The residual score is then normalized to its maximum value to obtain the anomaly score for each graph node at each time step.

[0123] Set a residual judgment threshold. If the abnormal score value is greater than or equal to the residual judgment threshold, the current graph node is judged as an abnormal node at the current time step.

[0124] Organize the time steps, node numbers, and anomaly scores of all abnormal nodes into an anomaly monitoring point set.

[0125] In this embodiment, step six specifically includes:

[0126] Based on the set of anomaly monitoring points, the time step, graph node number, and anomaly score of each anomaly node are obtained, and the spatial coordinates are obtained according to the graph node number.

[0127] At each time step, the bandwidth parameter for each abnormal node is set, and the abnormal score value is transformed into a weight coefficient through linear mapping;

[0128] Given a set of spatial centroid locations, and based on bandwidth parameters, weighting coefficients, and spatial coordinates, calculate the weighted kernel density estimate of the spatial centroid locations at the current time step using a Gaussian kernel function:

[0129] ;

[0130] in, Indicates the position of the spatial centroid at the current time step. The weighted kernel density estimate, where n represents the total number of anomalous nodes at the current time step. This represents the weight coefficient of the i-th abnormal node at the current time step. This represents the spatial coordinates of the i-th anomaly node at the current time step. This represents the bandwidth parameter of the i-th abnormal node at the current time step. Represents the Gaussian kernel function;

[0131] In the current clustering iteration, a weighted mean drift update is performed on the spatial centroid position at the current time step to obtain the spatial centroid position for the next clustering iteration. Specifically, the weighted distance of all outlier nodes relative to the current spatial centroid position is calculated, and a weighted average is performed on the spatial coordinates of each outlier node according to the inverse ratio of the weighted distance. The position obtained by the weighted average is then used as the new spatial centroid position.

[0132] If the Euclidean distance between the spatial centroid position of the current clustering iteration and the spatial centroid position of the next clustering iteration is less than or equal to the set convergence threshold, it indicates that the centroid has converged, and the current spatial centroid position is output as the cluster center.

[0133] All cluster centers corresponding to the set of spatial centroid locations are combined into a cluster center set at the current time step, and a cluster set is set according to the cluster center set, with each cluster center corresponding to a cluster.

[0134] Calculate the Euclidean distance between each outlier node and all cluster centers at the current time step, and assign each outlier node to the cluster containing the cluster center with the smallest Euclidean distance;

[0135] The spatial coordinates of all abnormal nodes in each cluster are used to construct the cluster space boundary through the convex hull algorithm, and the abnormal region mask unit of each cluster at the current time step is obtained. All abnormal region mask units are combined into an abnormal region mask map.

[0136] Calculate the mean of all outlier scores in each cluster as the regional risk level score for each cluster, and combine all regional risk level scores at the current time step into a regional risk score sequence;

[0137] Output the anomaly region mask map and region risk score sequence for all time steps.

[0138] This invention, by introducing weighted kernel density estimation and an improved mean drift algorithm, achieves spatial clustering and regional risk level modeling of surface subsidence anomaly monitoring points. This effectively captures the aggregation characteristics of subsidence anomalies in spatial distribution and constructs anomaly region mask maps with physical boundary significance based on anomaly scores. Simultaneously, by quantifying the mean anomaly intensity of each cluster, a regional risk score sequence is formed, providing a highly reliable spatial risk expression method for subsequent early warning level classification and response strategy formulation. This improves the interpretability of anomaly detection results, spatial positioning accuracy, and the targeted nature of early warning responses.

[0139] In this embodiment, step seven specifically includes:

[0140] The risk level scores of clusters are matched with the set abnormal warning intervals to generate warning level labels. The warning level labels include Level 1 warning, Level 2 warning, Level 3 warning and Level 4 warning; Level 1 warning is the highest warning and Level 4 warning is the lowest warning.

