A coal mine fault intelligent identification method and system

Abstract

This invention discloses an intelligent fault identification method and system for coal mines, relating to the field of coal mining. The intelligent fault identification method for coal mines includes the following steps: S1. Data loading; S2. Data preprocessing; S3. AI model construction and training; S4. Intelligent fault identification; S5. Optimization of identification results; S6. Result output and correlation. This intelligent fault identification method for coal mines establishes a multi-level cross-judgment system of primary core judgment factors + secondary auxiliary judgment factors, dynamically adjusting depth and angle thresholds to avoid misjudgment based on single data; through parameter correction, the deviation rate between the final fault parameters and the actual exposure is reduced.

Description

Technical Field

[0001] This invention relates to the field of coal mining, and in particular to a method and system for intelligent identification of coal mine faults. Background Technology

[0002] In the field of coal mining, faults, as key geological structures, directly affect the safety and efficiency of mining operations. Small faults can easily lead to coal seam mining interruptions and equipment wear and tear, while large faults may trigger major safety accidents such as water inrush and roof collapse. Therefore, accurate fault identification is a core aspect of coal mine geological exploration and production design. However, existing fault identification technologies have many limitations: Manual identification is inefficient and highly subjective: Traditional fault identification relies on geologists to manually interpret seismic profiles, requiring the tracking of phase axes and analysis of amplitude changes. When dealing with batches of SGY data from multiple work areas (the data volume of a single work area often exceeds 10GB), the processing cycle can take several weeks. Furthermore, due to differences in experience, the interpretation error of fault strike and depth can reach 5°-10m, which is difficult to meet the needs of efficient exploration. Data processing is fragmented and has poor compatibility: SGY seismic data, borehole results (geodetic coordinates), and tunnel line data (local coordinates) are often in different coordinate systems, and manual conversion can easily produce errors of ≥1m; at the same time, noise filtering of SGY data often uses simple mean denoising, which cannot effectively preserve the weak in-phase axis faulting characteristics of the fault, resulting in low data fusion quality. AI models have limited recognition capabilities: existing CNN-based fault identification models can only extract local seismic features (such as abrupt amplitude changes in a single profile), making it difficult to capture the continuity of fault strike across multiple profiles; while Transformer-based models can focus on global correlations, they lack sensitivity to local minor faulting of the same phase axis, and their training samples mostly rely solely on seismic data, failing to integrate key geological information such as lithological abrupt changes revealed by boreholes and the location of faults in roadway contact, resulting in weak model generalization ability. Summary of the Invention

[0003] The purpose of this invention is to provide a method and system for intelligent identification of coal mine faults, so as to solve the problems mentioned in the background art.

[0004] To achieve the above objectives, the present invention provides the following technical solution: a method for intelligent identification of coal mine faults, comprising the following steps: S1. Data Loading: Obtain the SGY seismic processing results data volume, borehole results data and coal mine roadway line data of the coal mining area, and load them into the preset processing environment; S2. Data Preprocessing: Perform signal optimization processing on the SGY seismic processing results data volume, perform coordinate adaptation processing on borehole data and coal mine roadway data, and construct a unified data space; S3. AI Model Construction and Training: Construct an intelligent fault recognition model, using data in a unified data space as the training basis, and enable the model to have the ability to extract and recognize fault features through iterative training; S4. Intelligent Fault Recognition: Input the preprocessed data of the area to be identified into the trained model, and output the preliminary fault recognition results; S5. Optimization of identification results: Extract the judgment factors from borehole data and coal mine roadway data, verify and correct the preliminary identification results through multi-level cross-judgment, and generate the final fault identification results; S6. Results Output and Correlation: The final fault identification results are correlated with the coal mine 3D seismic interpretation and modeling process, and the results map can be exported.

[0005] Preferably, the specific implementation steps for data loading in step S1 are as follows: S1.1 Steps for batch loading SGY data from multiple work areas: It adopts a multi-threaded parallel loading mode, supports the single batch import of SGY seismic processing data from multiple work areas, and the maximum data volume for a single import is 100GB. It has a built-in data cache unit with configurable capacity to store frequently accessed SGY data locally during the loading process, reducing the time spent on repeated readings and improving data retrieval efficiency.

[0006] S1.2 Real-time data integrity verification steps: Step 1: Format verification: Verify the consistency between the SGY file header information (including key fields such as survey line number, track number, and sampling interval) and the preset SEG-Y standard format. If there are missing fields or format mismatches, mark it as a format error. The second step is data volume verification: compare the actual number of channels in the SGY file with the number of channels marked at the top of the channel header. If the difference is ≥1 channel, it is determined that the data is missing. Only when both the format verification and the data volume verification pass (100% pass rate) can the data enter the subsequent processing stage.

[0007] S1.3, Steps for handling abnormal files: Step 1 Error Log Generation: For SGY files marked as format anomalies or missing data, automatically generate structured error logs. The logs must include the storage path of the abnormal file, the anomaly type (format anomaly / missing data), and the specific location of the anomaly (e.g., "Field 3 of track header is missing" or "Data missing in track 120"). The second step of the repair solution outputs targeted repairs for different anomaly types: if the data is missing, the adjacent channel data interpolation algorithm (based on linear interpolation of two valid channels before and after the missing channel) is used for repair; if the format is abnormal, a format conversion tool adapted to the SEG-Y2.0 standard is recommended, and a format correction parameter template is provided.

[0008] Preferably, in step S2, the signal optimization processing includes: wavelet threshold denoising (selecting the sym8 wavelet basis, decomposing 3-5 layers, and using the Birgé-Massart threshold rule) and amplitude normalization (adjusting the amplitude of each trace to the [0,1] interval based on the average amplitude of the seismic data in the work area). The coordinate adaptation process involves converting the geodetic coordinates (X,Y,Z) of the borehole data and the center coordinates of the mine roadway lines into an Inline-Crossline-Time coordinate system consistent with the SGY seismic data volume, with a conversion error ≤0.5m.

[0009] Preferably, in step S3, the intelligent tomography recognition model is a hybrid model that combines convolutional neural networks (CNN) and Transformers: The CNN module uses three convolutional layers (with kernel sizes of 3×3, 5×5, and 3×3 and a stride of 1) and two pooling layers (maximum pooling with a pooling kernel of 2×2) to extract local fault features (phase axis faulting and amplitude abrupt changes) from seismic data. The Transformer module employs a 4-layer encoder (8 attention heads, 256 hidden layer dimensions) to capture global fault spatial correlation features of seismic data (fault strike continuity, multi-section fault correspondence). The training is based on seismic data samples with fault locations marked (sample size 256×256 pixels, fault marking accuracy ≤1), fault exposure information (top and bottom plate elevation, lithological change depth) in borehole data, and fault contact information (intersection coordinates, strike angle) in coal mine roadway data.

