Deep learning-based mining area surface subsidence risk analysis system

Abstract

The invention discloses a mining area surface subsidence risk analysis system based on deep learning, and the system comprises a data construction module which is used for collecting multi-source spatio-temporal data of a mining area, and generating a coding feature tensor; the prediction reasoning module is used for inputting the coding feature tensor into an improved Earthformer model, introducing a causal sensitive cube attention structure and generating a subsidence prediction sequence and space-time hidden state features; the causal analysis module is used for constructing a reverse subsidence sequence and forming a mining area subsidence causal link tensor; the training updating module is used for executing parameter updating according to the joint loss; the risk assessment module is used for identifying a subsidence source point and generating a subsidence source area identification result; and calculating a risk score, carrying out grade division, and outputting a subsidence risk grade map and structured subsidence early warning information. According to the invention, fine identification and high-reliability early warning output of the mining area surface subsidence risk are realized.

Description

Technical Field

[0001] This invention relates to the field of surface subsidence monitoring and intelligent early warning technology, and in particular to a deep learning-based surface subsidence risk analysis system for mining areas. Background Technology

[0002] With the continuous increase in the intensity of underground resource extraction in mining areas, the demand for monitoring surface subsidence evolution, subsidence risk assessment, and source area identification has increased significantly. Subsidence processes are driven by multiple factors, including underground mining void morphology, changes in mining disturbance, differences in geological structure, and fluctuations in hydrological conditions, exhibiting obvious spatiotemporal nonlinear characteristics. Existing methods for analyzing subsidence in mining areas mainly rely on surface deformation monitoring sequences, empirical analysis models, or traditional spatial interpolation methods to estimate subsidence trends. However, these methods generally employ single-moment or local time-period data processing, making it difficult to characterize the temporal cumulative effect and spatial diffusion patterns during subsidence evolution. Existing interpolation and empirical models lack sensitivity to changes in geological properties, hydrological recharge, and abrupt changes in mining disturbance, often exhibiting prediction biases during periods of rapid subsidence increment, affecting the stability and continuity of subsidence trend assessments.

[0003] While existing deep learning prediction models possess some temporal modeling capabilities, they typically only directly fit feature sequences, lacking the ability to model the causal relationships between subsidence driving factors and failing to distinguish between the contribution of external disturbances and internal evolutionary trends. Furthermore, traditional spatiotemporal coding structures are generally based on regular grids or simple convolutional structures to construct feature representations, which are insufficiently adaptable to the heterogeneity of spatial structures in mining areas and slow to respond to abrupt changes in geological properties and fluctuations in hydrological conditions, leading to inaccurate identification of key subsidence source areas and diffusion paths. In the source area identification stage, existing technologies mostly rely on threshold screening or single-scale discrimination methods based on deformation gradients, lacking a systematic link combination mechanism for continuous time periods and adjacent spatial units. This makes the subsidence source tracking results prone to fragmentation, breakage, and omissions, making it difficult to support high-precision subsidence risk level classification.

[0004] Therefore, how to provide a deep learning-based system for analyzing the risk of surface subsidence in mining areas is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a deep learning-based system for analyzing the risk of surface subsidence in mining areas. This invention constructs multi-channel spatiotemporal features based on a multi-source spatiotemporal sample set, generating a spatiotemporal feature tensor and an encoded feature tensor. The encoded feature tensor is input into an improved Earthformer model, incorporating a causal-sensitive cube attention structure, performing cube partitioning and causal weighting processing to generate a subsidence prediction sequence and spatiotemporal latent state features. Based on the subsidence prediction sequence and spatiotemporal latent state features, a reverse subsidence sequence is constructed. A disturbance intensity estimation sequence and a causal contribution sequence are generated through a causal link network, forming a mining area subsidence causal link tensor. Consistency constraints are constructed based on the subsidence prediction sequence, disturbance intensity estimation sequence, and causal contribution sequence. Subsidence reconstruction consistency training and causal link consistency training are performed, updating the parameters of the improved Earthformer model and the causal link network. Based on the mining area subsidence causal link tensor, spatial units with disturbance contributions greater than a preset disturbance threshold are identified, generating subsidence source area identification results. Based on the subsidence source area identification results, a risk score is calculated, levels are classified, and a subsidence risk level map and structured subsidence early warning information are output. This invention has the advantages of high accuracy in subsidence prediction, strong ability to construct causal links, high accuracy in identifying subsidence source areas, and a high degree of structure in the subsidence risk output results.

[0006] A deep learning-based surface subsidence risk analysis system for mining areas according to an embodiment of the present invention includes:

[0007] The data construction module is used to collect multi-source spatiotemporal data from the mining area, perform time alignment and spatial alignment processing to form a multi-source spatiotemporal sample set; construct multi-channel spatiotemporal features to form a spatiotemporal feature tensor, and perform vector dimension mapping processing on the spatiotemporal feature tensor to generate an encoded feature tensor;

[0008] The prediction and inference module is used to input the encoded feature tensor into the improved Earthformer model. The improved Earthformer model introduces a causal sensitive cube attention structure, performs cube partitioning and causal weighting, and generates a sinking prediction sequence and spatiotemporal latent state features.

[0009] The causal analysis module is used to construct a reverse subsidence sequence based on the subsidence prediction sequence and spatiotemporal latent state characteristics, and to generate a disturbance intensity estimation sequence and a causal contribution sequence through a causal link network, thus forming a causal link tensor for subsidence in the mining area.

[0010] The training and update module is used to construct consistency constraints based on the subsidence prediction sequence, the disturbance intensity estimation sequence, and the causal contribution sequence, perform subsidence reconstruction consistency training and causal link consistency training, and update the parameters of the improved Earthformer model and the causal link network according to the joint loss.

[0011] The risk assessment module is used to identify spatial units whose disturbance contribution is greater than a preset disturbance threshold as subsidence source points based on the causal link tensor of the subsidence in the mining area. It performs link combination processing on continuous time segments and adjacent spatial units to generate subsidence source area identification results; calculates risk scores, classifies them into levels, and outputs subsidence risk level maps and structured subsidence early warning information.

[0012] Optionally, modules can be integrated using the following methods:

[0013] S1. Collect multi-source spatiotemporal data from the mining area, perform time alignment and spatial alignment processing, and form a multi-source spatiotemporal sample set;

[0014] S2. Construct multi-channel spatiotemporal features based on multi-source spatiotemporal sample sets to form a spatiotemporal feature tensor, and perform vector dimension mapping processing on the spatiotemporal feature tensor to generate an encoded feature tensor;

[0015] S3. Input the encoded feature tensor into the improved Earthformer model. The improved Earthformer model introduces a causal sensitive cube attention structure, performs cube partitioning and causal weighting, and generates a sinking prediction sequence and spatiotemporal hidden state features.

[0016] S4. Construct an inverse subsidence sequence based on the subsidence prediction sequence and spatiotemporal latent state characteristics. Generate a disturbance intensity estimation sequence and a causal contribution sequence through a causal link network to form a mining area subsidence causal link tensor.

[0017] S5. Based on the subsidence prediction sequence, disturbance intensity estimation sequence and causal contribution sequence, construct consistency constraints, perform subsidence reconstruction consistency training and causal link consistency training, and update the parameters of the improved Earthformer model and causal link network according to the joint loss.

[0018] S6. Based on the spatial unit whose disturbance contribution is greater than the preset disturbance threshold identified by the causal link tensor of the subsidence in the mining area, subsidence source points are identified. Link combination processing is performed on continuous time segments and adjacent spatial units to generate subsidence source area identification results.

[0019] S7. Calculate the risk score based on the subsidence source area identification results, classify the risk levels, and output the subsidence risk level map and structured subsidence early warning information.

[0020] Optionally, S1 specifically includes:

[0021] Collect multi-source spatiotemporal data of the mining area; the multi-source spatiotemporal data includes surface deformation observation data, mining disturbance data, underground rock strata structure data, and environmental hydrological data;

[0022] The multi-source spatiotemporal data is time-aligned using a unified timestamp to form a time index with a unified time step.

[0023] The multi-source spatiotemporal data is spatially aligned using a mining area spatial grid and underground layered structure to form a spatial index with fixed spatial division.

