Small fault prediction method based on large-scale fault distribution information in coal seam

Abstract

The invention relates to the technical field of fault detection in a coal mine, in particular to a small fault prediction method based on large-scale fault distribution information in a coal seam, which is characterized in that geometric features and spatial distribution features of a large-scale fault are analyzed based on actual fault data of the coal mine, and a small fault prediction result is obtained based on a mathematical distribution relation model of the large-scale fault and small fault distribution. Establishing a neural network prediction model of minor fault distribution, extracting a local cloud picture representing a working face area and an actual minor fault distribution cloud picture of the working face area on the basis of constructing a minor fault prediction distribution cloud picture of the whole mine, performing comparative analysis, establishing a loss function of a neural network, and training and optimizing the neural network. The prediction accuracy and robustness of the model in the target area are improved, high-precision prediction of the small fault distribution density of the working face area is achieved, advanced prediction support is provided for coal mine production, the sudden influence of the hidden fault on production is reduced, and the safety and reliability of mine production are effectively improved.

Description

Technical Field

[0001] The present invention relates to the technical field of fault detection in coal mines, and in particular to a method for predicting small faults based on large-scale fault distribution information in coal seams. Background Art

[0002] When small, previously unidentified faults are accidentally exposed during mine production, they can delay production and, in severe cases, trigger safety incidents such as gas outbursts and rock bursts. In recent years, as coal mining depths have increased, the threat posed by hidden small faults to mine safety has become increasingly significant. Therefore, the advanced prediction and forecasting of hidden faults has become a crucial task in coal mine production geology.

[0003] Traditional geophysical exploration methods mainly use geophysical principles to detect underground structures. For example, seismic exploration infers structural morphology by analyzing the reflection or refraction characteristics of artificially excited seismic waves at geological interfaces. Electrical exploration identifies anomalies by measuring the electric field distribution generated by current based on the differences in electrical properties (such as resistivity) between rocks, coal seams and structural belts. Alternatively, auxiliary methods such as ground-penetrating radar that detects shallow layers through high-frequency electromagnetic waves and well logging that directly measures the physical properties of rock formations in boreholes can be used.

[0004] Traditional geophysical methods are effective for detecting large-scale structures that significantly perturb the strata or physical fields. However, for small, hidden faults, due to their small size, potentially subtle differences in physical properties, and weak anomalous signals, they often fall below the resolution limit of traditional methods or are overwhelmed by noise, resulting in poor detection. Therefore, existing technologies for predicting small, hidden faults rely on traditional geophysical methods and empirical judgment, making them ineffective in addressing these small-scale, hidden geological problems.

[0005] Based on the problems in the prior art, the present invention provides a small fault prediction method based on the large-scale fault distribution information in the coal seam. Summary of the Invention

[0006] The purpose of the present invention is to provide a method for predicting small faults based on the distribution information of large-scale faults in coal seams, so as to solve the technical problem that conventional detection means in the prior art have poor detection effect on hidden small faults.

[0007] The technical solution of the present invention is: a method for predicting small faults based on the distribution information of large-scale faults in coal seams, comprising: collecting fault distribution data in coal seams, extracting fault characteristic information, and constructing a data set of large-scale faults and a data set of small faults; marking and dividing the actual working face of the mine to obtain several working faces, and based on the actual distribution of small faults in each working face area, drawing the actual cloud map Gi of the small fault distribution of each working face; according to the characteristic information of the large-scale faults, determining the probability distribution model of the influence of the large-scale faults on the surrounding small faults; based on the probability distribution model, building a neural network prediction model and defining a loss function; wherein, the input of the neural network prediction model is the characteristic information of the pre-processed large-scale faults, and the output is the standard deviation of the Gaussian distribution density of small faults around each large-scale fault; according to the standard deviation, constructing a predicted distribution cloud map under the influence of the large-scale faults of the entire mine, and extracting a locally enlarged predicted distribution cloud map Pi of the corresponding working face area; comparing and analyzing the actual distribution cloud map Gi of the working face with the predicted distribution cloud map Pi, and obtaining the correlation between the two, calculating the defined loss function, and optimizing the neural network prediction model.