[0141] The specific response strategy for abnormal surface subsidence early warning is as follows: Level 1 warning involves emergency investigation, suspension of construction, and site closure; Level 2 warning involves increased monitoring, on-site inspections, and risk notification; Level 3 warning involves risk alerts, remote alarms, and surrounding signage; and Level 4 warning involves continued observation, weekly reporting, and no intervention.

[0142] Map all warning level labels at the current time step to the abnormal area mask map to form the warning area label map at the current time step, and visualize the warning area label map using a pseudo-color encoding method.

[0143] Example 1

[0144] To verify the feasibility of this invention in practice, it was applied to a surface subsidence risk monitoring task along a subway line. This area is located in a typical soft soil region, and in recent years, due to intensive underground construction and frequent ground disturbance, a significant surface subsidence trend has emerged. Cracks and uneven ground subsidence have occurred in some sections, posing significant structural safety hazards. Traditional monitoring methods rely on manual inspections at fixed stations and data analysis from single devices, making it difficult to accurately identify surface subsidence anomalies and provide regional-level risk warnings. Furthermore, their timeliness is poor, failing to meet the high-frequency, high-precision monitoring requirements for urban underground space safety management.

[0145] In this scenario, the project team deployed a comprehensive monitoring network based on the method of this invention. 342 monitoring points within a 25-kilometer radius of the rail transit line were selected as graph nodes. A synthetic aperture radar interferometric antenna array, GNSS equipment, and a tiltmeter system were deployed to collect multi-source settlement data in real time, including changes in surface elevation, vertical displacement, and ground tilt. The data collection cycle was set to once every 12 hours, with a continuous monitoring period of 4 months, collecting approximately 82,000 data sets. During the data processing stage, standardized settlement datasets were generated through preprocessing methods such as Z-Score anomaly removal, sliding window smoothing, and minimum-maximum normalization. A settlement change rate vector was constructed for each monitoring point, and these vectors were further concatenated to obtain a settlement monitoring sample matrix. Using a graph construction method, geological structure encoding information and spatial distance weighted fusion were used to establish a surface settlement monitoring map. A Time2Vec model was employed for temporal embedding, jointly introducing temporal information and geological features into the multimodal feature tensors of the graph nodes.

[0146] An improved DCRNN network was used to perform spatial, temporal, and multi-scale joint modeling of multimodal feature tensors and land subsidence monitoring maps. In the multi-scale diffusion modeling module, the maximum diffusion order and the number of scale branches were set to 3. An adaptive graph structure weight update mechanism was used for dynamic diffusion map convolution to capture dynamic diffusion relationships. In the anomaly perception gated recursive unit, node-level prediction residual score vectors were introduced as adjustment factors to improve the response capability to sudden anomalies. After model training, the anomaly score values ​​of all graph nodes were calculated at each time step, and a set of anomaly monitoring points was output based on threshold judgment. During anomaly point clustering, an improved mean-shift algorithm was used, incorporating anomaly score weights and a Gaussian kernel function for weighted kernel density estimation, achieving spatial clustering and centroid migration of anomaly nodes, thereby generating anomaly area mask maps and regional risk score sequences. Finally, four-level warning level labels were set based on the score intervals, and a warning area label map was output, realizing dynamic visualization and strategy guidance for subsidence risk areas.

[0147] To compare the performance differences between the method of this invention and existing solutions in real-world scenarios, three typical solutions in the current industry were selected for comparison: the traditional InSAR anomaly identification method, the LSTM-based single-point subsidence trend prediction method, and the spatial anomaly detection method based on the standard GCN model. The methods were evaluated based on six indicators: anomaly detection accuracy, regional early warning positioning accuracy, false alarm rate, response time, spatial clustering quality, and regional early warning level matching rate. The experimental results are shown in Table 1.