[0024] Based on multi-source spatiotemporal data that has been aligned in time and space, subsidence sequences, disturbance sequences, geological sequences, and hydrological sequences are constructed to generate a multi-source spatiotemporal sample set; the multi-source spatiotemporal sample set satisfies a fixed time step and a fixed spatial index.

[0025] Optionally, S2 specifically includes:

[0026] The time index and spatial index are respectively associated with the subsidence sequence, disturbance sequence, geological sequence and hydrological sequence to generate multi-channel spatiotemporal features; the multi-channel spatiotemporal features include subsidence amount data, disturbance intensity data, geological attribute data and hydrological status data;

[0027] Based on the order of the time index and the order of the spatial index, the multi-channel spatiotemporal features are organized in three dimensions: time dimension, spatial dimension, and feature dimension, to form a spatiotemporal feature tensor; the first dimension of the spatiotemporal feature tensor is the time index, the second dimension is the spatial index, and the third dimension is the feature channel.

[0028] A linear mapping process is performed on the spatiotemporal feature tensor, and a unified dimension mapping is performed on the third-dimensional feature channel to map all feature channels into vectors of the same dimension, while maintaining the original order of time index and spatial index, to generate an encoded feature tensor; the first dimension of the encoded feature tensor is the time index, the second dimension is the spatial index, and the third dimension is the unified vector dimension.

[0029] Optionally, the improved Earthformer model includes an input encoding structure, a cube partitioning structure, a multi-layer spatiotemporal processing structure, and an output structure, and introduces a causal-sensitive cube attention structure, specifically:

[0030] The spatiotemporal feature tensor is vector encoded using an input encoding structure, and positional encoding is performed on the time index and spatial index to generate linear encoded vectors and positional encoded vectors. The linear encoded vectors and positional encoded vectors are then weighted and merged to generate encoded feature vectors.

[0031] The cube partitioning structure performs cube partitioning processing on the encoded feature vector using a preset time window length and a preset spatial block size. Continuous time indices are grouped into multiple time segments according to the preset time window length, and adjacent spatial indices are grouped into multiple spatial segments according to the preset spatial block size. Multiple time segments and multiple spatial segments are combined using a Cartesian combination method to generate multiple cube units. A unique cube index is generated for each cube unit, and the corresponding time range and spatial range of the cube unit are recorded.

[0032] The improved Earthformer model introduces a causal-sensitive cube attention structure, which includes: extracting encoded vectors corresponding to all temporal and spatial positions within a cube cell from the encoded feature tensor according to the cube index; generating query vectors, key vectors, and value vectors from each encoded vector to form query vector sequences, key vector sequences, and value vector sequences. The length of each sequence is consistent with the product of the number of temporal and spatial positions contained in the cube cell, and they are organized according to the order of the temporal and spatial dimensions; calculating matching causal weights based on the temporal and spatial ranges corresponding to the cube cells to form a cube causal weight sequence; weighting the corresponding positions of the cube causal weight sequence and the key vector sequence element-wise to generate a causal weighted key vector sequence; calculating attention coefficients using the query vector sequence and the causal weighted key vector sequence, and performing a weighted summation on the attention coefficients and the value vector sequence to generate causal weighted cube features.

[0033] A multi-layered spatiotemporal processing structure is used to perform spatiotemporal feature extraction processing on the causal weighted cube features in a hierarchical order, including a spatiotemporal interaction sub-layer, a feedforward transformation sub-layer, and a residual normalization sub-layer, to form hierarchical cube features;

[0034] The output structure is used to perform regression mapping on the hierarchical cube features to generate a sinking prediction sequence. The hierarchical cube features are then collected and combined to generate spatiotemporal latent state features.

[0035] Optionally, the step of calculating matching causal weights based on the time and spatial ranges corresponding to the cube units to form a cube causal weight sequence specifically involves:

[0036] The set of coverage time indices is determined based on the time range corresponding to the cube unit, and the set of coverage space indices is determined based on the spatial range corresponding to the cube unit.

[0037] Based on the coverage time index set and the coverage space index set, subsidence data, disturbance intensity data, geological attribute data and hydrological status data corresponding to each combination of time index and each space index are selected from the multi-channel spatiotemporal features to form a cubic candidate feature set;

[0038] Calculate the spatial influence factor for the spatial distance between each spatial index location and the center location of the cubic unit's spatial range;

[0039] Calculate the geological and hydrological changes in the geological attribute data and hydrological status data within the coverage time index set and coverage spatial index set between adjacent time index positions, and generate attribute sensitivity factors based on the geological and hydrological changes.

[0040] A weighted combination process is performed on the time influence factor, spatial influence factor, and attribute sensitivity factor to obtain the initial causal weight set covering the time index set and the spatial index set;

[0041] Normalization is performed on the initial causal weight set to ensure that all causal weight values ​​are within a preset range, thus forming a causal weight set within the cube cell.

[0042] Based on the order of the time index and spatial index in the encoded feature tensor within the cube cell, the causal weight values ​​in the causal weight set within the cube cell are sequentially arranged to form a cube causal weight sequence.

[0043] Optionally, S4 specifically includes:

[0044] The subsidence prediction sequence is reversed according to the time index, and the subsidence prediction sequence is rearranged in the time dimension according to the time index from back to front to generate the reverse subsidence sequence.

[0045] Alignment processing is performed on the reverse sinking sequence and spatiotemporal hidden state features according to the time index and spatial index. A correspondence between the reverse sinking value and the corresponding spatiotemporal hidden state feature is established at each time index and each spatial index position to form the causal analysis input sequence.

[0046] The input sequence for causal analysis is processed by feature concatenation, which concatenates the reverse sinking values ​​and spatiotemporal latent state features along the feature dimension to form a joint feature sequence containing sinking information and latent state information.

[0047] Temporal correlation modeling is performed on the joint feature sequence according to the time index order to construct the dependency relationship between different moments in the time dimension, thereby obtaining the temporal correlation feature sequence that characterizes the subsidence evolution trend.

[0048] Spatial neighborhood modeling is performed on the joint feature sequence according to the spatial index position. The spatial neighborhood set is determined based on the proximity relationship between spatial indices. The statistical features of the joint feature sequence are calculated in each spatial neighborhood to obtain the spatial correlation feature sequence that characterizes the spatial distribution difference.

[0049] The temporal correlation feature sequence and the spatial correlation feature sequence are paired according to their time index and spatial index, and the temporal correlation feature and spatial correlation feature at the corresponding positions are weighted and merged to generate the causal correlation feature sequence.

[0050] Based on the numerical relationship between the components related to disturbance change and the components related to subsidence change in the causal correlation feature sequence, the disturbance intensity estimate and causal contribution value corresponding to each time index and each spatial index position are calculated to form the disturbance intensity estimate sequence and the causal contribution sequence.

[0051] Tensor organization processing is performed on the disturbance intensity estimation sequence and the causal contribution sequence according to the time index and spatial index. The disturbance intensity estimation value and the causal contribution value are arranged according to the time dimension and the spatial dimension to construct a three-dimensional structure. The first dimension is set as the time index, the second dimension is set as the spatial index, and the third dimension is set as the feature dimension composed of the disturbance intensity estimation value and the causal contribution value, forming the causal link tensor of the mining area subsidence.

[0052] Optionally, S5 specifically includes:

[0053] Alignment processing is performed on the subsidence prediction sequence and subsidence data based on time index and spatial index. A correspondence between the subsidence prediction value and the subsidence observation value is established at each time index and each spatial index position to construct the subsidence reconstruction target sequence.

[0054] The numerical differences between the subsidence prediction sequence and the subsidence reconstruction target sequence at each time index and each spatial index position are calculated to generate a subsidence reconstruction deviation sequence. Weighted aggregation processing is performed on the subsidence reconstruction deviation sequence in the time dimension and the spatial dimension to obtain the subsidence reconstruction consistency loss term.

[0055] Numerical combination processing is performed on the disturbance intensity estimation sequence and the causal contribution sequence at each time index and each spatial index position. The disturbance intensity estimation value and the causal contribution value are multiplied to generate the causal-driven subsidence estimation sequence.