[0010] Preferably, step S4, intelligent identification of interrupted layers, is based on a CNN-Transformer hybrid model and is implemented according to the following steps: S4.1 Steps for Quantitative Extraction of Fault Features: Step 1: Extraction of in-phase axis misalignment features: Using a CNN module (3 convolutional layers with kernel sizes of 3×3, 5×5, and 3×3, stride 1; 2 max pooling layers with 2×2 pooling kernels), the tracking deviation values ​​of in-phase axes in adjacent channels are extracted. When the deviation value is ≥2 channels and the deviation is maintained for 5 consecutive sampling points, it is determined to be an in-phase axis misalignment feature. The second step is amplitude abrupt change feature extraction: The Transformer module (4-layer encoder, 8 attention heads, 256 hidden layer dimensions) is used to capture amplitude aberration regions, and the amplitude gradient value of these regions is calculated using the formula: G = |A| 当前采样点 -A 前一采样点| / Sampling interval, when G ≥ twice the mean amplitude gradient of the seismic data in the work area, it is judged as an amplitude mutation feature; The third step is waveform distortion feature extraction: The dynamic time warping algorithm is used to calculate the waveform similarity between the target seismic trace and the five adjacent seismic waves. When the similarity is ≤0.6 and this state lasts for ≥10 sampling points, it is determined to be a waveform distortion feature. S4.2 Model Collaborative Validation Steps: Based on the local fault features of a single profile extracted by the CNN module, the local features of at least three adjacent profiles are verified by global spatial correlation using the Transformer module. If a certain local feature exists only in a single profile (e.g., there is an amplitude abrupt change only in a single profile, and there is no corresponding waveform distortion in adjacent profiles), it is marked as an interfering feature and removed, and only the effective local features that correspond to multiple profiles are retained.

[0011] S4.3, Preliminary identification result output steps: The output includes preliminary identification results containing fault location coordinates, strike angle, and depth range, and meets the following accuracy constraints: Fault location coordinates: Planar coordinate error ≤ 1 trace, depth coordinate error ≤ 50ms; Direction angle: calculated based on vector fitting of at least 5 effective feature points, with accuracy controlled within ±1°; Depth range: Based on the seismic data sampling interval (2ms / sampling point), the upper and lower limits of the depth range should be ≤2m, and the confidence level of the results should be marked (when the number of valid feature points is ≥10, the confidence level is ≥90%).

[0012] Preferably, in step S5, the rule for extracting the judgment factors is as follows: First-level core judgment factors and second-level auxiliary judgment factors are extracted from the borehole data. The first-level core judgment factors are: A1: Elevation of exposed top and bottom plates of the fault, accuracy ±0.1m; A2: Depth of borehole lithological abrupt change interface, identified by abrupt change points in natural gamma logging curves, with an abrupt change amplitude ≥30 API; A3: Abrupt change value of borehole acoustic transit time, with an abrupt change amplitude ≥50μs / m; The secondary auxiliary judgment factor is: B1: Core fragmentation, classified as 0-100% based on the percentage of fragmented rock core; B2: Abnormal borehole fluid pressure, with a pressure difference ≥2MPa from the regional average pressure; Primary core decision factors and secondary auxiliary decision factors are extracted from coal mine roadway data. The primary core decision factors are: C1: Coordinates of the intersection of the tunnel and the fault, determined through tunnel excavation records and geological logging; C2: The strike of the fault exposed at the roadway heading is measured by a compass with an accuracy of ±0.5°. The secondary auxiliary determination factors are: D1: The sudden change value of the roadway driving speed, with the difference from the normal driving speed ≥ 0.5 m / d. D2: The abnormal value of the roadway support resistance, with the difference from the designed support resistance ≥ 5 MPa. D3: The difference in displacement between the roadway rib and roof, with the difference between the rib displacement and the roof displacement ≥ 10 mm. The priority of the primary core determination factor is higher than that of the secondary auxiliary determination factor, and the priority is reflected by the weight coefficient. Among them, the weight of the primary factor is 0.6, and the weight of the secondary factor is 0.4.

[0013] Preferably, in step S5, the specific process of the multi-level cross determination is as follows: S5a. Primary cross determination: ① Fault depth verification: Calculate the difference |H - A1| between the fault depth H in the preliminary fault identification result and A1. The depth threshold T1 is dynamically adjusted according to the coal seam thickness h (T1 = 5 m when h ≤ 2 m; T1 = 8 m when 2 m < h ≤ 5 m; T1 = 15 m when h > 5 m). If |H - A1| ≤ T1, the depth verification passes. ② Fault strike verification: Use the vector dot product method to calculate the included angle between the fault strike vector identified preliminarily and the strike vector corresponding to C2 (cos = vector dot product / (product of the moduli of the two vectors)), and the angle threshold T2 is dynamically adjusted according to the roadway length L (T2 = 5° when L ≤ 500 m; T2 = 10° when L > 500 m). If ≤ T2, the strike verification passes. ③ If both the depth and strike verifications pass, enter S5c for parameter correction; if any verification fails, start S5b for secondary cross determination. S5b. Secondary cross determination: ① Calculation of the borehole side support score S1: If the depth interval [A2 - , A2 + ( = 3 m when h ≤ 2 m; = 5 m when 2 m < h ≤ 5 m; = 8 m when h > 5 m) where A2 is located contains H, and A3 ≥ T3 (T3 = 50 μs / m for soft rock formations; T3 = 100 μs / m for hard rock formations), and at the same time B1 ≥ 50% or B2 ≥ 2 MPa, then S1 = 1; otherwise, S1 = 0.3×(A3 / T3) + 0.2×(B1 / 100%) + 0.5×(B2 / 2 MPa) (the value range of S1 is [0,1]). ②Calculation of the side support S2 of the roadway: If the vertical distance d from C1 to the preliminary identified fault strike line is ≤ T4 (T4 = 2m for nearly horizontal coal seams; T4 = 5m for inclined coal seams), and D1 ≤ T5 (T5 = 0.3m / d for rock roadways; T5 = 0.8m / d for coal roadways), D2 ≥ T6 (T6 = 25MPa for bolt-net support; T6 = 40MPa for brick lining support), and at the same time D3 ≥ 10mm, then S2 = 1; otherwise, S2 = 0.4×((T4 - d) / T4) + 0.3×((T5 - D1) / T5) + 0.2×(D2 / T6) + 0.1×(D3 / 10mm) (the value range of S2 is [0,1]); ③Calculation of the comprehensive support degree S: S = α×S1 + β×S2 (α + β = 1), and the weight coefficients α and β are dynamically adjusted according to the geological complexity of the working area (drilling density <000009�>1 ≥ 5 per km², α = 0.55 - 0.6, β = 0.4 - 0.45; roadway density 2 ≥ 2 km / km², β = 0.55 - 0.6, α = 0.4 - 0.45; <০০০০০৯৬>1 ≥ 5 and <০০০০০৯৭>2 ≥ 2, α = β = 0.5); the comprehensive threshold T7 is adjusted according to the fault development degree of the working area (T7 = 0.8 in the densely fault-developed area; T7 = 0.7 in the sparsely fault-developed area); <০০০০০৯৮>④If S ≥ T7, it is determined that the preliminary identification result is valid, and enter the S5c parameter correction; if S < T7, it is marked as a suspected fault, and an artificial review prompt is output; <০০০০০৯৯>S5c. Parameter correction: Based on the valid determination result, perform three-dimensional parameter calibration. <০০০০১০০><০০০০১০১>Preferably, in step S5c, the specific process of the parameter correction is as follows: <০০০০১০২>S5c1. Fault depth calibration: Taking A1 as the benchmark, if the deviation between the preliminary identified depth H and A1 is ≤ T1, directly take A1 as the final depth; if the deviation > T1, perform linear interpolation correction in combination with A2 (the final depth H_final = A1 + (A2 - A1)×(H - A1) / (H - A2)), and the depth error after correction is ≤ 0.5m; <০০০০১০৩>S5c2. Fault strike correction: Taking C1 as the control point, use the least squares method to fit the fault strike line. If there are multiple C1 (≥ 3), calculate the residuals from each C1 to the preliminary strike line,剔除 the abnormal C1 with residuals > 2 times the standard deviation, and then refit the strike line. The strike angle error after correction is ≤ 0.5°; <০০০০১০৪>S5c3. Fault dip verification: <০০০০১০৫>①Extract the dip angle of the lithological interface corresponding to A2 from the borehole logging data <০০০০১০৬>1 (measured by a well inclinometer, accuracy ±1°); <০০০০১০৭>②Fit the dip angle according to D3 <০০০০১০৮>2. The fitting relationship is as follows: 2 = k × D3 + b (k and b are the fitting coefficients of the work area, obtained by regression of more than 100 sets of measured data, such as k = 0.5° / mm and b = 25° for coal-bearing strata). ③Calculation _avg=( 1+ 2) / 2, if | 1- 2|≤15°, take _avg is the final tendency angle; if | 1- 2|>15°, activate the AI ​​model to re-extract the dip characteristics of the in-phase axis from the 3×3 neighborhood seismic profile of this area (extraction window size 512×512 pixels), and obtain the AI-predicted dip angle. AI, ultimately favoring an angle _final=0.4× 1+0.3× 2+0.3× _AI, after correction, the tilt angle error is ≤2°; S5c4. Fault Displacement Calculation: Based on the calibrated depth, strike, and dip parameters, and combined with the elevation difference between the hanging wall and footwall coal seams, calculate the fault displacement (displacement = |hanging wall coal seam elevation - footwall coal seam elevation| / sin (final), calculation accuracy ≤ 0.1m.

