Coal field seismic interpretation multi-mode hidden geological disaster information fusion and dynamic modeling method

Abstract

The invention discloses a coal field earthquake interpretation multi-mode hidden geological disaster information fusion and dynamic modeling method. The method comprises the following steps: acquiring multi-source heterogeneous geological data of a target coal field area; according to the method, the multi-modal feature tensor fusion model is constructed, so that the problem of multiplicity of solutions of a single data source is effectively solved, and the integrity and consistency of hidden geological disaster features are remarkably improved; conventional seismic attributes are extracted, high-order statistical characteristics of a logging curve and geologic body structure parameters are introduced, lithologic sudden change and tectonic deformation are described from multiple dimensions, and the recognition capacity of a micro-scale hidden disaster body is enhanced; a hybrid neural network combining CNN and LSTM is adopted, historical stratum stress field data and geologic features are fused, a space-time prediction module is constructed, the limitation of traditional static snapshot type interpretation is broken through, a geologic structure stability probability body of a future time sequence can be output, and dynamic prediction of a disaster evolution trend is achieved.

Description

Technical Field

[0001] This invention relates to the field of coalfield seismic interpretation technology, specifically to a method for multimodal fusion and dynamic modeling of hidden geological hazard information in coalfield seismic interpretation. Background Technology

[0002] As coal mining depths increase, the geological environment becomes increasingly complex. Hidden geological hazards such as faults, folds, collapse columns, and water-rich anomalies pose a serious threat to coal mine safety. These hidden geological hazards are often characterized by being "small-scale, highly concealed, and highly destructive," making them difficult to accurately identify and dynamically evaluate using traditional geological exploration methods.

[0003] Currently, seismic interpretation in coalfields mainly relies on single-mode seismic data or simple multi-data overlay analysis, which has the following significant technical limitations: 1. Low degree of fusion of multi-source heterogeneous data and insufficient information utilization: Existing geological interpretation methods typically analyze 3D seismic data, well logging curves, borehole lithology, and mining records as independent data sources. For example, seismic data is mainly used for structural interpretation, while well logging data is used for lithological naming, lacking deep-level feature interaction between the two. Traditional fusion methods often employ simple "data stitching" or "weighted overlay," failing to map the spatial continuity of seismic attributes, the high resolution of well logging responses, and the hard constraints of geological boreholes to a unified high-dimensional feature space. This results in the complementarity between multimodal data not being fully explored, especially in areas with poor seismic data quality or sparse boreholes, leading to multiple interpretations.

[0004] 2. The seismic data preprocessing and feature extraction algorithms lack specificity: In seismic data preprocessing, conventional denoising methods (such as wavelet denoising with fixed thresholds) are insufficient to effectively suppress complex coherent interference and random noise. Furthermore, time-depth conversion often uses average velocity models, neglecting the influence of local velocity anomalies, resulting in low depth domain conversion accuracy. Regarding feature extraction, existing technologies are mostly limited to extracting basic seismic attributes such as instantaneous amplitude and frequency, lacking in-depth quantitative analysis of higher-order statistical features of well logging curves (such as skewness and kurtosis) and geological body structural features (such as fault attitude and curvature tensor), making it difficult to effectively characterize lithological abrupt changes and subtle tectonic deformations.

[0005] 3. Lack of dynamic prediction models considering spatiotemporal evolution: Existing geological hazard identification is mostly based on static "snapshot" data, that is, using seismic inversion results at a certain moment to delineate the hazard range. However, coalfield mining is a dynamic process. As mining activities proceed, the stress field of the strata is constantly redistributed, and hidden hazard bodies (such as fault activation and floor water inrush channels) have obvious time-varying characteristics. Traditional machine learning models (such as simple CNNs) are good at spatial feature extraction, but they are difficult to capture the long-term dependence of geological structure stability on the evolution of time; while simple time series models ignore three-dimensional spatial structural information, resulting in low accuracy in predicting the evolution trend of hazards.

[0006] 4. The disaster risk assessment indicators are too simplistic and lack quantitative accuracy: In the disaster risk assessment stage, existing methods often only classify disaster bodies based on their volume or burial depth, without comprehensively considering the coupled effects of surrounding rock stress state, water-bearing capacity, and tectonic stability. For identified anomalies, there is a lack of precise boundary delineation algorithms based on three-dimensional connected domains, and the risk index calculation often uses linear weighting, failing to incorporate physical and mechanical models such as the Mohr-Coulomb criterion to calculate the probability of shear failure, and also failing to combine low resistivity anomalies for water-bearing weighted integration. This results in a significant deviation between the disaster risk index and the actual situation, making it difficult to guide accurate disaster prevention and control decisions. Summary of the Invention

[0007] To address the aforementioned technical issues, this technical solution provides a method for fusing and dynamically modeling multimodal hidden geological hazard information in coalfield seismic interpretation. The solution resolves the problems mentioned in the background section.