[0008] Preferably, based on a two-dimensional plane geological map of the mine fault distribution, the fault distribution data in the coal seam is collected, and the extracted fault characteristic information includes the geometric characteristics, spatial distribution characteristics and influence range of the large fault, and the distribution information of the small fault. The fault characteristic information specifically includes: the position L of the fault, the position coordinates of the two ends of the fault line segment (X1, Y1) and (X2, Y2), the fall H of the fault, the dip angle θ, the azimuth angle Φ and the direction;

[0009] The extracted fault feature information is preprocessed, including normalization, noise filtering and feature standardization.

[0010] Preferably, the loss function of the neural network prediction model is defined as:

[0011]

[0012] Among them, Loss represents the loss value, N* represents the total number of working faces involved in calculating the correlation, σ represents the standard deviation of the neural network output, and R is the correlation between the actual distribution cloud map Gi of the small faults on the working face and the predicted distribution cloud map Pi.

[0013] Preferably, the correlation R between the actual distribution cloud map Gi of the working face of the training set and the predicted distribution cloud map Pi is calculated, and the value range of R is [0, 1], where 0 means there is no correlation between the two sets of maps, and 1 means the two sets of maps are completely consistent.

[0014] Preferably, when calculating the correlation between Pi and Gi, the comparison includes comparing the spatial distribution pattern of small faults, the outlines of high and low density areas, and the consistency of density gradient change trends displayed by the two.

[0015] Preferably, the correlation is calculated by an algorithm, and the algorithm used should have an attention mechanism or the ability to filter out low-pixel areas to maintain the original color ratio of the work surface prediction cloud map;

[0016] The correlation between the predicted cloud map and the actual cloud map is determined based on the consistency of the color gradient changes in the local area.

[0017] Preferably, the neural network prediction model is iteratively trained, and the standard deviation is output each time to iteratively update the loss function, and the optimization training is repeated until the loss function is reduced to below a set threshold, or the number of iterations exceeds a maximum upper limit, and finally the standard deviation parameters of each fault corresponding to the minimum loss are output;

[0018] Combined with the overall distribution prediction cloud map of small faults in the mine, the construction of a Gaussian distribution model for predicting the density of small faults based on large-scale faults is completed.

[0019] Preferably, based on the distribution of working faces in an actual mine, the working face area of the entire mine is marked to obtain N working faces, and the N working faces are divided into a training set and a test set. The training set is used to train the neural network prediction model, and the test set is used to evaluate the performance of the test model.

[0020] Based on the loss function of the actual distribution cloud map and the predicted distribution cloud map of each working surface in the test set, the performance of the entire model is evaluated to avoid overfitting and monitor the training progress. The operations include:

[0021] The input data corresponding to the working surface of the test set is fed into the trained model to calculate and extract the corresponding predicted distribution cloud map Pi of the small fault;

[0022] Compare the predicted cloud map with the actual small fault distribution cloud map Gi corresponding to the test set, and calculate the loss value and / or correlation R of each test working surface;

[0023] The results obtained are the overall performance indicators on the test set. The performance indicators include average loss value and average correlation. The indicators quantify the generalization ability of the model on unseen data and can be used to evaluate whether the model is overfitting by comparing with the performance indicators of the training set.

[0024] Preferably, based on the geometric characteristics and spatial distribution characteristics of the large-scale fault, the influence range of the large-scale fault on the surrounding small faults and the probability distribution model obeyed by the number distribution of small faults around the large fault are determined;

[0025] The probability distribution model is one of Gaussian distribution, binomial distribution, Weibull distribution, and Poisson distribution.