[0015] Preferably, in step S6, the association process includes: ① Linked verification: The final fault identification results are linked with seismic horizon interpretation data (such as coal seam roof and floor horizons) to establish the spatial topological relationship between faults and horizons. When the horizon interpretation results are updated (such as horizon depth adjustment ≥0.5m), the secondary verification of the fault identification results is automatically triggered (S5a depth verification and S5c1 depth calibration are re-executed) to ensure the consistency of the spatial relationship between faults and horizons (the coordinate deviation of the intersection point of the fault and the horizon is ≤1m). ②Results Export: Supports saving seismic time profile results in CGM format (map resolution 300dpi, including fault labels (name, strike, dip, elevation difference), borehole / tunnel location labels, and coordinate grid); ③ Manual verification assistance: The manual verification prompts include seismic profile screenshots of suspected fault areas (within the range of ±50 traces and ±500ms from the fault center), a comparison table of the specific values ​​of each judgment factor and the threshold (indicating factors that do not meet the threshold and their deviation values), and the coordinates of the borehole / tunnel locations to be verified (with a coordinate jump link, which can be clicked to directly access the corresponding data display interface).

[0016] A coal mine fault intelligent identification system, used to execute a coal mine fault intelligent identification method, comprising: Data loading module: Used to load SGY seismic processing results data, borehole results data and coal mine roadway line data. It supports data format verification (verification of SGY roadway head and borehole coordinate format) and batch import (single import limit of 100GB). It has a built-in data cache unit (cache capacity is configurable, supports frequent access to local data storage, and improves call efficiency by more than 30%). Data preprocessing module: Connects to the data loading module and is used to perform wavelet threshold denoising and amplitude normalization on the SGY seismic data volume, perform coordinate system transformation on borehole and tunnel line data, and construct a unified data space (data space accuracy ≤ 0.1m). The AI ​​intelligent recognition module includes a model building submodule (supporting the construction of CNN-Transformer hybrid models, with customizable network layers and convolutional kernel size), a model training submodule (supporting batch sample training and built-in 5-fold cross-validation), and an inference recognition submodule (supporting parallel inference, improving single-area recognition efficiency by 50%), which is used to output preliminary tomographic recognition results. Result optimization module: Connects to the AI ​​intelligent recognition module, including a multi-factor extraction unit (automatically extracts primary / secondary judgment factors, supports factor numerical quantification and outlier removal), a cross-judgment unit (performs primary / secondary cross-judgment, with a built-in threshold self-updating subunit, automatically optimizing the threshold every 50 sets of manually reviewed samples), and a parameter correction unit (performs depth calibration, orientation correction, tendency verification and drop calculation), generating the final tomographic recognition result; Results Output and Correlation Module: Connects to the results optimization module to establish the linkage between faults and stratigraphic levels. It supports exporting CGM format result maps and outputs fault interpretation files (including fault parameter tables and judgment process reports). Display module: Used to display seismic time profiles (supports arbitrary line cutting, cutting accuracy ±1 line), borehole and tunnel baseline maps (supports layer overlay display), fault identification results (annotating fault parameters and judgment factor values), and supports multi-user collaborative viewing.

[0017] The technical effects and advantages of this invention are as follows: 1. This intelligent fault identification method for coal mines supports batch loading of SGY data from multiple work areas. During the loading process, it verifies the consistency of track head information and data track count in real time. It generates error logs containing abnormal locations and error types for damaged files and provides recommended solutions for interpolation repair of adjacent tracks and format conversion tools, thereby improving the data integrity assurance rate. It innovates a CNN-Transformer hybrid model, with the CNN module extracting local fault features and the Transformer module capturing global spatial correlations. Combined with labeled seismic samples, borehole fault exposure information, and roadway contact information for training, the model's fault location identification error is reduced and the accuracy of strike angle is improved.

[0018] 2. The intelligent fault identification method for coal mines establishes a multi-level cross-judgment system of primary core judgment factors and secondary auxiliary judgment factors, dynamically adjusts depth threshold and angle threshold to avoid misjudgment based on single data; through parameter correction, the deviation rate between the fault parameters and the actual exposure is reduced. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the overall process of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] This invention provides, for example Figure 1 The method for intelligent identification of coal mine faults, as shown, includes the following steps: S1. Data Loading: Obtain the SGY seismic processing results data volume, borehole results data and coal mine roadway line data of the coal mining area, and load them into the preset processing environment; S2. Data Preprocessing: Perform signal optimization processing on the SGY seismic processing results data volume, perform coordinate adaptation processing on borehole data and coal mine roadway data, and construct a unified data space; S3. AI Model Construction and Training: Construct an intelligent fault recognition model, using data in a unified data space as the training basis, and enable the model to have the ability to extract and recognize fault features through iterative training; S4. Intelligent Fault Recognition: Input the preprocessed data of the area to be identified into the trained model, and output the preliminary fault recognition results; S5. Optimization of identification results: Extract the judgment factors from borehole data and coal mine roadway data, verify and correct the preliminary identification results through multi-level cross-judgment, and generate the final fault identification results; S6. Results Output and Correlation: The final fault identification results are correlated with the coal mine 3D seismic interpretation and modeling process, and the results map can be exported.