[0008] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for fusing and dynamically modeling multimodal hidden geological hazard information in coalfield seismic interpretation, including: Acquire multi-source heterogeneous geological data of the target coalfield area, including three-dimensional seismic wavefield data, well logging curve data, geological borehole lithology data, and mining engineering record data; The three-dimensional seismic wavefield data is preprocessed to obtain the depth domain seismic impedance data volume. Seismic attribute features are extracted from the depth domain seismic impedance data volume, lithology-sensitive logging response features are extracted from the logging curve data, and geological body structure features are extracted from the geological borehole lithology data. The seismic attribute features include instantaneous amplitude, instantaneous frequency, coherence volume and curvature attributes, and the lithology-sensitive logging response features include the statistical distribution characteristics of natural gamma, sonic transit time and resistivity. A multimodal feature tensor fusion model is constructed to map the extracted seismic attribute features, well logging response features and geological body structure features to a unified high-dimensional feature space. Based on the low-rank sparse decomposition algorithm, the multimodal features are aligned and complemented to generate a fused feature body. Among them, the spatial distribution of geological bodies is constrained by seismic attributes, and the low-frequency components of seismic inversion are corrected by well logging and borehole data. A dynamic evolution model of hidden geological hazards is constructed based on the fusion feature body. A spatiotemporal prediction module is constructed by combining a long short-term memory network with a convolutional neural network. The input is historical stratigraphic stress field data and fusion feature body, and the output is the geological structural stability probability body of future time series. Threshold segmentation and three-dimensional connected domain analysis are performed on the probability volume of geological structural stability to identify the spatial location, boundary range and water-rich anomaly zone of hidden geological hazard bodies, and the disaster risk index is calculated based on the volume, burial depth and surrounding rock stress state of the hazard body.

[0009] Preferably, the preprocessing of the three-dimensional seismic wavefield data specifically includes the following steps: An adaptive wavelet basis function is used to decompose the seismic wavefield data at multiple scales. The threshold function is dynamically adjusted according to the signal-to-noise ratio at each scale, and random noise is suppressed by a soft-hard threshold trade-off algorithm. A prediction filter based on multinomial regression is introduced, and a prediction model is established by utilizing the linear correlation between adjacent channels. Coherent interference waves are suppressed by minimizing the prediction error function. A well-controlled velocity model is constructed, using layer velocities from geological borehole lithology data to constrain the velocity field in the time-depth conversion. An iterative inversion algorithm is employed to correct velocity anomalies, achieving high-precision conversion from the time domain to the depth domain. The time-depth conversion relationship is defined by the following integral formula: ; in, For round-trip travel, For depth, For depth The layer velocity function at the location is used, and the depth-domain seismic data is converted into a wave impedance data volume based on the wave impedance inversion algorithm. Calculated using the recursive formula: ; in, for The wave impedance value at time t. The reflection coefficient, This represents the sampling time interval.

[0010] Preferably, the extraction process of the seismic attribute characteristics and lithology-sensitive logging response characteristics includes the following data analysis logic: For earthquake attributes, Hilbert transform is used to calculate instantaneous amplitude and instantaneous frequency. Instantaneous amplitude Defined as the modulus of an analytic signal: ; in, This is a real earthquake signal. for Hilbert transform; The coherence volume is calculated using eigenvalue decomposition of the covariance matrix within a sliding time window, and the maximum positive curvature is calculated based on a three-dimensional surface fitting algorithm. With minimum negative curvature The curvature calculation formula is used to characterize the bending deformation features of geological bodies: ; in, , The depth domain construction surface is respectively in , The first derivative of the direction, , , It is the second derivative; For well logging response characteristics, the natural gamma ray, sonic transit time, and resistivity curves are standardized, and their higher-order statistical moments, such as mean, variance, skewness, and kurtosis, are calculated within the target layer. Skewness... and kurtosis The calculation is as follows: ; ; in, These are well logging sample values. The mean, Standard deviation, This represents the number of sampling points within the window. A probability density distribution function was constructed, and the kernel density estimation method was used to identify the abrupt change points in the logging response at the lithological interface.

[0011] Preferably, the extraction of the geological body structural features includes a geometrical parameterization description of faults and folds: Using lithological assemblage sequences from geological borehole lithological data, a Markov chain model was employed to analyze stratigraphic sedimentary cycles and state transition probability matrices. Defined as: ; in, For the first The lithological state of the layer, To be based on lithology Transfer to lithology The probability of; An initial geological interface model is constructed based on a three-dimensional spatial interpolation algorithm, and the strike of the fault is calculated. ,tendency ,inclination and discontinuity Parameters, fault attitude via structural surface normal vector calculate: ; ; The curvature attribute tensor voting method is used to identify the pivot line position of the fold, and the identified fault polygon and fold axis are added as hard constraints to the subsequent fusion model.

[0012] Preferably, the specific implementation method for constructing the multimodal feature tensor fusion model is as follows: Seismic attribute features, well logging response features, and geological structure features are each constructed as two-dimensional feature matrices, and then mapped to a unified high-dimensional feature space through tensor product operations to form a third-order feature tensor. , It consists of spatial dimension × spatial dimension × modal dimension; A low-rank sparse decomposition algorithm is introduced to decompose the high-dimensional feature tensor into a low-rank background matrix L, a sparse noise matrix S, and a task-specific salient feature matrix M. The optimization objective function is defined as follows: ; in, It is the nuclear norm (used to constrain low-rank properties). for Norms are used to constrain sparsity. It is the Frobenius norm. The regularization parameter is defined as follows: the low-rank term preserves the common potential structure of multimodal data, while the sparse term captures the unique anomalous geological information of each mode. The alternating direction multiplier method is used to solve the optimization objective function, thereby achieving feature alignment and complementary enhancement of multi-source heterogeneous data.