[0056] Alignment processing is performed on the causal-driven subsidence estimation sequence and the subsidence prediction sequence based on the time index and spatial index. The numerical difference is calculated at the corresponding position to generate the causal link deviation sequence. The causal link deviation sequence is weighted and aggregated in the time dimension and spatial dimension to obtain the causal link consistency loss term.

[0057] A weighted summation process is performed on the sinking reconstruction consistency loss term and the causal link consistency loss term to generate a joint loss scalar.

[0058] Gradient calculation is performed on all trainable parameters in the improved Earthformer model and causal link network based on the joint loss scalar, and numerical update is performed on each trainable parameter. The trainable parameters are cyclically updated based on the change of the joint loss scalar in consecutive training rounds.

[0059] Optionally, S6 specifically includes:

[0060] A threshold comparison process is performed on the disturbance contribution sequence at each time index and each spatial index position. Positions in the disturbance contribution sequence with values ​​greater than a preset disturbance threshold are marked as spatial unit subsidence source points, forming a subsidence source point marking matrix.

[0061] Based on the subsidence source point marking matrix, continuous time segments are identified in the time dimension, and time positions that are continuous in the time index and have valid subsidence source point markings are grouped into time segments.

[0062] Based on the subsidence source point marker matrix, spatial neighborhood sets are identified in the spatial dimension, and the valid spatial locations of subsidence source point markers are merged into spatially connected regions based on the proximity relationship between spatial indices.

[0063] Perform a correspondence matching process between time segments and spatially connected regions. Retrieve the corresponding spatially connected region at the time index position covered by the time segment. Record the time index and spatial index combination that satisfies the simultaneous validity of the time segment and the spatially connected region as the spatiotemporal sinking source point segment.

[0064] Link connection processing is performed on adjacent spatiotemporal subsidence source point segments based on temporal index continuity and spatial index proximity, and spatiotemporal subsidence source point segments that meet the connection conditions are merged into a spatiotemporal evolution link set.

[0065] The locations of all subsidence source points in the subsidence source point marker matrix are summarized according to the time index and spatial index. After removing duplicates, a subsidence source point set is formed. Together with the spatiotemporal evolution link set, it constitutes the subsidence source area identification result.

[0066] Optionally, S7 specifically includes:

[0067] Based on the subsidence prediction sequence, the set of subsidence source points and the set of spatiotemporal evolution links, a correspondence is established in the time and space dimensions. Subsidence prediction values, source point markers and link markers are extracted at each time index and each space index position to form a risk analysis data sequence.

[0068] A weighted combination process is performed on the risk analysis data sequence at each time index and each spatial index position to combine the subsidence prediction value, source point marker and link marker into a risk score sequence.

[0069] The risk scoring sequence is subjected to a level classification process. The values ​​in the risk scoring sequence are divided into multiple level intervals according to a preset level threshold. The risk level corresponding to each position is marked in the time and space dimensions according to the level intervals, forming a risk level matrix.

[0070] Based on the risk level matrix, a grid mapping process is performed in the spatial dimension to map the risk level of each spatial index position in the risk level matrix to the corresponding spatial grid cell of the mining area, thus forming a subsidence risk level map.

[0071] Based on the subsidence risk level map, the level values, level change trends and spatial location parameters are extracted in the time and space dimensions to form subsidence early warning information.

[0072] The beneficial effects of this invention are:

[0073] (1) By performing time alignment and spatial alignment processing on multi-source spatiotemporal data of mining area, a multi-source spatiotemporal sample set with fixed time step and fixed spatial index is constructed. After forming spatiotemporal feature tensors based on multi-channel spatiotemporal features, vector dimension mapping processing is performed to generate coded feature tensors, so that multi-source spatiotemporal data can obtain a unified structured expression, and significant improvements are made in the accuracy of subsidence feature construction and cross-source data consistency.

[0074] (2) Input the encoded feature tensor into the improved Earthformer model with causal sensitive cube attention structure, generate subsidence prediction sequence and spatiotemporal latent state features through cube partitioning and causal weighting, and construct an effective coupling relationship between time index and spatial index in the spatiotemporal dimension, so as to enhance the temporal correlation and spatial distribution sensitivity of subsidence prediction, and obtain higher accuracy in subsidence development trend characterization and spatial deformation response modeling.

[0075] (3) Construct a reverse subsidence sequence based on the subsidence prediction sequence and spatiotemporal latent state characteristics, and establish causal relationship feature expression in the time and space dimensions to generate a disturbance intensity estimation sequence and a causal contribution sequence, forming a mining area subsidence causal link tensor, so that the driving relationship between disturbance change and subsidence change can be quantitatively expressed, and an interpretable causal link structure can be formed in terms of subsidence cause identification, disturbance impact location and source area tracing.

[0076] (4) Based on the causal link tensor of the mining area subsidence, the subsidence source point is identified and a spatiotemporal evolution link set is constructed. The subsidence risk level map and structured subsidence early warning information are generated through risk scoring and level classification, making the subsidence source area identification more refined, the subsidence evolution path clearer, and the risk presentation more structured, thus significantly improving the reliability of subsidence risk early warning and the usability of risk expression. Attached Figure Description

[0077] 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:

[0078] Figure 1 This is a schematic diagram of a deep learning-based surface subsidence risk analysis system for mining areas proposed in this invention.

[0079] Figure 2 This is a flowchart of a deep learning-based surface subsidence risk analysis system for mining areas proposed in this invention.

[0080] Figure 3 This is a schematic diagram of the structure of the improved Earthformer model proposed in this invention. Detailed Implementation

[0081] 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.

[0082] refer to Figure 1 A deep learning-based system for analyzing the risk of surface subsidence in mining areas, comprising:

[0083] The data construction module is used to collect multi-source spatiotemporal data from the mining area, perform time alignment and spatial alignment processing to form a multi-source spatiotemporal sample set; construct multi-channel spatiotemporal features to form a spatiotemporal feature tensor, and perform vector dimension mapping processing on the spatiotemporal feature tensor to generate an encoded feature tensor;

[0084] The prediction and inference module is used to input the encoded feature tensor into the improved Earthformer model. The improved Earthformer model introduces a causal sensitive cube attention structure, performs cube partitioning and causal weighting, and generates a sinking prediction sequence and spatiotemporal latent state features.

[0085] The causal analysis module is used to construct a reverse subsidence sequence based on the subsidence prediction sequence and spatiotemporal latent state characteristics, and to generate a disturbance intensity estimation sequence and a causal contribution sequence through a causal link network, thus forming a causal link tensor for subsidence in the mining area.

[0086] The training and update module is used to construct consistency constraints based on the subsidence prediction sequence, the disturbance intensity estimation sequence, and the causal contribution sequence, perform subsidence reconstruction consistency training and causal link consistency training, and update the parameters of the improved Earthformer model and the causal link network according to the joint loss.

[0087] The risk assessment module is used to identify spatial units whose disturbance contribution is greater than a preset disturbance threshold as subsidence source points based on the causal link tensor of the subsidence in the mining area. It performs link combination processing on continuous time segments and adjacent spatial units to generate subsidence source area identification results; calculates risk scores, classifies them into levels, and outputs subsidence risk level maps and structured subsidence early warning information.