[0022] Preferably, the specific implementation steps for data loading in step S1 are as follows: S1.1 Steps for batch loading SGY data from multiple work areas: It adopts a multi-threaded parallel loading mode, supports the single batch import of SGY seismic processing data from multiple work areas, and the maximum data volume for a single import is 100GB. It has a built-in data cache unit with configurable capacity to store frequently accessed SGY data locally during the loading process, reducing the time spent on repeated readings and improving data retrieval efficiency.

[0023] S1.2 Real-time data integrity verification steps: Step 1: Format verification: Verify the consistency between the SGY file header information (including key fields such as survey line number, track number, and sampling interval) and the preset SEG-Y standard format. If there are missing fields or format mismatches, mark it as a format error. The second step is data volume verification: compare the actual number of channels in the SGY file with the number of channels marked at the top of the channel header. If the difference is ≥1 channel, it is determined that the data is missing. Only when both the format verification and the data volume verification pass (100% pass rate) can the data enter the subsequent processing stage.

[0024] S1.3, Steps for handling abnormal files: Step 1 Error Log Generation: For SGY files marked as format anomalies or missing data, automatically generate structured error logs. The logs must include the storage path of the abnormal file, the anomaly type (format anomaly / missing data), and the specific location of the anomaly (e.g., "Field 3 of track header is missing" or "Data missing in track 120"). The second step of the repair solution outputs targeted repairs for different anomaly types: if the data is missing, the adjacent channel data interpolation algorithm (based on linear interpolation of two valid channels before and after the missing channel) is used for repair; if the format is abnormal, a format conversion tool adapted to the SEG-Y2.0 standard is recommended, and a format correction parameter template is provided.

[0025] Preferably, in step S2, the signal optimization processing includes: wavelet threshold denoising (selecting the sym8 wavelet basis, decomposing 3-5 layers, and using the Birgé-Massart threshold rule) and amplitude normalization (adjusting the amplitude of each trace to the [0,1] interval based on the average amplitude of the seismic data in the work area). The coordinate adaptation process involves converting the geodetic coordinates (X,Y,Z) of the borehole data and the center coordinates of the mine roadway lines into an Inline-Crossline-Time coordinate system consistent with the SGY seismic data volume, with a conversion error ≤0.5m.

[0026] Preferably, in step S3, the intelligent tomography recognition model is a hybrid model that combines convolutional neural networks (CNN) and Transformers: The CNN module uses three convolutional layers (with kernel sizes of 3×3, 5×5, and 3×3 and a stride of 1) and two pooling layers (maximum pooling with a pooling kernel of 2×2) to extract local fault features (phase axis faulting and amplitude abrupt changes) from seismic data. The Transformer module employs a 4-layer encoder (8 attention heads, 256 hidden layer dimensions) to capture global fault spatial correlation features of seismic data (fault strike continuity, multi-section fault correspondence). The training is based on seismic data samples with fault locations marked (sample size 256×256 pixels, fault marking accuracy ≤1), fault exposure information (top and bottom plate elevation, lithological change depth) in borehole data, and fault contact information (intersection coordinates, strike angle) in coal mine roadway data.

[0027] Preferably, step S4, intelligent identification of interrupted layers, is based on a CNN-Transformer hybrid model and is implemented according to the following steps: S4.1 Steps for Quantitative Extraction of Fault Features: Step 1: Extraction of in-phase axis misalignment features: Using a CNN module (3 convolutional layers with kernel sizes of 3×3, 5×5, and 3×3, stride 1; 2 max pooling layers with 2×2 pooling kernels), the tracking deviation values ​​of in-phase axes in adjacent channels are extracted. When the deviation value is ≥2 channels and the deviation is maintained for 5 consecutive sampling points, it is determined to be an in-phase axis misalignment feature. The second step is amplitude abrupt change feature extraction: The Transformer module (4-layer encoder, 8 attention heads, 256 hidden layer dimensions) is used to capture amplitude aberration regions, and the amplitude gradient value of these regions is calculated using the formula: G = |A| 当前采样点 -A 前一采样点 | / Sampling interval, when G ≥ twice the mean amplitude gradient of the seismic data in the work area, it is judged as an amplitude mutation feature; The third step is waveform distortion feature extraction: The dynamic time warping algorithm is used to calculate the waveform similarity between the target seismic trace and the five adjacent seismic waves. When the similarity is ≤0.6 and this state lasts for ≥10 sampling points, it is determined to be a waveform distortion feature. S4.2 Model Collaborative Validation Steps: Based on the local fault features of a single profile extracted by the CNN module, the local features of at least three adjacent profiles are verified by global spatial correlation using the Transformer module. If a certain local feature exists only in a single profile (e.g., there is an amplitude abrupt change only in a single profile, and there is no corresponding waveform distortion in adjacent profiles), it is marked as an interfering feature and removed, and only the effective local features that correspond to multiple profiles are retained.

[0028] S4.3, Preliminary identification result output steps: The output includes preliminary identification results containing fault location coordinates, strike angle, and depth range, and meets the following accuracy constraints: Fault location coordinates: Planar coordinate error ≤ 1 trace, depth coordinate error ≤ 50ms; Direction angle: calculated based on vector fitting of at least 5 effective feature points, with accuracy controlled within ±1°; Depth range: Based on the seismic data sampling interval (2ms / sampling point), the upper and lower limits of the depth range should be ≤2m, and the confidence level of the results should be marked (when the number of valid feature points is ≥10, the confidence level is ≥90%).

[0029] Preferably, in step S5, the rule for extracting the judgment factors is as follows: First-level core judgment factors and second-level auxiliary judgment factors are extracted from the borehole data. The first-level core judgment factors are: A1: Elevation of exposed top and bottom plates of the fault, accuracy ±0.1m; A2: Depth of borehole lithological abrupt change interface, identified by abrupt change points in natural gamma logging curves, with an abrupt change amplitude ≥30 API; A3: Abrupt change value of borehole acoustic transit time, with an abrupt change amplitude ≥50μs / m; The secondary auxiliary judgment factor is: B1: Core fragmentation, classified as 0-100% based on the percentage of fragmented rock core; B2: Abnormal borehole fluid pressure, with a pressure difference ≥2MPa from the regional average pressure; Primary core decision factors and secondary auxiliary decision factors are extracted from coal mine roadway data. The primary core decision factors are: C1: Coordinates of the intersection of the tunnel and the fault, determined through tunnel excavation records and geological logging; C2: The strike of the fault at the tunnel face is revealed, measured with a compass, with an accuracy of ±0.5°; The secondary auxiliary judgment factor is: D1: Sudden change in tunnel excavation speed, with a difference from the normal excavation speed ≥ 0.5 m / d; D2: Abnormal value of roadway support resistance, the difference between the roadway support resistance and the design support resistance is ≥5MPa; D3: The displacement difference between the roadway rib and roof, the displacement difference between the rib displacement and the roof displacement ≥ 10 mm; The priority of the primary core determination factor is higher than that of the secondary auxiliary determination factor, and the priority is reflected by the weight coefficient. Among them, the weight of the primary factor is 0.6, and the weight of the secondary factor is 0.4.