[0013] Preferably, the processing logic for constraining the spatial distribution of geological bodies using seismic attributes and correcting low-frequency components using well logging and borehole data is as follows: Extracting seismic coherence With curvature properties As a spatial weighting factor, a prior probability field for the spatial distribution of geological bodies is constructed. To limit the activation intensity of fused features in non-geological anomaly areas, the weighting function... Defined as: ; in, , The attenuation coefficient is used; a low-frequency model is generated by kriging interpolation of well logging curves and geological borehole data. The low-frequency components in the seismic inversion results were extracted using spectral decomposition technology. Calculate its residual with the interpolated low-frequency model. : ; The low-frequency components of the fused feature body are iteratively updated using the residual backpropagation algorithm, and the update formula is as follows: This eliminates the multiple solutions in earthquake inversion.

[0014] Preferably, in the step of constructing a dynamic evolution model of hidden geological hazards based on fused feature volumes, the spatiotemporal prediction module architecture design includes: A hybrid architecture of cascaded convolutional neural networks and long short-term memory networks is constructed. First, a 3D CNN is used to extract spatial texture features from the fused feature volume. The 3D convolution operation is defined as follows: ; in, For the input feature map, It is a three-dimensional convolution kernel. Using spatial coordinates, the output high-dimensional feature map is flattened into a sequence of feature vectors. ; The eigenvector sequence is compared with historical time-series formation stress field data, including vertical stress. Maximum horizontal principal stress and minimum horizontal principal stress The data is spliced ​​together in the time dimension and then input into the LSTM gating unit. LSTM cells pass through the forget gate Input gate and output gate The mechanism learns the temporal evolution of geological structural stability with mining activities, and the gating calculation formula is as follows: ; ; ; ; ; in, This is the weight matrix. For bias vectors, For the Sigmoid activation function, It is the Hadamard product; it outputs the probability of a disaster occurring at a preset time step in the future.

[0015] Preferably, the training process of the spatiotemporal prediction module includes the following optimization strategies: A sample library containing historical mining records and actual disaster locations was constructed, and training sample pairs were generated using a sliding time window technique; In the loss function The system introduces a category weight balancing term to address the scarcity of samples for hidden geological hazards, thereby increasing the loss weight of positive samples. ; Add an L2 regularization term to constrain network weights and prevent the model from overfitting; By using Dropout layers to randomly mask some neuron connections, the robustness of the model to missing data is enhanced. The Adam optimizer is used to dynamically adjust the learning rate, accelerating the model's convergence to the optimal solution. The total loss function is defined as: ; in, The labels are real-world, with 1 indicating a disaster and 0 indicating normal conditions. To predict probabilities, For the sample size, These are the network weight parameters. This is the regularization coefficient.

[0016] Preferably, the specific algorithm for threshold segmentation and three-dimensional connected component analysis of the geological structure stability probability volume is as follows: An adaptive threshold segmentation algorithm is employed, which automatically determines the optimal segmentation threshold based on the global gray-level histogram and local gray-level distribution of the probability volume using Otsu's method. inter-class variance The calculation formula is: ; in, , The pixel ratio of the foreground to the background. , This corresponds to the average grayscale value. ; The continuous probability volumes are binarized into potential hazard volume masks. A three-dimensional connected component labeling algorithm, such as a disjoint-set data structure or run-length-based connected component analysis, is applied to assign unique labels to interconnected voxels, eliminating those with volumes smaller than a preset noise threshold. Isolated point clusters; Calculate the geometric center coordinates of each connected component based on the labeling results. equivalent volume Surface area and shape factor : ; in, For the volume of a single voxel, The number of voxels within the connected component; Based on the low resistivity anomaly characteristics of the water-rich anomaly zone, the final boundary of the hidden geological hazard body was delineated.

[0017] Preferably, the calculation model for the disaster risk index includes multi-parameter coupling analysis: Establish the formula for calculating the disaster risk index R: ; in, For the volume of the disaster body, For burial depth, The stress concentration factor of the surrounding rock is . The water-bearing capacity coefficient, , , , These are the normalized influence functions for each parameter; , , , These are the weight coefficients determined using the analytic hierarchy process (AHP). The stress state of the surrounding rock was used to calculate the probability of shear failure using the Mohr-Coulomb criterion. : ; in, For shear stress, It is normal stress. The coefficient of internal friction, For cohesion, It is the standard normal distribution function; Water-rich anomaly regions based on resistivity Below a certain threshold Weighted integral over the range: ; Finally, a three-dimensional risk index field is generated and rendered in a hierarchical manner.