[0148] Table 1 Performance Comparison Evaluation Table of the Method of the Invention and the Comparative Scheme

[0149] Method name Abnormality detection accuracy (%) Regional early warning positioning accuracy (%) False alarm rate (%) Average response time (hours) Spatial clustering profile coefficient Regional early warning level matching rate (%) The method of the present application 94.8 91.2 3.6 1.2 0.82 89.5 Traditional InSAR anomaly map spot identification method 79.5 72.3 11.8 12.0 0.59 63.1 Single-point subsidence trend prediction method based on LSTM 83.4 69.7 9.4 6.5 0.48 58.9 Spatial anomaly detection method based on standard GCN 88.7 76.1 7.1 4.8 0.65 71.4

[0150] As shown in Table 1, the method of this invention exhibits significant advantages in all indicators. The anomaly detection accuracy reaches 94.8%, which is 15.3% higher than the traditional InSAR anomaly patch identification method. The average response time is significantly shortened to only 1.2 hours, which is significantly better than the comparative scheme. The spatial clustering contour coefficient reaches 0.82, indicating that the method of this invention has clear boundaries and excellent clustering effect in the spatial clustering process. The method of this invention effectively reduces the false alarm rate to only 3.6% through anomaly scoring and spatial clustering. The matching rate between the warning level label and the settlement impact level recorded in the field investigation reaches 89.5%, which is significantly higher than the comparative scheme, indicating that the method of this invention has strong interpretability and credibility.

[0151] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for detecting surface subsidence anomalies based on graph neural networks, characterized in that, Includes the following steps: Step 1: Collect multi-source subsidence monitoring data within the surface subsidence monitoring area to form a standardized subsidence dataset; Step 2: Construct a surface subsidence monitoring map using the surface subsidence monitoring points as graph nodes; Step 3: Based on the standardized subsidence dataset, introduce the Time2Vec model and geological classification coding to construct a multimodal node feature tensor for the surface subsidence monitoring map; Step 4: Input the surface subsidence monitoring map and multimodal node feature tensor into the improved DCRNN model. The network structure includes the introduction of a multi-scale diffusion structure and a structural residual fusion path, and performs joint modeling of space, time and multi-scale to generate subsidence evolution feature tensors. Step 5: Based on the settlement evolution feature tensor, calculate the anomaly score value of each graph node at the current time step, identify abnormal nodes by threshold comparison, and output the set of abnormal monitoring points; Step 6: Use the improved mean-shift clustering algorithm to perform spatial clustering on the set of anomaly monitoring points, generating anomaly region mask maps and region risk score sequences for each time step; Step 7: Based on the anomaly area mask map and the regional risk score sequence, generate a land subsidence anomaly early warning response strategy and an early warning area label map.

2. The method for detecting surface subsidence anomalies based on graph neural networks according to claim 1, characterized in that, Step one specifically includes: Several surface subsidence monitoring points are set up within the target surface subsidence monitoring area, and multi-source subsidence monitoring data corresponding to the surface subsidence monitoring points are collected. The multi-source subsidence monitoring data includes the surface elevation change sequence obtained based on synthetic aperture radar interferometry, the vertical displacement sequence of surface subsidence monitoring points obtained based on the global satellite navigation system, and the ground tilt angle change data obtained based on the tiltmeter equipment; The multi-source settlement monitoring data were time-aligned, outliers were removed using the Z-Score method, smoothed using the sliding window averaging method, and normalized using the minimum-maximum normalization formula to obtain the normalized settlement monitoring data sequence. Based on the normalized settlement monitoring data sequence, the settlement change rate sequence is calculated. Specifically, for each surface settlement monitoring point, a time-step difference operation is performed on the normalized elevation change sequence, vertical displacement sequence, and ground tilt change sequence to obtain the elevation change rate value, vertical displacement change rate value, and ground tilt change rate value for the corresponding time step. The elevation change rate value, vertical displacement change rate value, and ground tilt change rate value are then concatenated to obtain the settlement change rate vector for each surface settlement monitoring point at each time step. Arrange the settlement rate vectors of all time steps in chronological order to form a settlement rate sequence: The normalized settlement monitoring sequence and the settlement change rate sequence are concatenated to obtain the settlement monitoring sample matrix of each surface settlement monitoring point, and the settlement monitoring sample matrices of all surface settlement monitoring points are used to form a standardized settlement dataset.