[0030] Preferably, in step S5, the specific process of the multi-level cross determination is as follows: S5a. Primary cross determination: ① Fault depth verification: Calculate the difference |H - A1| between the fault depth H in the preliminary fault recognition result and A1. The depth threshold T1 is dynamically adjusted according to the coal seam thickness h (T1 = 5 m when h ≤ 2 m; T1 = 8 m when 2 m < h ≤ 5 m; T1 = 15 m when h > 5 m). If |H - A1| ≤ T1, the depth verification passes; ② Fault strike verification: Use the vector dot product method to calculate the included angle between the fault strike vector in the preliminary recognition and the strike vector corresponding to C2 (cos = vector dot product / (product of the moduli of the two vectors)), and the angle threshold T2 is dynamically adjusted according to the roadway length L (T2 = 5° when L ≤ 500 m; T2 = 10° when L > 500 m). If ≤ T2, the strike verification passes; ③ If both the depth and strike verifications pass, enter S5c parameter correction; if any verification fails, start S5b secondary cross determination; S5b. Secondary cross determination: ① Calculation of the borehole side support score S1: If the depth interval [A2 - , A2 + ( = 3 m, when h ≤ 2 m; = 5 m, when 2 m < h ≤ 5 m; = 8 m, when h > 5 m) contains H, and A3 ≥ T3 (T3 = 50 μs / m for soft rock formations; T3 = 100 μs / m for hard rock formations), and at the same time B1 ≥ 50% or B2 ≥ 2 MPa, then S1 = 1; otherwise, S1 = 0.3×(A3 / T3) + 0.2×(B1 / 100%) + 0.5×(B2 / 2 MPa) (the value range of S1 is [0, 1]); ②Calculation of the side support S2 of the roadway: If the vertical distance d from C1 to the initially identified fault strike line is ≤ T4 (T4 = 2 m for nearly horizontal coal seams; T4 = 5 m for inclined coal seams), and D1 ≤ T5 (T5 = 0.3 m / d for rock roadways; T5 = 0.8 m / d for coal roadways), D2 ≥ T6 (T6 = 25 MPa for bolt - mesh support; T6 = 40 MPa for masonry support), and at the same time D3 ≥ 10 mm, then S2 = 1; otherwise, S2 = 0.4×((T4 - d) / T4)+0.3×((T5 - D1) / T5)+0.2×(D2 / T6)+0.1×(D3 / 10 mm) (the value range of S2 is [0,1]); ③Calculation of the comprehensive support degree S: S = α×S1+β×S2 (α + β = 1), and the weight coefficients α and β are dynamically adjusted according to the geological complexity of the working area (drilling density 1≥5 per km², α = 0.55 - 0.6, β = 0.4 - 0.45; roadway density 2≥2 km / km², β = 0.55 - 0.6, α = 0.4 - 0.45; When 1≥5 and 2≥2, α = β = 0.5); the comprehensive threshold T7 is adjusted according to the fault development degree in the working area (T7 = 0.8 in the area with dense fault development; T7 = 0.7 in the area with sparse fault development); ④If S≥T7, it is determined that the initially identified result is valid, and enter the S5c parameter correction; if S<T7, it is marked as a suspected fault, and an artificial review prompt is output; S5c. Parameter correction: Based on the valid determination result, perform three - dimensional parameter calibration.

[0031] Preferably, in step S5c, the specific process of the parameter correction is as follows: S5c1. Fault depth calibration: Taking A1 as the benchmark, if the deviation between the initially identified depth H and A1 is ≤ T1, directly take A1 as the final depth; if the deviation > T1, perform linear interpolation correction in combination with A2 (the final depth H_final = A1+(A2 - A1)×(H - A1) / (H - A2)), and the depth error after correction is ≤ 0.5 m; S5c2. Fault strike correction: Taking C1 as the control point, use the least - squares method to fit the fault strike line. If there are multiple C1 (≥3), calculate the residuals of each C1 to the initially fitted strike line, remove the abnormal C1 with residuals > 2 times the standard deviation, and then refit the strike line. The strike angle error after correction is ≤ 0.5°; S5c3. Fault dip verification: ①Extract the dip angle of the lithologic interface corresponding to A2 from the borehole logging data 1 (measured by a well - inclinometer, with an accuracy of ±1°); ②Fit the dip angle according to D3 2. The fitting relationship is as follows: 2 = k × D3 + b (k and b are the fitting coefficients of the work area, obtained by regression of more than 100 sets of measured data, such as k = 0.5° / mm and b = 25° for coal-bearing strata). ③Calculation _avg=( 1+ 2) / 2, if | 1- 2|≤15°, take _avg is the final tendency angle; if | 1- 2|>15°, activate the AI ​​model to re-extract the dip characteristics of the in-phase axis from the 3×3 neighborhood seismic profile of this area (extraction window size 512×512 pixels), and obtain the AI-predicted dip angle. AI, ultimately favoring an angle _final=0.4× 1+0.3× 2+0.3× _AI, after correction, the tilt angle error is ≤2°; S5c4. Fault Displacement Calculation: Based on the calibrated depth, strike, and dip parameters, and combined with the elevation difference between the hanging wall and footwall coal seams, calculate the fault displacement (displacement = |hanging wall coal seam elevation - footwall coal seam elevation| / sin (final), calculation accuracy ≤ 0.1m.

[0032] Preferably, in step S6, the association process includes: ① Linked verification: The final fault identification results are linked with seismic horizon interpretation data (such as coal seam roof and floor horizons) to establish the spatial topological relationship between faults and horizons. When the horizon interpretation results are updated (such as horizon depth adjustment ≥0.5m), the secondary verification of the fault identification results is automatically triggered (S5a depth verification and S5c1 depth calibration are re-executed) to ensure the consistency of the spatial relationship between faults and horizons (the coordinate deviation of the intersection point of the fault and the horizon is ≤1m). ②Results Export: Supports saving seismic time profile results in CGM format (map resolution 300dpi, including fault labels (name, strike, dip, elevation difference), borehole / tunnel location labels, and coordinate grid); ③ Manual verification assistance: The manual verification prompts include seismic profile screenshots of suspected fault areas (within the range of ±50 traces and ±500ms from the fault center), a comparison table of the specific values ​​of each judgment factor and the threshold (indicating factors that do not meet the threshold and their deviation values), and the coordinates of the borehole / tunnel locations to be verified (with a coordinate jump link, which can be clicked to directly access the corresponding data display interface).

[0033] It supports batch loading of SGY data from multiple work areas, and verifies the consistency of track head information and data number in real time during the loading process. It generates error logs containing abnormal locations and error types for damaged files, and provides recommended solutions for interpolation repair of adjacent tracks and format conversion tools, thereby improving the data integrity guarantee rate. It innovates the CNN-Transformer hybrid model, in which the CNN module extracts local fault features and the Transformer module captures global spatial correlations. Combined with labeled seismic samples, borehole fault exposure information, and roadway contact information for training, the model's fault location identification error is reduced and the accuracy of strike angle is improved.

[0034] A multi-level cross-judgment system consisting of a primary core judgment factor and a secondary auxiliary judgment factor was established to dynamically adjust the depth threshold and angle threshold, thereby avoiding misjudgment based on single data. Through parameter correction, the deviation rate between the fault parameters and the actual exposure was reduced.