[0018] Compared with existing technologies, this invention provides a method for fusing and dynamically modeling multimodal hidden geological hazard information in coalfield seismic interpretation, which has the following beneficial effects: This invention constructs a multimodal feature tensor fusion model, mapping seismic, well logging, borehole, and mining data to a unified high-dimensional space. It utilizes low-rank sparse decomposition to achieve feature alignment and complementarity, constrains spatial distribution using seismic attributes, and corrects low-frequency components using well logging and borehole data. This effectively solves the problem of multiple solutions from a single data source, significantly improving the completeness and consistency of hidden geological hazard characteristics. In addition to extracting conventional seismic attributes, it also introduces higher-order statistical features of well logging curves (such as skewness and kurtosis) and geological body structural parameters (fault attitude and curvature) to characterize lithological abrupt changes and tectonic deformations from multiple dimensions, enhancing the identification of microscale hidden hazard bodies. Capabilities: Employing a hybrid neural network combining CNN and LSTM, integrating historical stratigraphic stress field data with geological features, a spatiotemporal prediction module is constructed. This overcomes the limitations of traditional static "snapshot" interpretation, enabling the output of a probability volume of geological structural stability for future time series, thus achieving dynamic prediction of disaster evolution trends. Furthermore, based on three-dimensional connected domain analysis, disaster boundaries are accurately delineated, and the probability of shear failure is calculated using the Mohr-Coulomb criterion. The disaster risk index is calculated by coupling multiple parameters including volume, burial depth, stress, and water-bearing capacity, achieving a leap from "qualitative identification" to "quantitative evaluation," providing precise decision support for coal mine water control and roof management. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the method flow for S101-S105 in this invention; Figure 2 This is a schematic diagram of the method flow for S201-S203 in this invention; Figure 3 This is a schematic diagram of the method flow for S301-S304 in this invention; Figure 4 This is a schematic diagram of the method flow for S401-S403 in this invention; Figure 5 This is a schematic diagram of the method flow for S501-S503 in this invention; Figure 6 This is a schematic diagram of the method flow for S601-S605 in this invention. Detailed Implementation

[0020] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.

[0021] Example 1 Please refer to Figure 1 As shown, the method for information fusion and dynamic modeling of multimodal hidden geological hazards in coalfield seismic interpretation includes: S101. Obtain multi-source heterogeneous geological data of the target coalfield area, including three-dimensional seismic wavefield data, well logging curve data, geological borehole lithology data, and mining engineering record data; S102. Preprocess the three-dimensional seismic wavefield data to obtain the depth domain seismic impedance data volume. Extract seismic attribute features from the depth domain seismic impedance data volume, extract lithology-sensitive logging response features from the logging curve data, and extract geological body structure features from the geological borehole lithology data. The seismic attribute features include instantaneous amplitude, instantaneous frequency, coherence volume and curvature attributes. The lithology-sensitive logging response features include the statistical distribution characteristics of natural gamma, sonic transit time and resistivity. S103. Construct a multimodal feature tensor fusion model, map the extracted seismic attribute features, well logging response features and geological body structure features to a unified high-dimensional feature space, and perform alignment and complementary enhancement of multimodal features based on a low-rank sparse decomposition algorithm to generate a fused feature body. Among them, seismic attributes are used to constrain the spatial distribution of geological bodies, and well logging and borehole data are used to correct the low-frequency components of seismic inversion. S104. Based on the fusion feature body, a dynamic evolution model of hidden geological hazards is constructed. A spatiotemporal prediction module is constructed by combining a long short-term memory network with a convolutional neural network. The input is the historical stratigraphic stress field data and the fusion feature body, and the output is the geological structural stability probability body of the future time series. S105. Threshold segmentation and three-dimensional connected domain analysis are performed on the probability volume of geological structure stability to identify the spatial location, boundary range and water-rich anomaly zone of hidden geological hazard bodies, and the disaster risk index is calculated based on the volume, burial depth and surrounding rock stress state of the hazard body.

[0022] As will be understood by those skilled in the art, this invention maps seismic, well logging, borehole, and mining data to a unified high-dimensional space by constructing a multimodal feature tensor fusion model. It utilizes low-rank sparse decomposition to achieve feature alignment and complementarity, constrains spatial distribution using seismic attributes, and corrects low-frequency components using well logging and borehole data. This effectively solves the problem of multiple solutions from a single data source and significantly improves the integrity and consistency of hidden geological hazard characteristics. Furthermore, it not only extracts conventional seismic attributes but also introduces higher-order statistical features of well logging curves (such as skewness and kurtosis) and geological body structural parameters (fault attitude and curvature), providing multi-dimensional... This method characterizes lithological abrupt changes and tectonic deformations, enhancing the ability to identify hidden hazards at the microscale. Employing a hybrid neural network combining CNN and LSTM, it integrates historical stratigraphic stress field data with geological features to construct a spatiotemporal prediction module. This overcomes the limitations of traditional static "snapshot" interpretation, outputting a probability volume of geological structural stability for future time series, enabling dynamic prediction of hazard evolution trends. Furthermore, it accurately delineates hazard boundaries based on three-dimensional connected domain analysis and calculates shear failure probability using the Mohr-Coulomb criterion, comprehensively coupling multiple parameters such as volume, burial depth, stress, and water-bearing capacity to calculate the hazard risk index. This represents a leap from "qualitative identification" to "quantitative evaluation," providing precise decision support for coal mine water control and roof management.