[0026] Preferably, Gaussian distribution is used as the probability distribution model of the impact of a large-scale fault on surrounding small faults, and a neural network prediction model is built. The mean of the Gaussian distribution is set as the position of the fault, which is expressed as a constant.

[0027] Compared with the prior art, the advantages of the present invention are:

[0028] (1) The present invention is based on actual fault data of mines, analyzes the geometric characteristics and spatial distribution characteristics of large-scale faults, and establishes a neural network prediction model for the distribution of small faults based on the mathematical distribution relationship model of large-scale faults and small faults. It does not directly rely on the weak physical signals generated by the detection of small faults themselves, but predicts their possible distribution areas and densities through statistical laws and machine learning models. The method is simple and clear.

[0029] In addition, combined with the automatic optimization capability of the neural network, on the basis of constructing the small fault prediction distribution cloud map of the entire mine, the local cloud map representing the working face area and the actual small fault distribution cloud map of the working face area are extracted for comparative analysis, the loss function of the neural network is established, and the neural network training is optimized, which improves the accuracy and robustness of the model's prediction in the target area (such as the working face), and realizes high-precision prediction of the distribution density of small faults in the working face area, providing advanced forecast support for coal mine production, reducing the sudden impact of hidden faults on production, and effectively improving the safety and reliability of mine production.

[0030] (2) The method is concise and clear. The steps can be divided into fault data acquisition and processing, model construction and training, and distribution prediction and evaluation, forming a systematic small fault prediction process. No additional equipment or new processes are required, effectively reducing the cost of small fault prediction. The prediction results can provide scientific guidance for working face layout and safe production, significantly improving production efficiency and safety.

[0031] (3) It is applied to intelligent mine production and is suitable for the prediction of hidden small faults in coal mine production. It has high practical application value and can further promote the informatization and precision development of coal mine geological work, providing strong support for mine disaster warning and prevention. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] The present invention will be further described below with reference to the accompanying drawings and embodiments:

[0033] Figure 1 This is a schematic diagram of the process flow of the small fault prediction method of the present invention;

[0034] Figure 2 This is a flowchart of the implementation of the small fault prediction method of the present invention;

[0035] Figure 3A schematic diagram of a mathematical calculation model using Gaussian distribution to describe the impact of a fault on the surrounding area in an embodiment of the present invention;

[0036] Figure 4 A schematic diagram of a local calculation process for forming a prediction model by superimposing Gaussian probability density fields of multiple large-scale fault influence areas in an embodiment of the present invention;

[0037] Figure 5 Schematic diagram of comparative analysis between predicted distribution cloud map and actual distribution cloud map. DETAILED DESCRIPTION

[0038] The present invention will be described in further detail below with reference to specific embodiments:

[0039] Minor faults are small faults (drops less than 3-5 meters) that do not interrupt coal seam continuity, have minimal impact on working face layout and mining, and are difficult to accurately detect using conventional geophysical methods. These faults are often discovered during the production process, have the most direct impact on coal production, and are the focus of prediction and forecasting in coal mine geology.

[0040] Faults larger than this range are considered large faults. Based on the large faults that can be detected in advance, the probability density of small fault development and high-risk areas are predicted, providing an effective reference for actual mining.

[0041] The present invention provides a small fault prediction method based on the large-scale fault distribution information in coal seams, which is suitable for coal seam geological information characterization and hidden fault prediction, and is particularly suitable for analyzing the impact of small fault distribution on production safety and disaster warning during coal mining.