[0035] A coal mine fault intelligent identification system, used to execute a coal mine fault intelligent identification method, comprising: Data loading module: Used to load SGY seismic processing results data, borehole results data and coal mine roadway line data. It supports data format verification (verification of SGY roadway head and borehole coordinate format) and batch import (single import limit of 100GB). It has a built-in data cache unit (cache capacity is configurable, supports frequent access to local data storage, and improves call efficiency by more than 30%). Data preprocessing module: Connects to the data loading module and is used to perform wavelet threshold denoising and amplitude normalization on the SGY seismic data volume, perform coordinate system transformation on borehole and tunnel line data, and construct a unified data space (data space accuracy ≤ 0.1m). The AI ​​intelligent recognition module includes a model building submodule (supporting the construction of CNN-Transformer hybrid models, with customizable network layers and convolutional kernel size), a model training submodule (supporting batch sample training and built-in 5-fold cross-validation), and an inference recognition submodule (supporting parallel inference, improving single-area recognition efficiency by 50%), which is used to output preliminary tomographic recognition results. Result optimization module: Connects to the AI ​​intelligent recognition module, including a multi-factor extraction unit (automatically extracts primary / secondary judgment factors, supports factor numerical quantification and outlier removal), a cross-judgment unit (performs primary / secondary cross-judgment, with a built-in threshold self-updating subunit, automatically optimizing the threshold every 50 sets of manually reviewed samples), and a parameter correction unit (performs depth calibration, orientation correction, tendency verification and drop calculation), generating the final tomographic recognition result; Results Output and Correlation Module: Connects to the results optimization module to establish the linkage between faults and stratigraphic levels. It supports exporting CGM format result maps and outputs fault interpretation files (including fault parameter tables and judgment process reports). Display module: Used to display seismic time profiles (supports arbitrary line cutting, cutting accuracy ±1 line), borehole and tunnel baseline maps (supports layer overlay display), fault identification results (annotating fault parameters and judgment factor values), and supports multi-user collaborative viewing.

[0036] The overall workflow and specific steps are as follows: Step 1: Loading and Validating Multi-Source Data 1.1 Data Acquisition: Collect three types of core data from the target area of ​​the coal mine: SGY seismic processing results data (recording seismic wave signals of underground geological structures), borehole results data (including geodetic coordinates X / Y / Z, natural gamma logging curves, sonic transit time, etc.), and coal mine roadway line data (including roadway center coordinates, tunneling records, support resistance, etc.). 1.2 Batch Loading: Import data through the data loading module, supporting batch loading of SGY data from multiple work areas (maximum of 100GB per load), with a built-in cache unit to store frequently accessed data; 1.3 Data Verification and Repair: Real-time verification of the consistency between SGY file header information, data trace number and preset format, and the standardization of borehole data coordinate format; for damaged / format-abnormal files, generate an error log containing "abnormal location (e.g., SGY trace 120) and error type (e.g., missing trace header)" and suggest repair solutions (e.g., repairing damaged traces based on interpolation of adjacent trace data, recommending format conversion tools to correct coordinate format). Step 2: Multi-source data preprocessing and unified space construction (corresponding to claims 1, 2, and 3) 2.1 SGY Signal Optimization: Perform two-step processing on the SGY data volume. ① Wavelet thresholding: Select the sym8 wavelet basis, decompose it into 3-5 layers, and use the Birgé-Massart strategy to determine the threshold and filter random noise; ② Amplitude normalization: Based on the average amplitude of the seismic data in the work area, the amplitude of each trace is adjusted to the [0,1] interval to eliminate the interference of amplitude differences on identification; 2.2 Coordinate Adaptation and Transformation: The geodetic coordinates (X,Y,Z) and center coordinates of the tunnel lines in the borehole results are converted into an Inline-Crossline-Time coordinate system consistent with the SGY data through a coordinate transformation algorithm. The transformation error is strictly controlled to ≤0.5m. 2.3 Unified Data Space Construction: Integrate and optimize SGY data and transformed borehole / tunnel data to construct a unified data space with an accuracy of ≤0.1m, ensuring spatial alignment of multi-source data. Step 3: Construction and Training of the CNN-Transformer Hybrid Model 3.1 Model Architecture Setup: A hybrid model is built using the model building submodule of the AI ​​intelligent recognition module. ①CNN module: 3 convolutional layers (3×3 convolutional kernels → 5×5 → 3×3, stride 1) + 2 max pooling layers (2×2 pooling kernels), responsible for extracting local fault features; ②Transformer module: 4-layer encoder (8 attention heads, 256 hidden layer dimensions), responsible for capturing global fault spatial correlation features; 3.2 Training Sample Preparation: Based on the unified data space, three types of training samples were prepared. ① Mark earthquake samples: 256×256 pixel seismic profiles, with fault locations marked (accuracy ≤ 1 trace); ② Drilling fault information: fault top and bottom plate elevations, lithological abrupt change depths, and sonic transit time abrupt change values; ③ Tunnel contact information: fault intersection coordinates and strike angles; 3.3 Model Iterative Training: Load samples through the model training submodule, enable the 5-fold cross-validation function, and iteratively adjust model parameters (such as learning rate and number of convolutional kernels) until the model's accuracy in extracting tomographic features is ≥95%, and it has a stable ability to extract and recognize tomographic features. Step 4: Intelligent Fault Identification and Preliminary Result Output 4.1 Input of data to be identified: Input the preprocessed data (in a uniform data space format) of the target region to be identified into the trained model; 4.2 Fault Feature Extraction: The model automatically extracts three types of core fault features. ① Phase axis misalignment: When tracing adjacent phase axes, a deviation value of ≥2 tracks is considered a misalignment; ②Amplitude mutation: Calculate the amplitude gradient value; a mutation is defined as ≥ twice the mean. ③ Waveform distortion: The waveform similarity is calculated using a dynamic time warping algorithm; a similarity of ≤0.6 is considered distortion. 4.3 Preliminary Result Output: Generate preliminary fault identification results, including fault location coordinates, strike angle (accuracy ±1°), and depth range (accuracy ±2m). Step 5: Multi-level cross-determination and result optimization 5.1 Judgment Factor Extraction: Using the multi-factor extraction unit of the result optimization module, primary core factors (A1: fault top and bottom plate elevation ±0.1m; A2: lithological abrupt change depth, gamma change ≥30API; A3: sonic transit time change ≥50μs / m) and secondary auxiliary factors (B1: core fragmentation; B2: fluid pressure anomaly ≥2MPa) are extracted from borehole data; primary core factors (C1: fault intersection coordinates; C2: strike angle ±0.5°) and secondary auxiliary factors (D1: tunneling speed change ≥0.5m / d; D2: support resistance anomaly ≥5MPa; D3: sidewall and roof displacement difference ≥10mm) are extracted from roadway data. 5.2 Multi-level cross-determination: Level 1 crossover determination: ① Depth verification: Calculate the difference between the initial identification depth H and A1, and take the threshold T1 (5m / 8m / 15m) according to the coal seam thickness h. If |H-A1|≤T1, it passes; ② Direction verification: Calculate the direction angle using the vector dot product method. The threshold T2 (5° / 10°) is taken based on the tunnel length L. If the value is less than or equal to T2, the test passes; if all values ​​pass, parameter correction is initiated; otherwise, a secondary judgment is activated. Secondary cross-referencing: ① Calculate the borehole side support component S1 (combined with A2 / A3 / B1 / B2, with values ​​[0,1]); ② Calculate the roadway side support component S2 (combined with C1 / D1 / D2 / D3, with values ​​[0,1]); ③ Calculate the comprehensive support S=αS1+βS2 (α / β are dynamically adjusted according to the borehole / tunnel density), and take the threshold T7 (0.7 / 0.8) according to the degree of fault development. If S≥T7, it is considered valid; otherwise, it is marked as a suspected fault and prompts for manual review. 5.3 Parameter Correction: ① Depth Calibration: Using A1 as the benchmark, if the deviation is ≤ T1, then A1 is used; otherwise, linear interpolation with A2 is used (error ≤ 0.5m); ② Strike Correction: Least squares fitting of C1 is used, and outliers are removed before refitting (error ≤ 0.5°); ③ Dip Verification: Combined with borehole logging dip verification. 1. Roadway displacement fitting 2. If the deviation is ≤15°, the mean value is taken; otherwise, the AI ​​prediction tends to be biased. _AI (error ≤ 2°); ④ Drop calculation: Based on the calibration parameters and the elevation difference between the upper and lower coal seams, calculate the drop (accuracy ≤ 0.1m); finally, generate the optimized fault identification results. Step 6: Linking Results Output with 3D Flowchart 6.1 Linkage Verification: Through the result output and association module, the final result is linked with the seismic horizon interpretation data (coal seam top / bottom) to establish spatial topological relationships; when the horizon depth adjustment is ≥0.5m, depth verification and depth calibration are automatically triggered to ensure that the fault-horizontal intersection coordinate deviation is ≤1m; 6.2 Output Export: Supports exporting two types of outputs. ①CGM format seismic time profile results (300dpi, including fault parameter annotations, borehole / tunnel locations, and coordinate grid); ② Fault interpretation documents (including parameter tables and determination process reports); 6.3 Manual Verification Assistance: For suspected faults, output auxiliary information. ① Seismic profile screenshot (fault center ±50 traces, ±500ms); ② Factor threshold comparison table (marking factors that do not meet the threshold and their deviation values). ③ Check the coordinate jump link (click to navigate to the data display interface); 6.4 Multi-user collaborative viewing: Through the display module, multiple users can simultaneously view seismic profiles (arbitrary line cuts), borehole / tunnel base maps (layer overlay), and fault identification results (annotated parameters and factor values).