[0023] Please refer to Figure 2 As shown, the preprocessing of three-dimensional seismic wavefield data includes the following steps: S201. Adaptive wavelet basis functions are used to decompose seismic wavefield data into multiple scales. The threshold function is dynamically adjusted according to the signal-to-noise ratio at each scale, and random noise is suppressed by a soft-hard threshold trade-off algorithm. S202. Introduce a prediction filter based on multinomial regression, establish a prediction model by utilizing the linear correlation between adjacent channels, and suppress coherent interference waves by minimizing the prediction error function. S203. Construct a well control velocity model, using the layer velocity in geological borehole lithology data to constrain the velocity field of the time-depth conversion, and employ an iterative inversion algorithm to correct velocity anomalies, achieving high-precision conversion from the time domain to the depth domain. The time-depth conversion relationship is defined by the following integral formula: ; in, For round-trip travel, For depth, For depth The layer velocity function at the location is used, and the depth-domain seismic data is converted into a wave impedance data volume based on the wave impedance inversion algorithm. Calculated using the recursive formula: ; in, for The wave impedance value at time t. The reflection coefficient, This represents the sampling time interval.

[0024] Please refer to Figure 3 As shown, the extraction process of seismic attribute characteristics and lithology-sensitive logging response characteristics includes the following data analysis logic: S301. For earthquake attributes, the Hilbert transform is used to calculate the instantaneous amplitude and instantaneous frequency. Instantaneous amplitude Defined as the modulus of an analytic signal: ; in, This is a real earthquake signal. for Hilbert transform; S302. Calculate the coherent volume using the eigenvalue decomposition of the covariance matrix within the sliding time window, and calculate the maximum positive curvature based on a three-dimensional surface fitting algorithm. With minimum negative curvature The curvature calculation formula is used to characterize the bending deformation features of geological bodies: ; in, , The depth domain construction surface is respectively in , The first derivative of the direction, , , It is the second derivative; S303. For well logging response characteristics, the natural gamma ray, sonic transit time, and resistivity curves are standardized, and their higher-order statistical moments, such as mean, variance, skewness, and kurtosis, are calculated within the target layer. Skewness... and kurtosis The calculation is as follows: ; ; in, These are well logging sample values. The mean, Standard deviation, This represents the number of sampling points within the window. S304, and construct a probability density distribution function, and identify the abrupt change points in logging response at the lithological interface by using the kernel density estimation method.

[0025] Please refer to Figure 4 As shown, the extraction of geological body structural features includes a geometrical parameterized description of faults and folds: S401. Using lithological assemblage sequences from geological borehole lithological data, a Markov chain model is employed to analyze stratigraphic sedimentary cycles and state transition probability matrices. Defined as: ; in, For the first The lithological state of the layer, To be based on lithology Transfer to lithology The probability of; S402. Construct an initial geological interface model based on a three-dimensional spatial interpolation algorithm and calculate the strike of the fault. ,tendency ,inclination and discontinuity Parameters, fault attitude via structural surface normal vector calculate: ; ; S403, and the curvature attribute tensor voting method is used to identify the pivot line position of the fold, and the identified fault polygon and fold axis are added as hard constraints to the subsequent fusion model.

[0026] Please refer to Figure 5 As shown, the specific implementation method for constructing the multimodal feature tensor fusion model is as follows: S501. Seismic attribute features, well logging response features, and geological body structure features are each constructed as two-dimensional feature matrices, and then mapped to a unified high-dimensional feature space through tensor product operations to form a third-order feature tensor. , It consists of spatial dimension × spatial dimension × modal dimension; S502. Introduce a low-rank sparse decomposition algorithm to decompose the high-dimensional feature tensor into a low-rank background matrix L, a sparse noise matrix S, and a task-specific salient feature matrix M. The optimization objective function is defined as: ; in, It is the nuclear norm (used to constrain low-rank properties). for Norms are used to constrain sparsity. It is the Frobenius norm. The regularization parameter is defined as follows: the low-rank term preserves the common potential structure of multimodal data, while the sparse term captures the unique anomalous geological information of each mode. S503. The alternating direction multiplier method is used to solve the optimization objective function, thereby achieving feature alignment and complementary enhancement of multi-source heterogeneous data.

[0027] The processing logic for constraining the spatial distribution of geological bodies using seismic attributes and correcting low-frequency components using well logging and borehole data is as follows: Extracting seismic coherence With curvature properties As a spatial weighting factor, a prior probability field for the spatial distribution of geological bodies is constructed. To limit the activation intensity of fused features in non-geological anomaly areas, the weighting function... Defined as: ; in, , The attenuation coefficient is used; a low-frequency model is generated by kriging interpolation of well logging curves and geological borehole data. The low-frequency components in the seismic inversion results were extracted using spectral decomposition technology. Calculate its residual with the interpolated low-frequency model. : ; The low-frequency components of the fused feature body are iteratively updated using the residual backpropagation algorithm, and the update formula is as follows: This eliminates the multiple solutions in earthquake inversion.