[0042] In coal mine geological work, in order to improve the accuracy and reliability of small fault prediction, refer to the attached Figure 1 The method flow chart provided includes the acquisition and processing of large-scale fault data, the division of training set and test set, the design of loss function, the construction and training of neural network model, the prediction and evaluation of small fault distribution density, and the Figure 2 The overall construction flow chart provided, the specific steps and methods are as follows:

[0043] (1) Acquisition and processing of large-scale fault data;

[0044] Based on a two-dimensional geological map of the mine's fault distribution, data on the distribution of coal seam faults was collected. The coordinates (X1, Y1) and (X2, Y2) at the two ends of the fault segment were extracted. These coordinates corresponded one-to-one with the actual position of the fault in the geological coordinate system. The fault's drop (H) (in meters) and dip (θ) (in degrees) were extracted, and the azimuth (Φ) (in degrees) and directional characteristics of each fault were calculated. Based on the fault's size and impact on production, each fault was classified as a large-scale fault that can be monitored, and a small fault that is difficult to monitor using conventional geophysical methods. These data were then assigned labels. A complete fault dataset was constructed based on this data, preparing for subsequent analysis and model training.

[0045] The spatial distribution characteristics of large-scale faults, such as location, L, H, θ, and Φ, are extracted, and the characteristic data of large-scale faults are preprocessed, including normalization, noise filtering, and feature standardization, to ensure data quality.

[0046] Based on the actual distribution of mine faults, the geometric characteristics, spatial distribution characteristics and impact range of large faults are analyzed, and large-scale fault data of the entire mine are extracted to provide multi-dimensional data support for subsequent model training.

[0047] The formation and distribution of fault systems are inherently interconnected and regular, and the spatial pattern of large faults significantly controls the development of smaller faults. By acquiring large amounts of high-quality (incomplete and flawed data are eliminated) large-scale fault data, we reveal statistical distribution patterns, such as trends and superposition effects, and combine this with actual uncovered data from the working face to construct and optimize the prediction model.

[0048] (2) Acquisition and construction of small fault data training and test sets;

[0049] Small faults are mainly exposed during the working face mining process. Therefore, based on the actual working face distribution of the mine, the working face area of the entire mine is marked to obtain N working faces. Each working face is given a number #i (indicating the i-th working face). Then, the N working faces are divided into training sets and test sets in an m:n ratio to prepare for the subsequent training and verification of the neural network model.

[0050] The coordinates of the small fault positions in each working face area are extracted, including the distribution information of the small faults in the working face and the surrounding local areas, including the aforementioned characteristic information, as well as other characteristic information of the small faults themselves.

[0051] Based on the actual distribution of small faults in each working face area, that is, according to the actual distribution density of small faults in different areas of the working face, the actual distribution cloud map Gi of small faults in each working face is drawn to provide pseudo labels for subsequent neural network training.

[0052] (3) Construction and training of neural network models;

[0053] A regression prediction model is selected, guided by pseudo labels, that is, the distribution cloud map of small faults in the actual working face is used as the optimization target, and the loss function is defined by calculating the correlation between the predicted cloud map and the actual cloud map, so as to gradually optimize the model parameters and improve the prediction accuracy. Choosing a suitable network can obtain better prediction results.

[0054] For example, a multi-layer perceptron (MLP) is used to take the characteristics of each large-scale fault as input and predict the standard deviation σ of its corresponding Gaussian distribution. To ensure that the model focuses on the actual fault distribution area and is robust to small positional deviations, a loss function is designed based on the structural similarity index (SSIM) between the predicted cloud image Pi and the actual cloud image Gi, combined with a masking technique to ignore the influence of background areas.

[0055] Alternatively, a graph neural network (GNN) can be used to construct a graph structure of large-scale faults and their spatial relationships, which is then fed into the network to better learn how interactions between faults affect the distribution of smaller faults, ultimately predicting the σ for each fault. Furthermore, methods can be explored to extract image features using convolutional neural networks (CNNs) and construct an optimization objective based on perceptual loss to more precisely measure the structural and pattern similarity between the predicted and actual cloud images.