[0037] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for intelligent identification of coal mine faults, characterized in that, Includes the following steps: S1. Data Loading: Obtain the SGY seismic processing results data volume, borehole results data and coal mine roadway line data of the coal mining area, and load them into the preset processing environment; S2. Data Preprocessing: Perform signal optimization processing on the SGY seismic processing results data volume, perform coordinate adaptation processing on borehole data and coal mine roadway data, and construct a unified data space; S3. AI Model Construction and Training: Construct an intelligent fault recognition model, using data in a unified data space as the training basis, and enable the model to have the ability to extract and recognize fault features through iterative training; S4. Intelligent Fault Recognition: Input the preprocessed data of the area to be identified into the trained model, and output the preliminary fault recognition results; S5. Optimization of identification results: Extract the judgment factors from borehole data and coal mine roadway data, verify and correct the preliminary identification results through multi-level cross-judgment, and generate the final fault identification results; S6. Results Output and Correlation: The final fault identification results are correlated with the coal mine 3D seismic interpretation and modeling process, and the export of the results maps is supported; Step S4, intelligent identification of interrupted layers, is implemented according to the following steps: S4.1 Steps for Quantitative Extraction of Fault Features: Step 1: Extraction of in-phase shaft misalignment features: Extract the tracking deviation value of the in-phase shaft of adjacent channels. When the deviation value is ≥2 channels and the deviation value is maintained for 5 consecutive sampling points, it is determined to be an in-phase shaft misalignment feature. The second step is amplitude abrupt change feature extraction: capturing amplitude anomaly regions and calculating the amplitude gradient value of these regions using the formula: G=|A 当前采样点 -A 前一采样点 | / Sampling interval, when G ≥ twice the mean amplitude gradient of the seismic data in the work area, it is determined to be an amplitude mutation feature; The third step is waveform distortion feature extraction: calculate the waveform similarity between the target seismic trace and the five adjacent seismic traces. When the similarity is ≤0.6 and lasts for ≥10 sampling points, it is determined to be a waveform distortion feature. S4.2 Model Collaborative Validation Steps: Based on the local fault features of a single profile extracted by the CNN module, global spatial correlation verification is performed on the local features of at least three adjacent profiles. If a local feature exists only in a single profile, it is marked as an interfering feature and removed, retaining only the effective local features that correspond to multiple profiles; S4.3, Preliminary identification result output steps: The output includes preliminary identification results containing fault location coordinates, strike angle, and depth range, and meets the following accuracy constraints: Fault location coordinates: Planar coordinate error ≤ 1 trace, depth coordinate error ≤ 50ms; Direction angle: calculated based on vector fitting of at least 5 effective feature points, with accuracy controlled within ±1°; Depth range: Based on the seismic data sampling interval, the upper and lower limits of the depth range should have a deviation of ≤2m, and the confidence level of the results should be marked. In step S5, the rule for extracting the judgment factors is as follows: First-level core judgment factors and second-level auxiliary judgment factors are extracted from the borehole data. The first-level core judgment factors are: A1: Elevation of exposed top and bottom plates of the fault; A2: Depth of the borehole lithological transition interface; A3: Abrupt change value of borehole acoustic transit time; The secondary auxiliary judgment factor is: B1: Core fragmentation degree; B2: Abnormal borehole fluid pressure; Primary core decision factors and secondary auxiliary decision factors are extracted from coal mine roadway data. The primary core decision factors are: C1: Coordinates of the intersection of the tunnel and the fault; C2: The fault at the tunnel face reveals the direction; The secondary auxiliary judgment factor is: D1: Sudden change value in tunnel excavation speed; D2: Abnormal value of tunnel support resistance; D3: Displacement difference between the roadway rib and roof The priority of the first-level core decision factors is higher than that of the second-level auxiliary decision factors, and the priority is reflected by the weight coefficient. Among them, the weight of the first-level factors is 0.6, and the weight of the second-level factors is 0.