[0028] In the steps of constructing a dynamic evolution model of hidden geological hazards based on fused feature volumes, the architecture design of the spatiotemporal prediction module includes: A hybrid architecture of cascaded convolutional neural networks and long short-term memory networks is constructed. First, a 3D CNN is used to extract spatial texture features from the fused feature volume. The 3D convolution operation is defined as follows: ; in, For the input feature map, It is a three-dimensional convolution kernel. Using spatial coordinates, the output high-dimensional feature map is flattened into a sequence of feature vectors. ; The eigenvector sequence is compared with historical time-series formation stress field data, including vertical stress. Maximum horizontal principal stress and minimum horizontal principal stress The data is spliced ​​together in the time dimension and then input into the LSTM gating unit. LSTM cells pass through the forget gate Input gate and output gate The mechanism learns the temporal evolution of geological structural stability with mining activities, and the gating calculation formula is as follows: ; ; ; ; ; in, This is the weight matrix. For bias vectors, For the Sigmoid activation function, It is the Hadamard product; it outputs the probability of a disaster occurring at a preset time step in the future.

[0029] Please refer to Figure 6 As shown, the training process of the spatiotemporal prediction module includes the following optimization strategies: S601. Construct a sample library containing historical mining records and actual disaster locations, and use sliding time window technology to generate training sample pairs; S602, in the loss function The system introduces a category weight balancing term to address the scarcity of samples for hidden geological hazards, thereby increasing the loss weight of positive samples. ; S603. Add an L2 regularization term to constrain network weights to prevent model overfitting; S604. Use the Dropout layer to randomly block some neuron connections to enhance the model's robustness to missing data. S605. The Adam optimizer is used to dynamically adjust the learning rate, accelerating the model's convergence to the optimal solution. The total loss function is defined as: ; in, The labels are real-world, with 1 indicating a disaster and 0 indicating normal conditions. To predict probabilities, For the sample size, These are the network weight parameters. is the regularization coefficient.

[0030] The specific algorithm for threshold segmentation and three-dimensional connected component analysis of the probability volume of geological structural stability is as follows: An adaptive threshold segmentation algorithm is employed, which automatically determines the optimal segmentation threshold based on the global gray-level histogram and local gray-level distribution of the probability volume using Otsu's method. inter-class variance The calculation formula is: ; in, , The pixel ratio of the foreground to the background. , This corresponds to the average grayscale value. ; The continuous probability volumes are binarized into potential hazard volume masks. A three-dimensional connected component labeling algorithm, such as a disjoint-set data structure or run-length-based connected component analysis, is applied to assign unique labels to interconnected voxels, eliminating those with volumes smaller than a preset noise threshold. Isolated point clusters; Calculate the geometric center coordinates of each connected component based on the labeling results. equivalent volume Surface area and shape factor : ; in, For the volume of a single voxel, The number of voxels within the connected component; Based on the low resistivity anomaly characteristics of the water-rich anomaly zone, the final boundary of the hidden geological hazard body was delineated.

[0031] The calculation model for the disaster risk index includes multi-parameter coupling analysis: Establish the formula for calculating the disaster risk index R: ; in, For the volume of the disaster body, For burial depth, The stress concentration factor of the surrounding rock is . The water-bearing capacity coefficient, , , , These are the normalized influence functions for each parameter; , , , These are the weight coefficients determined using the analytic hierarchy process (AHP). The stress state of the surrounding rock was used to calculate the probability of shear failure using the Mohr-Coulomb criterion. : ; in, For shear stress, It is normal stress. The coefficient of internal friction, For cohesion, It is the standard normal distribution function; Water-rich anomaly regions based on resistivity Below a certain threshold Weighted integral over the range: ; Finally, a three-dimensional risk index field is generated and rendered in a hierarchical manner.

[0032] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A method for fusing and dynamically modeling multimodal hidden geological hazard information in coalfield seismic interpretation, characterized in that: include: Acquire multi-source heterogeneous geological data of the target coalfield area, including three-dimensional seismic wavefield data, well logging curve data, geological borehole lithology data, and mining engineering record data; The three-dimensional seismic wavefield data is preprocessed to obtain the depth domain seismic impedance data volume. Seismic attribute features are extracted from the depth domain seismic impedance data volume, lithology-sensitive logging response features are extracted from the logging curve data, and geological body structure features are extracted from the geological borehole lithology data. The seismic attribute features include instantaneous amplitude, instantaneous frequency, coherence volume and curvature attributes, and the lithology-sensitive logging response features include the statistical distribution characteristics of natural gamma, sonic transit time and resistivity. A multimodal feature tensor fusion model is constructed to map the extracted seismic attribute features, well logging response features and geological body structure features to a unified high-dimensional feature space. Based on the low-rank sparse decomposition algorithm, the multimodal features are aligned and complemented to generate a fused feature body. Among them, the spatial distribution of geological bodies is constrained by seismic attributes, and the low-frequency components of seismic inversion are corrected by well logging and borehole data. A dynamic evolution model of hidden geological hazards is constructed based on the fusion feature body. A spatiotemporal prediction module is constructed by combining a long short-term memory network with a convolutional neural network. The input is historical stratigraphic stress field data and fusion feature body, and the output is the geological structural stability probability body of future time series. Threshold segmentation and three-dimensional connected domain analysis are performed on the probability volume of geological structural stability to identify the spatial location, boundary range and water-rich anomaly zone of hidden geological hazard bodies, and the disaster risk index is calculated based on the volume, burial depth and surrounding rock stress state of the hazard body.