[0088] In this implementation, the data construction module collects multi-source spatiotemporal data from the mining area and performs time and spatial alignment processing to form a multi-source spatiotemporal sample set with a unified structure. This ensures that subsidence data, disturbance data, geological data, and hydrological data are consistent in both time and spatial dimensions, improving the input standardization of subsequent modeling stages. It also constructs multi-channel spatiotemporal features and generates coded feature tensors, allowing subsidence evolution information to be expressed within a unified vector dimension, thus enhancing the deep model's ability to jointly analyze multi-source heterogeneous information. The prediction and inference module inputs the coded feature tensors into the improved Earthformer model, introducing a causal-sensitive cube attention structure to perform spatiotemporal region partitioning and causal weighting. This enables the model to focus on the dominant factors of subsidence evolution within local spatiotemporal regions, improving the accuracy of subsidence prediction sequences and the completeness of spatiotemporal latent state feature expression. The causal analysis module constructs a reverse subsidence sequence based on the subsidence prediction sequence and spatiotemporal latent state characteristics. It generates a disturbance intensity estimation sequence and a causal contribution sequence through a causal link network, forming a causal link tensor for mining area subsidence. This clearly characterizes the relationship between subsidence changes and disturbance sources in both time and space dimensions, improving the traceability of subsidence causal analysis. The training and update module constructs subsidence reconstruction consistency constraints and causal link consistency constraints, updating model parameters to simultaneously improve subsidence prediction accuracy and causal inference reliability, further enhancing model stability in complex geological environments. The risk assessment module identifies spatial units with disturbance contributions exceeding a preset disturbance threshold and constructs subsidence source area identification results. It performs link combination processing on continuous time segments and adjacent spatial units, making the subsidence source area present a coherent structure in the spatiotemporal dimensions. Simultaneously, it calculates risk scores and performs level classification, outputting a subsidence risk level map and structured subsidence early warning information. This makes the expression of subsidence risk more readable, interpretable, and applicable, achieving full-process quantitative analysis and early warning output of mining area subsidence risk.

[0089] refer to Figure 2 In this embodiment, the modules are interconnected through the following method:

[0090] S1. Collect multi-source spatiotemporal data from the mining area, perform time alignment and spatial alignment processing, and form a multi-source spatiotemporal sample set;

[0091] S2. Construct multi-channel spatiotemporal features based on multi-source spatiotemporal sample sets to form a spatiotemporal feature tensor, and perform vector dimension mapping processing on the spatiotemporal feature tensor to generate an encoded feature tensor;

[0092] S3. Input the encoded feature tensor into the improved Earthformer model. The improved Earthformer model introduces a causal sensitive cube attention structure, performs cube partitioning and causal weighting, and generates a sinking prediction sequence and spatiotemporal hidden state features.

[0093] S4. Construct an inverse subsidence sequence based on the subsidence prediction sequence and spatiotemporal latent state characteristics. Generate a disturbance intensity estimation sequence and a causal contribution sequence through a causal link network to form a mining area subsidence causal link tensor.

[0094] S5. Based on the subsidence prediction sequence, disturbance intensity estimation sequence and causal contribution sequence, construct consistency constraints, perform subsidence reconstruction consistency training and causal link consistency training, and update the parameters of the improved Earthformer model and causal link network according to the joint loss.

[0095] S6. Based on the spatial unit whose disturbance contribution is greater than the preset disturbance threshold identified by the causal link tensor of the subsidence in the mining area, subsidence source points are identified. Link combination processing is performed on continuous time segments and adjacent spatial units to generate subsidence source area identification results.

[0096] S7. Calculate the risk score based on the subsidence source area identification results, classify the risk levels, and output the subsidence risk level map and structured subsidence early warning information.

[0097] In this embodiment, S1 specifically refers to:

[0098] Collect multi-source spatiotemporal data of the mining area; the multi-source spatiotemporal data includes surface deformation observation data, mining disturbance data, underground rock strata structure data, and environmental hydrological data;

[0099] The multi-source spatiotemporal data is time-aligned using a unified timestamp to form a time index with a unified time step.

[0100] The multi-source spatiotemporal data is spatially aligned using a mining area spatial grid and an underground stratified structure to form a spatial index with a fixed spatial division. The mining area spatial grid represents a set of two-dimensional grid cells obtained by dividing the surface area of ​​the mining area according to a preset spatial resolution, with each two-dimensional grid cell having a unique spatial index. The underground stratified structure represents a set of multi-layered structures obtained by dividing the underground area of ​​the mining area according to a preset depth level, with each depth level having a unique level index, and a spatial correspondence is established with the two-dimensional grid cells.

[0101] Based on multi-source spatiotemporal data that has been aligned in time and space, subsidence sequences, disturbance sequences, geological sequences, and hydrological sequences are constructed to generate a multi-source spatiotemporal sample set; the multi-source spatiotemporal sample set satisfies a fixed time step and a fixed spatial index.

[0102] In this embodiment, S2 specifically refers to:

[0103] The time index and spatial index are respectively associated with the subsidence sequence, disturbance sequence, geological sequence and hydrological sequence to generate multi-channel spatiotemporal features; the multi-channel spatiotemporal features include subsidence amount data, disturbance intensity data, geological attribute data and hydrological status data;

[0104] Based on the order of the time index and the order of the spatial index, the multi-channel spatiotemporal features are organized in three dimensions: time dimension, spatial dimension, and feature dimension, to form a spatiotemporal feature tensor; the first dimension of the spatiotemporal feature tensor is the time index, the second dimension is the spatial index, and the third dimension is the feature channel.

[0105] A linear mapping process is performed on the spatiotemporal feature tensor, and a unified dimension mapping is performed on the third-dimensional feature channel to map all feature channels into vectors of the same dimension, while maintaining the original order of time index and spatial index, to generate an encoded feature tensor; the first dimension of the encoded feature tensor is the time index, the second dimension is the spatial index, and the third dimension is the unified vector dimension.

[0106] refer to Figure 3 In this embodiment, the improved Earthformer model includes an input encoding structure, a cube partitioning structure, a multi-layer spatiotemporal processing structure, and an output structure, and introduces a causal-sensitive cube attention structure, specifically:

[0107] An input encoding structure is used to perform vector encoding processing on the spatiotemporal feature tensor, projecting the feature channel values ​​corresponding to each time index and each spatial index in the spatiotemporal feature tensor into an encoded vector of a unified vector dimension. Position encoding processing is performed on the time index and spatial index, constructing position sequences by constructing position sequences from the time index sequence and the spatial index sequence respectively, and performing sine and cosine position transformation processing on the position sequences to form a position encoded vector sequence. Element-wise weighted merging processing is performed on the linear encoded vector and the position encoded vector, performing element-wise addition or element-wise multiplication operations on the linear encoded vector and the position encoded vector at corresponding positions according to preset coefficients to generate encoded feature vectors. This enables multi-source spatiotemporal features to have temporal and spatial position information in a unified vector space, improving the model's ability to express spatiotemporal dependencies.

[0108] The cube partitioning structure performs cube partitioning on the encoded feature vector using a preset time window length and a preset spatial block size. The preset time window length is the length parameter for segmenting according to continuous time indices, and the preset spatial block size is the spatial scale parameter for grouping according to adjacent spatial indices. Multiple time segments are constructed from continuous time indices according to the preset time window length, and multiple spatial segments are constructed from spatial indices according to the preset spatial block size. A Cartesian combination is performed on all time segments and all spatial segments, and each pair of time segments and spatial segments in the combination result constitutes a cube unit. A unique cube index is generated for each cube unit, and the time and spatial ranges covered by each cube unit are recorded. This ensures that the encoded feature vector is divided into structurally regular and well-defined cube units in the spatiotemporal dimension, improving the ability to aggregate local spatiotemporal information.

[0109] Based on the cube index, the encoded vectors corresponding to all time and spatial locations within the cube cell are extracted from the encoded feature tensor. For each encoded vector, a query vector, key vector, and numerical vector are generated. The generation method involves performing three sets of linear mapping processes on the encoded vectors, each forming a target vector dimension that is the same as or different from the input dimension. All query vectors are organized into a query vector sequence according to time index priority and spatial index order; all key vectors are organized into a key vector sequence; and all numerical vectors are organized into a numerical vector sequence. Causal relationship weighting is performed on the time and spatial ranges corresponding to the cube cell. The temporal and spatial causal degrees are obtained by calculating the sequential relationship between time indices and the proximity relationship between spatial indices. The magnitude of changes in geological and hydrological attributes is calculated to obtain attribute sensitivity factors. A weighted combination process is then performed on the temporal causal degree, spatial causal degree, and attribute sensitivity factors. The three types of factors are added element-wise according to their corresponding weights to obtain the initial... A causal weight set is generated. The initial causal weight set is normalized by performing a linear transformation on all causal weight values ​​to minimize and maximize them, ensuring all weights are between zero and one, forming a causal weight set within each cube cell. A causal weight sequence is formed according to the order of the encoded vectors within the cube cell. An element-wise weighted combination is performed on the key vector sequence and the causal weight sequence, multiplying the value at each position of the key vector sequence with its corresponding causal weight to form a causal weighted key vector sequence. Attention coefficients are calculated using the query vector sequence and the causal weighted key vector sequence. These coefficients are calculated by performing a dot product on the query vector and the causal weighted key vector, followed by normalization of all dot product results. A weighted summation is performed on the attention coefficients and the numerical vector sequence, and each numerical vector is linearly combined according to the attention coefficient to generate causal weighted cube features. This process enhances the causal sensitivity of feature extraction within the spatiotemporal region, improving the ability of deep structures to capture the causes of subsidence.