3. The method for detecting surface subsidence anomalies based on graph neural networks according to claim 1, characterized in that, Step two specifically includes: All surface subsidence monitoring points are used as graph nodes to form a graph node set; Set a spatial connectivity threshold, calculate the spatial distance between two graph nodes, and if the spatial distance is less than the spatial connectivity threshold, establish a graph edge between the two graph nodes to form a graph edge set. An exponential decay function is used to weight the spatial distances between graph nodes to obtain the spatial distance decay weights between graph nodes, and all spatial distance decay weights are combined into a spatial decay weight matrix. Set geological classification codes and obtain the geological classification code to which each surface subsidence monitoring point belongs; Establish a geological structure similarity matrix. If the geological classification codes between graph nodes are the same, the corresponding element value in the geological structure similarity matrix is ​​equal to 1; otherwise, the corresponding element value in the geological structure similarity matrix is ​​0. The spatial attenuation weight matrix and the geological structure similarity matrix are weighted and fused to generate the graph structure weight matrix; The graph structure for monitoring land subsidence is composed of the graph node set, graph edge set, and graph structure weight matrix.

4. The method for detecting surface subsidence anomalies based on graph neural networks according to claim 1, characterized in that, Step three specifically includes: Based on the standardized settlement dataset, a settlement monitoring sample matrix is ​​extracted for each surface settlement monitoring point. The Time2Vec model is used to embed time at each time step, generating a time embedding matrix. The settlement monitoring sample matrix and the time-embedded vector sequence are concatenated along the feature dimension to form a graph node time-series feature matrix; The geological classification code of each surface subsidence monitoring point is uniquely encoded to generate an auxiliary feature vector. The auxiliary feature vector is then repeatedly extended in the time dimension and concatenated with the time-series feature matrix of the graph node in the feature dimension to obtain the multimodal feature matrix of each graph node. The multimodal feature matrices of all graph nodes are arranged in the order of the graph nodes to form the multimodal node feature tensor of the land subsidence monitoring map.

5. The method for detecting surface subsidence anomalies based on graph neural networks according to claim 1, characterized in that, Step four specifically includes: The surface subsidence monitoring map structure and multimodal node feature tensor are input into the improved DCRNN model. The improved DCRNN model introduces three parallel diffusion modeling branches and structural residual fusion paths. Each diffusion modeling branch consists of a set of dynamic diffusion convolution modules and anomaly perception gated recursive units. Each diffusion modeling branch independently executes the dynamic diffusion convolution and recursive modeling processes. In each diffusion modeling branch, the multimodal node feature tensor is split at each time step to obtain node feature matrices for several time steps. The node feature matrices and graph structure weight matrices are then input into the dynamic diffusion convolution module. Spatial diffusion modeling is performed based on the dynamic diffusion graph convolution to obtain the spatial diffusion node feature matrix for each time step. The update process of the graph structure weight matrix is ​​as follows: the graph structure weight matrix of the current time step is concatenated with the node feature matrices of the corresponding two nodes, and then input into the MLP structure for mapping. Normalization is performed through the Softmax function to obtain the graph structure weight matrix of the next time step. The spatial diffusion feature matrix of each time step is input into the anomaly perception gated recursive unit to perform temporal modeling and obtain the node hidden state matrix. The anomaly perception gated recursive unit introduces the node-level anomaly scoring vector as an additional modulation term in the update gate. The node-level anomaly scoring vector is obtained by calculating and normalizing the prediction residual of the previous time step. Specifically, the difference between the node-level model prediction vector and the actual observation vector of each graph node in the previous time step is taken as the node residual vector, and the maximum value of all node residual vectors is normalized to form the node-level anomaly scoring vector. The hidden state matrix of each node generated by each diffusion modeling branch is used to generate a scale-aware feature matrix through an MLP structure. The scale-aware feature matrix is ​​then normalized using the Softmax function to obtain the scale adjustment weight matrix of each diffusion modeling branch. Perform element-wise multiplication of the hidden state matrices of all nodes at all scales with the corresponding scale adjustment weight matrices and then perform a weighted summation to obtain the node fusion feature matrix; In the structural residual fusion path, the node fusion feature matrix is ​​residually fused with the spatial diffusion node feature matrix of the first-order diffusion in the first diffusion modeling branch to obtain the settlement evolution feature matrix at each time step. The settlement evolution feature matrices of all time steps are concatenated in chronological order to form a settlement evolution feature tensor.