4. In step S5, the specific process of the multi-level cross decision is as follows: S5a. First-level cross decision: ① Fault depth verification: Calculate the difference |H - A1| between the fault depth H in the preliminary fault recognition result and A1. The depth threshold T1 is dynamically adjusted according to the coal seam thickness h. If |H - A1| ≤ T1, the depth verification passes. ② Fault strike verification: The angle between the initially identified fault strike vector and the strike vector corresponding to C2 is calculated using the vector dot product method. The angle threshold T2 is dynamically adjusted according to the tunnel length L. If the value is less than or equal to T2, the check passes. ③ If both the depth and strike verifications pass, enter S5c parameter correction; if any verification fails, start S5b second-level cross decision. S5b. Second-level cross decision: ① Calculation of borehole side support S1: If the depth interval of A2 is [A2- A2+ If H is included and A3≥T3, then in soft rock formations, T3=50μs / m; in hard rock formations, T3=100μs / m; and simultaneously B1≥50% or B2≥2MPa, then S1=1; otherwise, S1=0.3×(A3 / T3)+0.2×(B1 / 100%)+0.5×(B2 / 2MPa). ② Calculation of roadway side support score S2: If the perpendicular distance d from C1 to the preliminary recognized fault strike line ≤ T4, in nearly horizontal coal seams, T4 = 2m; in inclined coal seams, T4 = 5m, and D1 ≤ T5. In rock roadways, T5 = 0.3m / d; in coal roadways, T5 = 0.8m / d; D2 ≥ T6. When bolt-net support is used, T6 = 25MPa; when concrete lining support is used, T6 = 40MPa; at the same time, D3 ≥ 10mm, then S2 = 1; otherwise, S2 = 0.4×((T4 - d) / T4) + 0.3×((T5 - D1) / T5) + 0.2×(D2 / T6) + 0.1×(D3 / 10mm); ③ Calculation of comprehensive support score S: S = α×S1 + β×S2. The weight coefficients α and β are dynamically adjusted according to the geological complexity of the working area, and α + β = 1; The comprehensive threshold T7 is adjusted according to the fault development degree of the working area; ④ If S ≥ T7, it is determined that the preliminary recognition result is valid, and enter S5c parameter correction; if S < T7, it is marked as a suspected fault, and an artificial review prompt is output; S5c. Parameter correction: Based on the valid decision result, perform three-dimensional parameter calibration.

2. The intelligent identification method for coal mine faults according to claim 1, characterized in that, The specific implementation steps of data loading in step S1 are as follows: S1.

1. Batch loading step of SGY data in multiple working areas: Adopt a multi-thread parallel loading mode, and multiple working areas' SGY seismic processing result data volumes can be imported in batches at one time; Build a configurable-capacity data cache unit to locally store the frequently accessed SGY data during the loading process; S1.

2. Real-time verification step of data integrity: The first step is format verification: Verify the consistency between the trace header information of the SGY file and the preset SEG-Y standard format. If there are missing fields or format mismatches, it is directly marked as format exception; The second step is data volume verification: Compare the actual number of data traces in the SGY file with the number of traces marked in the trace header. If the deviation between the two is ≥ 1 trace, it is determined as data missing; only when both the format verification and the data volume verification pass can the data enter the subsequent processing link; S1.

3. Abnormal file processing step: The first step is error log generation: For the SGY files marked as format exception or data missing, automatically generate a structured error log, which needs to include the storage path of the abnormal file, the type of exception, and the specific location of the exception; The second step of the repair solution outputs targeted repairs for different anomaly types: if the data is missing, use the adjacent channel data interpolation algorithm for repair; if the format is abnormal, it is recommended to use a format conversion tool that adapts to the SEG-Y2.0 standard and provide a format correction parameter template.

3. The intelligent identification method for coal mine faults according to claim 1, characterized in that, In step S2, the signal optimization processing includes: wavelet threshold denoising and amplitude normalization; The coordinate adaptation process involves converting the geodetic coordinates of the borehole data and the center coordinates of the mine roadway lines into an Inline-Crossline-Time coordinate system consistent with the SGY seismic data volume, with a conversion error ≤ 0.5m.

4. The intelligent identification method for coal mine faults according to claim 1, characterized in that, In step S3, the intelligent tomography identification model is a hybrid model that integrates convolutional neural networks and Transformers: The CNN module uses 3 convolutional layers and 2 pooling layers to extract local fault features from seismic data. The Transformer module employs a 4-layer encoder to capture the global fault spatial correlation characteristics of seismic data. The training is based on seismic data samples marked with fault locations, fault exposure information in borehole data, and fault contact information in coal mine roadway data.

5. The intelligent identification method for coal mine faults according to claim 1, characterized in that, In step S5c, the specific process for parameter correction is as follows: S5c1. Fault Depth Calibration: Using A1 as the benchmark, if the initial identified depth H deviates from A1 by less than or equal to T1, A1 is directly taken as the final depth; if the deviation exceeds T1, linear interpolation correction is performed in conjunction with A2, and the depth error after correction is less than or equal to 0.5m. S5c2. Fault strike correction: Using C1 as the control point, the least squares method is used to fit the fault strike line. If there are multiple C1s, the residuals from each C1 to the initial strike line are calculated. Outlier C1s with residuals > 2 times the standard deviation are removed, and the strike line is refitted. After correction, the strike angle error is ≤ 0.5°. S5c3. Fault Dip Verification: ① Extract the dip angle of the lithological interface corresponding to A2 from the borehole logging data. 1; ②Based on the D3 fitting tendency angle 2. The fitting relationship is as follows: 2 = k × D³ + b; ③Calculation _avg=( 1+ 2) / 2, if | 1- 2|≤15°, take _avg is the final tendency angle; if | 1- 2|>15°, activate the AI ​​model to re-extract the dip characteristics of the in-phase axis of the 3×3 neighborhood seismic profile in this area, and obtain the AI-predicted dip angle. AI, ultimately favoring an angle _final=0.4× 1+0.3× 2+0.3× _AI, after correction, the tilt angle error is ≤2°; S5c4. Fault Displacement Calculation: Based on the calibrated depth, strike, and dip parameters, combined with the elevation difference between the upper and lower coal seams, the fault displacement is calculated with an accuracy of ≤0.1m.

6. The intelligent identification method for coal mine faults according to claim 1, characterized in that, In step S6, the step of associating the final fault identification results with the coal mine 3D seismic interpretation and modeling process includes: ① Linked verification: The final fault identification results are linked with the seismic horizon interpretation data to establish the spatial topological relationship between faults and horizons. When the horizon interpretation results are updated, a secondary verification of the fault identification results is automatically triggered to ensure the consistency of the spatial relationship between faults and horizons. ②Results export: Supports saving seismic time profile results in CGM format; ③ Manual verification assistance: The manual verification prompts include seismic profile screenshots of suspected fault areas, a comparison table of the specific values ​​and thresholds of each judgment factor, and the coordinates of the borehole / tunnel locations to be verified.

7. A coal mine fault intelligent identification system, used to execute the coal mine fault intelligent identification method according to any one of claims 1-6, characterized in that, include: Data loading module: Used to load SGY seismic processing results data, borehole data and coal mine roadway data, supports data format verification and batch import, and has a built-in data caching unit; Data preprocessing module: Connected to the data loading module, it is used to perform wavelet threshold denoising and amplitude normalization on the SGY seismic data volume, perform coordinate system transformation on borehole and tunnel line data, and construct a unified data space; AI intelligent recognition module: includes model building submodule, model training submodule, and inference recognition submodule, used to output preliminary tomographic recognition results; Results optimization module: Connects to the AI ​​intelligent recognition module, including multi-factor extraction unit, cross-judgment unit, and parameter correction unit, to generate the final tomographic recognition results; Results Output and Correlation Module: Connects to the results optimization module, used to establish the linkage between faults and stratigraphic levels, supports exporting CGM format result maps, and outputs fault interpretation files; Display module: Used to display seismic time profiles, borehole and tunnel baseline maps, and fault identification results, supporting multi-user collaborative viewing.

Images