[0110] A multi-layer spatiotemporal processing structure is employed to extract spatiotemporal features from causal weighted cube features in a hierarchical order. Within each spatiotemporal interaction sub-layer, spatiotemporal interaction mapping is performed on the causal weighted cube features, constructing spatiotemporal dependencies by performing sequence mapping operations on the temporal dimension and neighborhood mapping operations on the spatial dimension. Within the feedforward transformation sub-layer, a two-layer vector transformation is performed on the spatiotemporal interaction output, where the first layer expands the vector dimension and the second layer compresses the vector dimension. Within the residual normalization sub-layer, the input features and the feedforward sub-layer output are added element-wise to form residual connections, and the residual results are normalized to form hierarchical cube features, enabling the model to maintain stable gradients and robust spatiotemporal expressive power in the deep structure.

[0111] The output structure performs regression mapping processing on the hierarchical cube features, calculates the subsidence prediction sequence based on the time and spatial indices of the hierarchical cube features, and performs convergence processing on the hierarchical cube features to aggregate the spatiotemporal local features along the time and spatial dimensions to generate spatiotemporal latent state features. This allows the spatiotemporal representation of subsidence prediction and subsidence evolution to be output synchronously, improving the overall prediction stability and spatial anomaly identification capability. The corresponding technical effects are to improve the accuracy of the subsidence prediction sequence, make the expression of spatiotemporal latent state features more complete, and make the subsidence risk analysis more reliable in the spatiotemporal dimension.

[0112] In this embodiment, the step of calculating matching causal weights based on the time and spatial ranges corresponding to the cube units to form a cube causal weight sequence specifically involves:

[0113] The time index set is determined based on the time range corresponding to the cube unit, and all time indices between the start and end time indices of the time range are included in the time index set in order; the spatial index set is determined based on the spatial range corresponding to the cube unit, and all spatial grid indices covered by the spatial range are included in the spatial index set in spatial order.

[0114] Based on the coverage time index set and the coverage space index set, corresponding subsidence data, disturbance intensity data, geological attribute data and hydrological status data are selected from the multi-channel spatiotemporal features to form all feature entries under the two-dimensional combination of coverage time index and coverage space index. All feature entries are then organized into a cube candidate feature set according to the priority of time index and the order of spatial index, so that the cube candidate feature set can completely map all attribute information of the cube unit in the time dimension and spatial dimension.

[0115] For each spatial location within the covered spatial index set and the center of the cubic unit's spatial range, a distance metric calculation is performed. The spatial distance can be obtained by summing the square roots of the squared differences in the three-dimensional coordinates, generating a spatial distance value. For each spatial distance value, a distance decay process is performed by transforming it into a negative exponential function with the distance value as the exponent, so that the closer the distance, the larger the corresponding spatial influence factor, forming a set of spatial influence factors for the spatial location of the cubic unit.

[0116] The geological attribute data and hydrological status data within the coverage time index set and coverage spatial index set are processed by calculating the difference between adjacent time positions. The geological change is obtained by calculating the absolute value of the difference between the geological attribute values ​​at adjacent time index positions, and the hydrological change is obtained by calculating the absolute value of the difference between the hydrological status values ​​at adjacent time index positions. The geological change and hydrological change are then processed by weighted combination. The geological change and hydrological change are added element-wise according to preset coefficients to form a set of attribute sensitive factors, so that the attribute sensitive factors reflect the comprehensive sensitivity of geological and hydrological changes within the cubic cell.

[0117] A weighted combination process is performed on the time influence factor, spatial influence factor, and attribute sensitivity factor. The time influence factor is constructed based on the time difference between the time index within the cube cell and the starting position of the time segment; the spatial influence factor is constructed based on spatial distance; and the attribute sensitivity factor is constructed based on geological and hydrological changes. Time weights, spatial weights, and attribute weights are assigned to each of the three factors respectively. Combination processing is performed through element-wise addition or multiplication to ensure that each time position and each spatial position within the covered time index set and the covered spatial index set corresponds to an initial causal weight value, resulting in an initial causal weight set. This initial causal weight set is not uniform in value and has different magnitudes.

[0118] Normalization is performed on the initial causal weight set, mapping all initial causal weight values ​​to the range of zero to one according to the minimum and maximum value linear transformation steps. The normalization transformation steps include taking the minimum value in the initial causal weight set as the lower bound and the maximum value as the upper bound, performing a linear transformation on each initial weight value by subtracting the minimum value from the value and dividing by the difference between the maximum value and the minimum value above, forming a causal weight set within the cube cell, so that the causal weight values ​​have a uniform scale.

[0119] Based on the order of temporal and spatial indices within the cube cell in the encoded feature tensor, the causal weight values ​​in the causal weight set within the cube cell are sequentially arranged. The sequence of causal weight values ​​organized according to temporal index priority and spatial index order is used as the final cube causal weight sequence. This ensures that the cube causal weight sequence is strictly aligned with the internal structure of the cube cell of the encoded feature tensor, thereby improving the weight expression ability of the causal sensitive cube attention structure for the internal correlation of the cube and the stability of the causal sensitivity of sinking prediction.

[0120] In this embodiment, S4 specifically refers to:

[0121] The subsidence prediction sequence is reversed according to the time index, and the subsidence prediction sequence is rearranged in the time dimension according to the time index from back to front to generate the reverse subsidence sequence.

[0122] Alignment processing is performed on the reverse sinking sequence and spatiotemporal hidden state features according to the time index and spatial index. A correspondence between the reverse sinking value and the corresponding spatiotemporal hidden state feature is established at each time index and each spatial index position to form the causal analysis input sequence.

[0123] The input sequence for causal analysis is processed by feature concatenation, which concatenates the reverse sinking values ​​and spatiotemporal latent state features along the feature dimension to form a joint feature sequence containing sinking information and latent state information.

[0124] Temporal correlation modeling is performed on the joint feature sequence according to the time index order to construct the dependency relationship between different moments in the time dimension, thereby obtaining the temporal correlation feature sequence that characterizes the subsidence evolution trend.

[0125] Spatial neighborhood modeling is performed on the joint feature sequence according to the spatial index position. The spatial neighborhood set is determined based on the proximity relationship between spatial indices. The statistical features of the joint feature sequence are calculated in each spatial neighborhood to obtain the spatial correlation feature sequence that characterizes the spatial distribution difference.

[0126] The temporal correlation feature sequence and the spatial correlation feature sequence are paired according to their time index and spatial index, and the temporal correlation feature and spatial correlation feature at the corresponding positions are weighted and merged to generate the causal correlation feature sequence.

[0127] Based on the numerical relationship between the components related to disturbance change and the components related to subsidence change in the causal correlation feature sequence, the disturbance intensity estimate and causal contribution value corresponding to each time index and each spatial index position are calculated to form the disturbance intensity estimate sequence and the causal contribution sequence.

[0128] Tensor organization processing is performed on the disturbance intensity estimation sequence and the causal contribution sequence according to the time index and spatial index. The disturbance intensity estimation value and the causal contribution value are arranged according to the time dimension and the spatial dimension to construct a three-dimensional structure. The first dimension is set as the time index, the second dimension is set as the spatial index, and the third dimension is set as the feature dimension composed of the disturbance intensity estimation value and the causal contribution value, forming the causal link tensor of the mining area subsidence.

[0129] In this embodiment, S5 specifically refers to:

[0130] Alignment processing is performed on the subsidence prediction sequence and subsidence data based on time index and spatial index. A correspondence between the subsidence prediction value and the subsidence observation value is established at each time index and each spatial index position to construct the subsidence reconstruction target sequence.