[0056] The distribution of the number of small faults around a large fault follows a specific distribution law, such as Gaussian distribution, binomial distribution, Weibull distribution, and Poisson distribution. Therefore, based on the geometric characteristics and spatial distribution characteristics of the large-scale fault, its impact range and probability distribution form on the surrounding small faults are determined. This embodiment uses Gaussian distribution as the probability distribution model for the impact of large-scale faults on surrounding small faults. That is, the Gaussian distribution is used to quantitatively describe the probabilistic impact field (i.e., probability density field) of a single fault on its surrounding area.

[0057] Among them, since the position of each large-scale fault is fixed, the mean of the Gaussian distribution is directly taken as the position of the fault, which is a constant.

[0058] Based on the Gaussian distribution probability model, a neural network prediction model is built. The input is the preprocessed large-scale fault characteristic data L, H, θ, Φ and the position of the fault. The output is the standard deviation of the Gaussian distribution density of small faults around each large-scale fault. The standard deviation is used as a variable to describe the variation range and density of the small fault distribution.

[0059] Refer to the attached Figure 3 The figure shows a schematic diagram of the mathematical model for calculating the shortest distance from a point to a fault segment in this embodiment. The figure shows the method for calculating the shortest distance from a point P(x, y) within a region to a fault segment AB, and combines it with a Gaussian distribution to describe the impact of the fault on the surrounding area.

[0060] Specifically, if Figure 3 As shown on the left, for any grid point P(x, y) in space and a known large (or medium or small) scale fault line segment AB, first calculate the shortest spatial distance d(x, y, AB) from the point (grid point) to the fault line segment AB. The formula for the shortest spatial distance is:

[0061]

[0062] Then, the impact intensity or probability density value Z(x,y) of the fault on the point P(x,y) is modeled as a Gaussian function (Gaussian kernel function) with the shortest distance d as the variable: Z(x,y) = e ^ (-d(x,y,AB) 2 / (2σ 2 In this Gaussian function model, the impact intensity z ranges from (0, 1]. When point P is located on the fault line, d = 0, and the impact intensity is maximum (z = 1). As the distance d from point P to the fault line increases, the impact intensity z decays exponentially according to the Gaussian curve.

[0063] The parameter σ (standard deviation) controls the "width" or "diffusion" of the impact range. The larger the σ value, the wider the impact range of the fault and the slower the decay; the smaller the σ value, the more concentrated the impact range and the faster the decay. In this embodiment, the parameter σ value is not fixed, but serves as the output of the neural network, that is, σ = f θ (x), x represents the characteristics of the fault, which is dynamically determined by the fault's own properties (such as length, drop, inclination, etc.) through the trained neural network.

[0064] Finally, as attached Figure 3 As shown on the right (the probability density field caused by the influence of a certain fault), by calculating the influence field based on Gaussian distribution for each large, medium and small scale fault listed in the mine (large, medium and small here are the classification of large scale faults), each fault obtains its own σ value according to its characteristics, and these individual influence fields are superimposed in space to generate a comprehensive reflection of the probability density prediction cloud map of the entire mine's micro-faults under the joint action of all faults.

[0065] The Gaussian prediction distribution cloud map of small faults under the influence of large-scale faults is controlled by the standard deviation obtained for each large-scale fault. If the distribution of large-scale faults is extremely uneven, when superimposing the influence of different faults, it is possible to consider limiting the upper limit of superposition to control the influence of different regions within a reasonable range.

[0066] Based on the position coordinates of each working surface circled in step (2), the local enlarged prediction cloud map of the corresponding working surface area can be extracted to obtain the prediction distribution cloud map Pi.

[0067] (4) Calculation of the correlation between the predicted distribution cloud map of small faults and the actual distribution cloud map;

[0068] According to the standard deviation of the output of the neural network prediction model, the Gaussian distribution prediction cloud map under the influence of the large-scale fault of the entire mine is first constructed to reflect the overall impact of the mine fault on the small fault, as shown in the attached figure. Figure 4 As shown, Figure 4 This demonstrates the Gaussian density calculation logic based on the fault's impact range. The left side shows the Gaussian probability density field for the impact of a single fault, while the right side shows the probability density fields for the impact ranges of large, medium, and small faults (large, medium, and small are all classifications of large-scale faults) calculated in this example. By calculating the Gaussian distribution of fault impacts at different scales and superimposing them in space, the final fault impact prediction model is formed.