6. The method for detecting surface subsidence anomalies based on graph neural networks according to claim 1, characterized in that, Step five specifically includes: Based on the settlement evolution feature tensor, the settlement evolution feature vector of each graph node at each time step is obtained; The settlement evolution feature vector is input into the MLP structure to generate the model prediction vector for each graph node at each time step; Perform vector difference operation between the model prediction vector and the actual monitoring data at the corresponding time step in the standardized settlement dataset to obtain the prediction residual vector of each graph node at each time step; For each graph node, the predicted residual vector is subjected to vector norm calculation to obtain a residual score. The residual score is then normalized to its maximum value to obtain the anomaly score for each graph node at each time step. Set a residual judgment threshold. If the abnormal score value is greater than or equal to the residual judgment threshold, the current graph node is judged as an abnormal node at the current time step. Organize the time steps, node numbers, and anomaly scores of all abnormal nodes into an anomaly monitoring point set.

7. The method for detecting surface subsidence anomalies based on graph neural networks according to claim 1, characterized in that, Step six specifically includes: Based on the set of anomaly monitoring points, the time step, graph node number, and anomaly score of each anomaly node are obtained, and the spatial coordinates are obtained according to the graph node number. At each time step, the bandwidth parameter for each abnormal node is set, and the abnormal score value is transformed into a weight coefficient through linear mapping; Define a set of spatial centroid locations, and calculate the weighted kernel density estimate of the spatial centroid locations at the current time step using a Gaussian kernel function based on bandwidth parameters, weight coefficients, and spatial coordinates. In the current clustering iteration, a weighted mean drift update is performed on the spatial centroid position at the current time step to obtain the spatial centroid position for the next round of clustering iteration; If the Euclidean distance between the spatial centroid position of the current clustering iteration and the spatial centroid position of the next clustering iteration is less than or equal to the set convergence threshold, it indicates that the centroid has converged, and the current spatial centroid position is output as the cluster center. All cluster centers corresponding to the set of spatial centroid locations are combined into a cluster center set at the current time step, and a cluster set is set according to the cluster center set, with each cluster center corresponding to a cluster. Calculate the Euclidean distance between each outlier node and all cluster centers at the current time step, and assign each outlier node to the cluster containing the cluster center with the smallest Euclidean distance; The spatial coordinates of all abnormal nodes in each cluster are used to construct the cluster space boundary through the convex hull algorithm, and the abnormal region mask unit of each cluster at the current time step is obtained. All abnormal region mask units are combined into an abnormal region mask map. Calculate the mean of all outlier scores in each cluster as the regional risk level score for each cluster, and combine all regional risk level scores at the current time step into a regional risk score sequence; Output the anomaly region mask map and region risk score sequence for all time steps.

8. The method for detecting surface subsidence anomalies based on graph neural networks according to claim 1, characterized in that, Step seven specifically includes: The risk level scores of clusters are matched with the set abnormal warning intervals to generate warning level labels. The warning level labels include Level 1 warning, Level 2 warning, Level 3 warning and Level 4 warning; Level 1 warning is the highest warning and Level 4 warning is the lowest warning. The specific response strategy for abnormal surface subsidence early warning is as follows: Level 1 warning involves emergency investigation, suspension of construction, and site closure; Level 2 warning involves increased monitoring, on-site inspections, and risk notification; Level 3 warning involves risk alerts, remote alarms, and surrounding signage; and Level 4 warning involves continued observation, weekly reporting, and no intervention. Map all warning level labels at the current time step to the abnormal area mask map to form the warning area label map at the current time step, and visualize the warning area label map using a pseudo-color encoding method.

Images