[0131] The numerical differences between the subsidence prediction sequence and the subsidence reconstruction target sequence at each time index and each spatial index position are calculated to generate a subsidence reconstruction deviation sequence. Weighted aggregation processing is performed on the subsidence reconstruction deviation sequence in the time dimension and the spatial dimension to obtain the subsidence reconstruction consistency loss term.

[0132] Numerical combination processing is performed on the disturbance intensity estimation sequence and the causal contribution sequence at each time index and each spatial index position. The disturbance intensity estimation value and the causal contribution value are multiplied to generate the causal-driven subsidence estimation sequence.

[0133] Alignment processing is performed on the causal-driven subsidence estimation sequence and the subsidence prediction sequence based on the time index and spatial index. The numerical difference is calculated at the corresponding position to generate the causal link deviation sequence. The causal link deviation sequence is weighted and aggregated in the time dimension and spatial dimension to obtain the causal link consistency loss term.

[0134] A weighted summation process is performed on the sinking reconstruction consistency loss term and the causal link consistency loss term to generate a joint loss scalar.

[0135] Gradient calculation is performed on all trainable parameters in the improved Earthformer model and causal link network based on the joint loss scalar, and numerical update is performed on each trainable parameter. The trainable parameters are cyclically updated based on the change of the joint loss scalar in consecutive training rounds.

[0136] In this embodiment, S6 specifically refers to:

[0137] A threshold comparison process is performed on the disturbance contribution sequence at each time index and each spatial index position. Positions in the disturbance contribution sequence with values ​​greater than a preset disturbance threshold are marked as spatial unit subsidence source points, forming a subsidence source point marking matrix.

[0138] Based on the subsidence source point marking matrix, continuous time segments are identified in the time dimension, and time positions that are continuous in the time index and have valid subsidence source point markings are grouped into time segments.

[0139] Based on the subsidence source point marker matrix, spatial neighborhood sets are identified in the spatial dimension, and the valid spatial locations of subsidence source point markers are merged into spatially connected regions based on the proximity relationship between spatial indices.

[0140] Perform a correspondence matching process between time segments and spatially connected regions. Retrieve the corresponding spatially connected region at the time index position covered by the time segment. Record the time index and spatial index combination that satisfies the simultaneous validity of the time segment and the spatially connected region as the spatiotemporal sinking source point segment.

[0141] Link connection processing is performed on adjacent spatiotemporal subsidence source point segments based on temporal index continuity and spatial index proximity, and spatiotemporal subsidence source point segments that meet the connection conditions are merged into a spatiotemporal evolution link set.

[0142] The locations of all subsidence source points in the subsidence source point marker matrix are summarized according to the time index and spatial index. After removing duplicates, a subsidence source point set is formed. Together with the spatiotemporal evolution link set, it constitutes the subsidence source area identification result.

[0143] In this embodiment, S7 specifically refers to:

[0144] Based on the subsidence prediction sequence, the set of subsidence source points and the set of spatiotemporal evolution links, a correspondence is established in the time and space dimensions. Subsidence prediction values, source point markers and link markers are extracted at each time index and each space index position to form a risk analysis data sequence.

[0145] A weighted combination process is performed on the risk analysis data sequence at each time index and each spatial index position to combine the subsidence prediction value, source point marker and link marker into a risk score sequence.

[0146] The risk scoring sequence is subjected to a level classification process. The values ​​in the risk scoring sequence are divided into multiple level intervals according to a preset level threshold. The risk level corresponding to each position is marked in the time and space dimensions according to the level intervals, forming a risk level matrix.

[0147] Based on the risk level matrix, a grid mapping process is performed in the spatial dimension to map the risk level of each spatial index position in the risk level matrix to the corresponding spatial grid cell of the mining area, thus forming a subsidence risk level map.

[0148] Based on the subsidence risk level map, the level values, level change trends and spatial location parameters are extracted in the time and space dimensions to form subsidence early warning information.

[0149] Example 1:

[0150] To verify the feasibility of this invention in practice, it was applied to the East Second District of a large coal mine. This area is buried at a depth of 150–320 meters, with strata consisting of interbedded sandstone, siltstone, and mudstone, and in contact with local aquifers. Over the past five years, this mining area has experienced several irregular subsidences, particularly above the goaf and near transportation roads, where sudden increases in localized subsidence have significantly impacted the safety of mine buildings and equipment. Existing subsidence analysis methods rely on empirical formulas and single monitoring curves, resulting in insufficient spatial resolution, difficulty in timely identification of subsidence source areas, and a high false alarm rate, significantly increasing the workload of inspection teams.

[0151] In this embodiment, surface monitoring stations, InSAR observation products, mining activity recording systems, geological model databases, and hydrological observation well networks deployed in the mining area are connected through a unified data interface. The system collects subsidence data, disturbance intensity data, geological attribute data, and hydrological status data daily at fixed time steps, aligns them according to a unified timestamp, and establishes a spatial index using a regular grid within the mining area, forming a three-month multi-source spatiotemporal sample set. This multi-source spatiotemporal sample set is processed to construct multi-channel spatiotemporal features, further organized into spatiotemporal feature tensors, and then processed by linear mapping to generate coded feature tensors. These coded feature tensors serve as unified data inputs to an improved Earthformer model deployed on edge computing nodes within the mining area.

[0152] The model internally performs positional encoding on the time and spatial indices and performs dimensional unification on the vector channels. Within the mining area, the encoded feature tensor is divided into a large number of cubic units according to a preset window. Time, space, and attribute data are extracted from each cubic unit, and the time influence factor, spatial influence factor, and attribute sensitivity factor are calculated. These are then weighted to form a cubic causal weight sequence. The cubic causal weight sequence is used to perform weighted processing on the key vector, thereby generating causal weighted cubic features. Finally, the sinking prediction sequence and spatiotemporal latent state features are output.

[0153] In the causal analysis phase, the system reverses the subsidence prediction sequence to generate a reverse subsidence sequence. This reverse subsidence sequence is then aligned and spatiotemporally latent features to generate a joint feature sequence. The system identifies subsidence evolution relationships in the temporal dimension and neighborhood differences in the spatial dimension; a weighted average of these two identifications yields a causal association feature sequence. Based on the numerical relationship between disturbance changes and subsidence changes, the system calculates disturbance intensity estimates and causal contribution values ​​for each time and spatial index, organizing them into a mining area subsidence causal link tensor. The system performs threshold judgment on the causal contribution sequence, identifying spatial units exceeding the threshold as subsidence source points, and combines temporal continuity and spatial proximity to form subsidence source area identification results. Finally, based on the subsidence source area identification results, a risk score is generated, risk levels are classified, and subsidence early warning information is output.

[0154] To verify the effectiveness of this system in real-world scenarios, a three-month monitoring period was used as the test window to compare the prediction errors and early warning results of traditional methods and the method of this invention.

[0155] Table 1 Comparison of Prediction Accuracy for Mining Area Subsidence

[0156]

[0157] As can be observed from the comparison results in Table 1, the average absolute error of the method of this invention remained around 3 cm over three calendar months, a stable reduction of approximately 55% to 65% compared to the traditional method. Regarding the maximum error, the method of this invention controlled the error to within 12 cm, while the traditional method often exhibited significant deviations exceeding 20 cm above localized goaf areas. The recall rate of the method of this invention in high-risk areas consistently exceeded 92%, an average improvement of over 13 percentage points compared to the traditional method. This indicates that the subsidence prediction sequence generated by the system based on the causal sensitive cube attention structure can more accurately track the subsidence evolution trend of disturbed areas, especially in locations with drastic changes in geological attributes and hydrological conditions. The system can improve prediction responsiveness by introducing attribute-sensitive factors. Furthermore, in several boundary areas with large subsidence gradients, the prediction results generated by the method of this invention are closer to the high-precision InSAR inversion results, with smoother spatial transitions, resulting in significantly more accurate identification of high-risk areas. This demonstrates the robustness of the improved Earthformer model and causal weighting mechanism in multi-source disturbance environments.