[0069] The actual distribution cloud map Gi of the small faults of each working face is compared and analyzed with the predicted distribution cloud map Pi, and the correlation R between the actual distribution cloud map Gi and the predicted distribution cloud map Pi of the training set working face is calculated. The value range is [0, 1], where 0 means there is no correlation between the two sets of maps, and 1 means that the two sets of maps are completely consistent.

[0070] The local prediction distribution cloud map Pi extracted from the overall prediction cloud map has a color gradient and relative density range consistent with its corresponding part in the overall map.

[0071] However, the color scale (i.e., the mapping between color values and actual density values) of the actual distribution cloud map Gi generated based on actual small fault data is inherently and systematically different from Pi. It is important to emphasize that forced adjustments to the color scale should not be used to align Pi and Gi numerically or visually. Such alignment will distort the true relative density relationships within each cloud map and between different working surfaces (incorrectly amplifying or reducing the predicted density differences between different areas), leading to incorrect assessments of the model's prediction performance.

[0072] Therefore, when calculating the correlation or similarity between Pi and Gi, we focus on comparing the spatial distribution patterns of small faults, the outlines of high and low density areas, and the consistency of density gradient change trends shown by the two images, so as to judge the correlation between the two images, rather than relying on direct matching of pixel colors or the absolute density values they represent.

[0073] As attached Figure 5A schematic diagram of the comparative analysis of the Gaussian distribution prediction cloud map (left side) calculated based on this embodiment and the actual small fault distribution cloud map (right side) is provided. The two show significant correspondence and similar trends in the position, shape, extension direction of high-risk areas (high-density areas) and the distribution contours of low-risk areas in multiple working face areas. There is a high degree of consistency, which verifies the rationality and feasibility of this method.

[0074] (5) Calculation of loss function;

[0075] The loss function of the neural network is defined as:

[0076]

[0077] Where Loss represents the loss value, σ represents the standard deviation of the neural network output, R represents the correlation between the actual distribution cloud map Gi (representing the actual distribution cloud map of small faults in the i-th working face) and the predicted distribution cloud map Pi (representing the predicted distribution cloud map of small faults in the i-th working face), and N* represents the total number of working faces involved in the correlation calculation, which is the actual number of samples used to calculate the loss in the loss function calculation formula.

[0078] Based on the correlation R obtained in the previous step, the loss function is calculated and the neural network model is iteratively trained. The actual cloud map Gi and predicted cloud map Pi corresponding to the working surface of the training set are used to optimize and update the network parameters to improve the prediction accuracy.

[0079] (6) Prediction and evaluation of small fault distribution density;

[0080] Based on the loss function of the actual and predicted distribution cloud maps of each working surface in the test set, the performance of the entire model is evaluated, while avoiding overfitting and monitoring the training progress. Operations on the test set include:

[0081] The input data corresponding to the test set working face is fed into the trained model to generate the corresponding predicted distribution cloud maps Pi of small faults. These predicted cloud maps are then compared with the actual distribution cloud maps Gi of small faults corresponding to the test set, and the loss value and / or correlation R are calculated for each test working face. The key results obtained are overall performance metrics on the test set, such as average loss value or average correlation. These metrics quantify the model's generalization ability to unseen data and can be used to assess whether the model is overfitting by comparing it with the performance metrics of the training set.

[0082] If the training effect is not good, further consideration should be given to improving the division of the working face data set, supplementing the validation set, and dividing the working face into training set, validation set, and test set according to a certain ratio.

[0083] Repeat steps (4), (5) and (6) and continuously iterate and update until the loss function drops below the set threshold or the number of iterations exceeds the maximum upper limit, and finally output the standard deviation parameters of each fault corresponding to the minimum loss.