2. The method for multimodal fusion and dynamic modeling of hidden geological hazard information in coalfield seismic interpretation according to claim 1, characterized in that, The preprocessing of the three-dimensional seismic wavefield data specifically includes the following steps: An adaptive wavelet basis function is used to decompose the seismic wavefield data at multiple scales. The threshold function is dynamically adjusted according to the signal-to-noise ratio at each scale, and random noise is suppressed by a soft-hard threshold trade-off algorithm. A prediction filter based on multinomial regression is introduced, and a prediction model is established by utilizing the linear correlation between adjacent channels. Coherent interference waves are suppressed by minimizing the prediction error function. A well-controlled velocity model is constructed, using layer velocities from geological borehole lithology data to constrain the velocity field in the time-depth conversion. An iterative inversion algorithm is employed to correct velocity anomalies, achieving high-precision conversion from the time domain to the depth domain. The time-depth conversion relationship is defined by the following integral formula: ； in, For round-trip travel, For depth, For depth The layer velocity function at the location is used, and the depth-domain seismic data is converted into a wave impedance data volume based on the wave impedance inversion algorithm. Calculated using the recursive formula: ； in, for The wave impedance value at time t. The reflection coefficient, This represents the sampling time interval.

3. The method for multimodal fusion and dynamic modeling of hidden geological hazard information in coalfield seismic interpretation according to claim 2, characterized in that, The extraction process of seismic attribute characteristics and lithology-sensitive logging response characteristics includes the following data analysis logic: For earthquake attributes, the Hilbert transform is used to calculate instantaneous amplitude and instantaneous frequency. Instantaneous amplitude Defined as the modulus of an analytic signal: ； in, This is a real earthquake signal. for Hilbert transform; The coherence volume is calculated using eigenvalue decomposition of the covariance matrix within a sliding time window, and the maximum positive curvature is calculated based on a three-dimensional surface fitting algorithm. With minimum negative curvature The curvature calculation formula is used to characterize the bending deformation features of geological bodies: ； in, , The depth domain construction surface is respectively in , The first derivative of the direction, , , It is the second derivative; For well logging response characteristics, the natural gamma ray, sonic transit time, and resistivity curves are standardized, and their higher-order statistical moments, such as mean, variance, skewness, and kurtosis, are calculated within the target layer. Skewness... and kurtosis The calculation is as follows: ； ； in, These are well logging sample values. The mean, Standard deviation This represents the number of sampling points within the window. A probability density distribution function was constructed, and the kernel density estimation method was used to identify the abrupt change points in the logging response at the lithological interface.

4. The method for multimodal fusion and dynamic modeling of hidden geological hazard information in coalfield seismic interpretation according to claim 3, characterized in that, The extraction of the geological body's structural features includes a geometrical parameterization description of faults and folds: Using lithological assemblage sequences from geological borehole lithological data, a Markov chain model was employed to analyze stratigraphic sedimentary cycles and state transition probability matrices. Defined as: ； in, For the first The lithological state of the layer, To be based on lithology Transfer to lithology The probability of; An initial geological interface model is constructed based on a three-dimensional spatial interpolation algorithm, and the strike of the fault is calculated. ,tendency ,inclination and discontinuity Parameters, fault attitude via structural surface normal vector calculate: ； ； The curvature attribute tensor voting method is used to identify the pivot line position of the fold, and the identified fault polygon and fold axis are added as hard constraints to the subsequent fusion model.

5. The method for multimodal fusion and dynamic modeling of hidden geological hazard information in coalfield seismic interpretation according to claim 4, characterized in that, The specific implementation method for constructing the multimodal feature tensor fusion model is as follows: Seismic attribute features, well logging response features, and geological structure features are each constructed as two-dimensional feature matrices, and then mapped to a unified high-dimensional feature space through tensor product operations to form a third-order feature tensor. , It consists of spatial dimension × spatial dimension × modal dimension; A low-rank sparse decomposition algorithm is introduced to decompose the high-dimensional feature tensor into a low-rank background matrix L, a sparse noise matrix S, and a task-specific salient feature matrix M. The optimization objective function is defined as follows: ； in, It is the nuclear norm (used to constrain low-rank properties). for Norms are used to constrain sparsity. It is the Frobenius norm. The regularization parameter is defined as follows: the low-rank term preserves the common potential structure of multimodal data, while the sparse term captures the unique anomalous geological information of each mode. The alternating direction multiplier method is used to solve the optimization objective function, thereby achieving feature alignment and complementary enhancement of multi-source heterogeneous data.