[0158] Table 2 Comparison of Subsidence Risk Warning and Inspection Workload

[0159]

[0160] Table 2 shows that the method of this invention provided early warnings for 27 subsidence events within 3 months, while the traditional method only provided early warnings for 17. The method of this invention had only 3 missed warnings, while the traditional method had 13, showing a significant difference in the missed warning rate. Regarding false alarms, the method of this invention had 42, while the traditional method had 88, reducing the number of false alarms by about half. The reduction in false alarms means that inspection routes no longer need to cover a large number of non-risk areas, allowing inspection personnel to focus their efforts on areas where spatiotemporal evolution links truly exist. In terms of inspection workload, the traditional method averaged approximately 57 inspection points per day, while the method of this invention reduced this to approximately 40, a reduction of about 30%, significantly shortening the inspection route length and reducing operational pressure. From an engineering perspective, the method of this invention identifies subsidence source points through causal links, effectively pinpointing the areas with the greatest disturbance contribution, making the early warning range more focused and the risk direction clearer. It maintains stable performance in continuous prediction tasks, demonstrating the practical application value of this system in mine safety production management.

[0161] 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 deep learning-based system for analyzing the risk of surface subsidence in mining areas, characterized in that, include: The data construction module is used to collect multi-source spatiotemporal data from the mining area, perform time alignment and spatial alignment processing to form a multi-source spatiotemporal sample set; construct multi-channel spatiotemporal features to form a spatiotemporal feature tensor, and perform vector dimension mapping processing on the spatiotemporal feature tensor to generate an encoded feature tensor; The prediction and inference module is used to input the encoded feature tensor into the improved Earthformer model. The improved Earthformer model introduces a causal sensitive cube attention structure, performs cube partitioning and causal weighting, and generates a sinking prediction sequence and spatiotemporal latent state features. The causal analysis module is used to construct a reverse subsidence sequence based on the subsidence prediction sequence and spatiotemporal latent state characteristics, and to generate a disturbance intensity estimation sequence and a causal contribution sequence through a causal link network, thus forming a causal link tensor for subsidence in the mining area. The training and update module is used to construct consistency constraints based on the subsidence prediction sequence, the disturbance intensity estimation sequence, and the causal contribution sequence, perform subsidence reconstruction consistency training and causal link consistency training, and update the parameters of the improved Earthformer model and the causal link network according to the joint loss. The risk assessment module is used to identify spatial units whose disturbance contribution is greater than a preset disturbance threshold as subsidence source points based on the causal link tensor of the subsidence in the mining area. It performs link combination processing on continuous time segments and adjacent spatial units to generate subsidence source area identification results; calculates risk scores, classifies them into levels, and outputs subsidence risk level maps and structured subsidence early warning information.

2. The deep learning-based surface subsidence risk analysis system for mining areas according to claim 1, characterized in that, Inter-module communication is achieved through the following methods: S1. Collect multi-source spatiotemporal data from the mining area, perform time alignment and spatial alignment processing, and form a multi-source spatiotemporal sample set; S2. Construct multi-channel spatiotemporal features based on multi-source spatiotemporal sample sets to form a spatiotemporal feature tensor, and perform vector dimension mapping processing on the spatiotemporal feature tensor to generate an encoded feature tensor; S3. Input the encoded feature tensor into the improved Earthformer model. The improved Earthformer model introduces a causal sensitive cube attention structure, performs cube partitioning and causal weighting, and generates a sinking prediction sequence and spatiotemporal hidden state features. S4. Construct an inverse subsidence sequence based on the subsidence prediction sequence and spatiotemporal latent state characteristics. Generate a disturbance intensity estimation sequence and a causal contribution sequence through a causal link network to form a mining area subsidence causal link tensor. S5. Based on the subsidence prediction sequence, disturbance intensity estimation sequence and causal contribution sequence, construct consistency constraints, perform subsidence reconstruction consistency training and causal link consistency training, and update the parameters of the improved Earthformer model and causal link network according to the joint loss. S6. Based on the spatial unit whose disturbance contribution is greater than the preset disturbance threshold identified by the causal link tensor of the subsidence in the mining area, subsidence source points are identified. Link combination processing is performed on continuous time segments and adjacent spatial units to generate subsidence source area identification results. S7. Calculate the risk score based on the subsidence source area identification results, classify the risk levels, and output the subsidence risk level map and structured subsidence early warning information.

3. The deep learning-based surface subsidence risk analysis system for mining areas according to claim 2, characterized in that, Specifically, S1 is: Collect multi-source spatiotemporal data of the mining area; the multi-source spatiotemporal data includes surface deformation observation data, mining disturbance data, underground rock strata structure data, and environmental hydrological data; The multi-source spatiotemporal data is time-aligned using a unified timestamp to form a time index with a unified time step. The multi-source spatiotemporal data is spatially aligned using a mining area spatial grid and underground layered structure to form a spatial index with fixed spatial division. Based on multi-source spatiotemporal data that has been aligned in time and space, subsidence sequences, disturbance sequences, geological sequences, and hydrological sequences are constructed to generate a multi-source spatiotemporal sample set; the multi-source spatiotemporal sample set satisfies a fixed time step and a fixed spatial index.

4. The deep learning-based surface subsidence risk analysis system for mining areas according to claim 2, characterized in that, Specifically, S2 is: The time index and spatial index are respectively associated with the subsidence sequence, disturbance sequence, geological sequence and hydrological sequence to generate multi-channel spatiotemporal features; the multi-channel spatiotemporal features include subsidence amount data, disturbance intensity data, geological attribute data and hydrological status data; Based on the order of the time index and the order of the spatial index, the multi-channel spatiotemporal features are organized in three dimensions: time dimension, spatial dimension, and feature dimension, to form a spatiotemporal feature tensor; the first dimension of the spatiotemporal feature tensor is the time index, the second dimension is the spatial index, and the third dimension is the feature channel. A linear mapping process is performed on the spatiotemporal feature tensor, and a unified dimension mapping is performed on the third-dimensional feature channel to map all feature channels into vectors of the same dimension, while maintaining the original order of time index and spatial index, to generate an encoded feature tensor; the first dimension of the encoded feature tensor is the time index, the second dimension is the spatial index, and the third dimension is the unified vector dimension.

5. The deep learning-based surface subsidence risk analysis system for mining areas according to claim 2, characterized in that, The improved Earthformer model includes an input encoding structure, a cube partitioning structure, a multi-layer spatiotemporal processing structure, and an output structure. It also introduces a causal-sensitive cube attention structure, specifically: The spatiotemporal feature tensor is vector encoded using an input encoding structure, and positional encoding is performed on the time index and spatial index to generate linear encoded vectors and positional encoded vectors. The linear encoded vectors and positional encoded vectors are then weighted and merged to generate encoded feature vectors. The cube partitioning structure performs cube partitioning processing on the encoded feature vector using a preset time window length and a preset spatial block size. Continuous time indices are grouped into multiple time segments according to the preset time window length, and adjacent spatial indices are grouped into multiple spatial segments according to the preset spatial block size. Multiple time segments and multiple spatial segments are combined using a Cartesian combination method to generate multiple cube units. A unique cube index is generated for each cube unit, and the corresponding time range and spatial range of the cube unit are recorded. The improved Earthformer model introduces a causal-sensitive cube attention structure, which includes: extracting encoded vectors corresponding to all temporal and spatial positions within a cube cell from the encoded feature tensor according to the cube index; generating query vectors, key vectors, and value vectors from each encoded vector to form query vector sequences, key vector sequences, and value vector sequences. The length of each sequence is consistent with the product of the number of temporal and spatial positions contained in the cube cell, and they are organized according to the order of the temporal and spatial dimensions; calculating matching causal weights based on the temporal and spatial ranges corresponding to the cube cells to form a cube causal weight sequence; weighting the corresponding positions of the cube causal weight sequence and the key vector sequence element-wise to generate a causal weighted key vector sequence; calculating attention coefficients using the query vector sequence and the causal weighted key vector sequence, and performing a weighted summation on the attention coefficients and the value vector sequence to generate causal weighted cube features. A multi-layered spatiotemporal processing structure is used to perform spatiotemporal feature extraction processing on the causal weighted cube features in a hierarchical order, including a spatiotemporal interaction sub-layer, a feedforward transformation sub-layer, and a residual normalization sub-layer, to form hierarchical cube features; The output structure is used to perform regression mapping on the hierarchical cube features to generate a sinking prediction sequence. The hierarchical cube features are then collected and combined to generate spatiotemporal latent state features.

6. The deep learning-based surface subsidence risk analysis system for mining areas according to claim 5, characterized in that, The step of calculating matching causal weights based on the time and spatial ranges corresponding to the cube units to form a cube causal weight sequence is as follows: The set of coverage time indices is determined based on the time range corresponding to the cube unit, and the set of coverage space indices is determined based on the spatial range corresponding to the cube unit. Based on the coverage time index set and the coverage space index set, subsidence data, disturbance intensity data, geological attribute data and hydrological status data corresponding to each combination of time index and each space index are selected from the multi-channel spatiotemporal features to form a cubic candidate feature set; Calculate the spatial influence factor for the spatial distance between each spatial index location and the center location of the cubic unit's spatial range; Calculate the geological and hydrological changes in the geological attribute data and hydrological status data within the coverage time index set and coverage spatial index set between adjacent time index positions, and generate attribute sensitivity factors based on the geological and hydrological changes. A weighted combination process is performed on the time influence factor, spatial influence factor, and attribute sensitivity factor to obtain the initial causal weight set covering the time index set and the spatial index set; Normalization is performed on the initial causal weight set to ensure that all causal weight values ​​are within a preset range, thus forming a causal weight set within the cube cell. Based on the order of the time index and spatial index in the encoded feature tensor within the cube cell, the causal weight values ​​in the causal weight set within the cube cell are sequentially arranged to form a cube causal weight sequence.

7. The deep learning-based surface subsidence risk analysis system for mining areas according to claim 2, characterized in that, Specifically, S4 is: The subsidence prediction sequence is reversed according to the time index, and the subsidence prediction sequence is rearranged in the time dimension according to the time index from back to front to generate the reverse subsidence sequence. Alignment processing is performed on the reverse sinking sequence and spatiotemporal hidden state features according to the time index and spatial index. A correspondence between the reverse sinking value and the corresponding spatiotemporal hidden state feature is established at each time index and each spatial index position to form the causal analysis input sequence. The input sequence for causal analysis is processed by feature concatenation, which concatenates the reverse sinking values ​​and spatiotemporal latent state features along the feature dimension to form a joint feature sequence containing sinking information and latent state information. Temporal correlation modeling is performed on the joint feature sequence according to the time index order to construct the dependency relationship between different moments in the time dimension, thereby obtaining the temporal correlation feature sequence that characterizes the subsidence evolution trend. Spatial neighborhood modeling is performed on the joint feature sequence according to the spatial index position. The spatial neighborhood set is determined based on the proximity relationship between spatial indices. The statistical features of the joint feature sequence are calculated in each spatial neighborhood to obtain the spatial correlation feature sequence that characterizes the spatial distribution difference. The temporal correlation feature sequence and the spatial correlation feature sequence are paired according to their time index and spatial index, and the temporal correlation feature and spatial correlation feature at the corresponding positions are weighted and merged to generate the causal correlation feature sequence. Based on the numerical relationship between the components related to disturbance change and the components related to subsidence change in the causal correlation feature sequence, the disturbance intensity estimate and causal contribution value corresponding to each time index and each spatial index position are calculated to form the disturbance intensity estimate sequence and the causal contribution sequence. Tensor organization processing is performed on the disturbance intensity estimation sequence and the causal contribution sequence according to the time index and spatial index. The disturbance intensity estimation value and the causal contribution value are arranged according to the time dimension and the spatial dimension to construct a three-dimensional structure. The first dimension is set as the time index, the second dimension is set as the spatial index, and the third dimension is set as the feature dimension composed of the disturbance intensity estimation value and the causal contribution value, forming the causal link tensor of the mining area subsidence.

8. The deep learning-based surface subsidence risk analysis system for mining areas according to claim 2, characterized in that, Specifically, S5 is: Alignment processing is performed on the subsidence prediction sequence and subsidence data based on time index and spatial index. A correspondence between the subsidence prediction value and the subsidence observation value is established at each time index and each spatial index position to construct the subsidence reconstruction target sequence. The numerical differences between the subsidence prediction sequence and the subsidence reconstruction target sequence at each time index and each spatial index position are calculated to generate a subsidence reconstruction deviation sequence. Weighted aggregation processing is performed on the subsidence reconstruction deviation sequence in the time dimension and the spatial dimension to obtain the subsidence reconstruction consistency loss term. Numerical combination processing is performed on the disturbance intensity estimation sequence and the causal contribution sequence at each time index and each spatial index position. The disturbance intensity estimation value and the causal contribution value are multiplied to generate the causal-driven subsidence estimation sequence. Alignment processing is performed on the causal-driven subsidence estimation sequence and the subsidence prediction sequence based on the time index and spatial index. The numerical difference is calculated at the corresponding position to generate the causal link deviation sequence. The causal link deviation sequence is weighted and aggregated in the time dimension and spatial dimension to obtain the causal link consistency loss term. A weighted summation process is performed on the sinking reconstruction consistency loss term and the causal link consistency loss term to generate a joint loss scalar. Gradient calculation is performed on all trainable parameters in the improved Earthformer model and causal link network based on the joint loss scalar, and numerical update is performed on each trainable parameter. The trainable parameters are cyclically updated based on the change of the joint loss scalar in consecutive training rounds.

9. A deep learning-based surface subsidence risk analysis system for mining areas according to claim 2, characterized in that, Specifically, S6 is: A threshold comparison process is performed on the disturbance contribution sequence at each time index and each spatial index position. Positions in the disturbance contribution sequence with values ​​greater than a preset disturbance threshold are marked as spatial unit subsidence source points, forming a subsidence source point marking matrix. Based on the subsidence source point marking matrix, continuous time segments are identified in the time dimension, and time positions that are continuous in the time index and have valid subsidence source point markings are grouped into time segments. Based on the subsidence source point marker matrix, spatial neighborhood sets are identified in the spatial dimension, and the valid spatial locations of subsidence source point markers are merged into spatially connected regions based on the proximity relationship between spatial indices. Perform a correspondence matching process between time segments and spatially connected regions. Retrieve the corresponding spatially connected region at the time index position covered by the time segment. Record the time index and spatial index combination that satisfies the simultaneous validity of the time segment and the spatially connected region as the spatiotemporal sinking source point segment. Link connection processing is performed on adjacent spatiotemporal subsidence source point segments based on temporal index continuity and spatial index proximity, and spatiotemporal subsidence source point segments that meet the connection conditions are merged into a spatiotemporal evolution link set. The locations of all subsidence source points in the subsidence source point marker matrix are summarized according to the time index and spatial index. After removing duplicates, a subsidence source point set is formed. Together with the spatiotemporal evolution link set, it constitutes the subsidence source area identification result.

10. A deep learning-based surface subsidence risk analysis system for mining areas according to claim 2, characterized in that, Specifically, S7 is: Based on the subsidence prediction sequence, the set of subsidence source points and the set of spatiotemporal evolution links, a correspondence is established in the time and space dimensions. Subsidence prediction values, source point markers and link markers are extracted at each time index and each space index position to form a risk analysis data sequence. A weighted combination process is performed on the risk analysis data sequence at each time index and each spatial index position to combine the subsidence prediction value, source point marker and link marker into a risk score sequence. The risk scoring sequence is subjected to a level classification process. The values ​​in the risk scoring sequence are divided into multiple level intervals according to a preset level threshold. The risk level corresponding to each position is marked in the time and space dimensions according to the level intervals, forming a risk level matrix. Based on the risk level matrix, a grid mapping process is performed in the spatial dimension to map the risk level of each spatial index position in the risk level matrix to the corresponding spatial grid cell of the mining area, thus forming a subsidence risk level map. Based on the subsidence risk level map, the level values, level change trends and spatial location parameters are extracted in the time and space dimensions to form subsidence early warning information.

Images