6. The method for multimodal fusion and dynamic modeling of hidden geological hazard information in coalfield seismic interpretation according to claim 5, characterized in that, The processing logic for constraining the spatial distribution of geological bodies using seismic attributes and correcting low-frequency components using well logging and borehole data is as follows: Extracting seismic coherence With curvature properties As a spatial weighting factor, a prior probability field for the spatial distribution of geological bodies is constructed. To limit the activation intensity of fused features in non-geological anomaly areas, the weighting function... Defined as: ； in, , The attenuation coefficient is used; a low-frequency model is generated by kriging interpolation of well logging curves and geological borehole data. The low-frequency components in the seismic inversion results were extracted using spectral decomposition technology. Calculate its residual with the interpolated low-frequency model. : ； The low-frequency components of the fused feature body are iteratively updated using the residual backpropagation algorithm, and the update formula is as follows: This eliminates the multiple solutions in earthquake inversion.

7. The method for multimodal fusion and dynamic modeling of hidden geological hazard information in coalfield seismic interpretation according to claim 6, characterized in that, In the step of constructing a dynamic evolution model of hidden geological hazards based on fused feature volumes, the architecture design of the spatiotemporal prediction module includes: A hybrid architecture of cascaded convolutional neural networks and long short-term memory networks is constructed. First, a 3D CNN is used to extract spatial texture features from the fused feature volume. The 3D convolution operation is defined as follows: ； in, For the input feature map, It is a three-dimensional convolution kernel. Using spatial coordinates, the output high-dimensional feature map is flattened into a sequence of feature vectors. ; The eigenvector sequence is compared with historical time-series formation stress field data, including vertical stress. Maximum horizontal principal stress and minimum horizontal principal stress The data is spliced ​​together in the time dimension and then input into the LSTM gating unit. LSTM cells pass through the forget gate Input gate and output gate The mechanism learns the temporal evolution of geological structural stability with mining activities, and the gating calculation formula is as follows: ； ； ； ； ； in, This is the weight matrix. For bias vectors, The Sigmoid activation function is used. It is the Hadamard product; it outputs the probability of a disaster occurring at a preset time step in the future.

8. The method for multimodal fusion and dynamic modeling of hidden geological hazard information in coalfield seismic interpretation according to claim 7, characterized in that, The training process of the spatiotemporal prediction module includes the following optimization strategies: A sample library containing historical mining records and actual disaster locations was constructed, and training sample pairs were generated using a sliding time window technique; In the loss function The system introduces a category weight balancing term to address the scarcity of samples for hidden geological hazards, thereby increasing the loss weight of positive samples. ; Add an L2 regularization term to constrain network weights and prevent the model from overfitting; By using Dropout layers to randomly mask some neuron connections, the robustness of the model to missing data is enhanced. The Adam optimizer is used to dynamically adjust the learning rate, accelerating the model's convergence to the optimal solution. The total loss function is defined as: ； in, The labels are real-world, with 1 indicating a disaster and 0 indicating normal conditions. To predict probabilities, For the sample size, These are the network weight parameters. This is the regularization coefficient.

9. The method for multimodal fusion and dynamic modeling of hidden geological hazard information in coalfield seismic interpretation according to claim 8, characterized in that, The specific algorithm for threshold segmentation and three-dimensional connected component analysis of the probability volume of geological structural stability is as follows: An adaptive threshold segmentation algorithm is employed, which automatically determines the optimal segmentation threshold based on the global gray-level histogram and local gray-level distribution of the probability volume using Otsu's method. inter-class variance The calculation formula is: ； in, , The pixel ratio of the foreground to the background. , This corresponds to the average grayscale value. ; The continuous probability volumes are binarized into potential hazard volume masks. A three-dimensional connected component labeling algorithm, such as a disjoint-set data structure or run-length-based connected component analysis, is applied to assign unique labels to interconnected voxels, and voxels with volumes smaller than a preset noise threshold are removed. Isolated point clusters; Calculate the geometric center coordinates of each connected component based on the labeling results. equivalent volume Surface area and shape factor : ； in, For the volume of a single voxel, The number of voxels within the connected component; Based on the low resistivity anomaly characteristics of the water-rich anomaly zone, the final boundary of the hidden geological hazard body was delineated.

10. The method for multimodal fusion and dynamic modeling of hidden geological hazard information in coalfield seismic interpretation according to claim 9, characterized in that, The calculation model for the disaster risk index includes multi-parameter coupling analysis: Establish the formula for calculating the disaster risk index R: ； in, For the volume of the disaster body, For burial depth, The stress concentration factor of the surrounding rock is . The water-bearing capacity coefficient, , , , These are the normalized influence functions for each parameter; , , , These are the weight coefficients determined using the analytic hierarchy process (AHP). The stress state of the surrounding rock was used to calculate the probability of shear failure using the Mohr-Coulomb criterion. : ； in, For shear stress, It is normal stress. The coefficient of internal friction, For cohesion, It is the standard normal distribution function; Water-rich anomaly regions based on resistivity Below a certain threshold Weighted integral over the range: ； Finally, a three-dimensional risk index field is generated and rendered in a hierarchical manner.

Images