[0084] Based on the geometric characteristics and spatial distribution information of large-scale faults, the present invention describes the impact of large-scale faults on small faults through a mathematical distribution model, combines the actual fault data of the mine to train a neural network model, predicts the distribution of small faults in the working face area, and combines the automatic optimization capability of the neural network to achieve high-precision prediction of the distribution density of small faults in the working face area, providing a scientific basis for safe production and geological survey of coal mines.

[0085] Rather than relying directly on detecting the weak physical signals generated by small faults themselves, this method uses statistical laws and machine learning models to predict their likely distribution and density. This concise and clear approach consists of a systematic process: fault data acquisition and processing, model building and training, and distribution prediction and evaluation. Its application in intelligent mine production can further promote the informatization and precision of coal mine geological work, providing strong support for mine disaster warning and prevention.

[0086] Compared with traditional small fault prediction methods, this method leverages existing mine geological data to more accurately identify the distribution density of small faults. It does not rely on complex geophysical methods, making it particularly suitable for practical application in coal mine engineering. It does not require additional equipment or new processes, effectively reducing the cost of small fault prediction. The prediction results can provide scientific guidance for working face layout and safe production, significantly improving production efficiency and safety.

[0087] The above embodiments are only for illustrating the technical concept and features of the present invention, and their purpose is to enable people familiar with this technology to understand the content of the present invention and implement it accordingly, and they are not intended to limit the scope of protection of the present invention. For those skilled in the art, it is obvious that the present invention is not limited to the details of the above exemplary embodiments, and the present invention can be implemented in other specific forms without departing from the spirit or basic characteristics of the present invention. Therefore, no matter from which point of view, the embodiments should be regarded as exemplary and non-restrictive. The scope of the present invention is defined by the appended claims rather than the above description, and it is intended that all changes that fall within the meaning and scope of the equivalent elements of the claims are included in the present invention.

Claims

1. A method for predicting small faults based on the distribution information of large-scale faults in coal seams, characterized in that: include: Collect coal seam fault distribution data, extract fault characteristic information, and construct large-scale fault data sets and small fault data sets; Mark and divide the actual working face of the mine to obtain several working faces. Based on the actual distribution of small faults in each working face area, draw the actual cloud map Gi of the small fault distribution of each working face; Based on the characteristic information of the large-scale fault, a probability distribution model of the impact of the large-scale fault on the surrounding small faults is determined; based on the probability distribution model, a neural network prediction model is built and a loss function is defined; wherein the input of the neural network prediction model is the preprocessed characteristic information of the large-scale fault, and the output is the standard deviation of the Gaussian distribution density of the small faults around each large-scale fault; According to the standard deviation, a prediction distribution cloud map under the influence of large-scale faults in the entire mine is constructed, and a locally enlarged prediction distribution cloud map Pi of the corresponding working face area is extracted; The actual distribution cloud map Gi of the working face is compared and analyzed with the predicted distribution cloud map Pi, and the correlation between the two is obtained. The defined loss function is calculated to optimize the neural network prediction model.

2. A method for predicting small faults based on large-scale fault distribution information in coal seams according to claim 1, characterized in that: Based on the two-dimensional geological map of the mine fault distribution, the fault distribution data in the coal seam is collected. The extracted fault characteristic information includes the geometric characteristics, spatial distribution characteristics and influence range of large faults, as well as the distribution information of small faults. The fault characteristic information specifically includes: the fault location L, the position coordinates of the two ends of the fault line segment (X1, Y1) and (X2, Y2), the fault height H, the dip angle θ, the azimuth angle Φ and the direction; The extracted fault feature information is preprocessed, including normalization, noise filtering and feature standardization.

3. The method for predicting small faults based on large-scale fault distribution information in coal seams according to claim 2, characterized in that: The loss function of the neural network prediction model is defined as: Among them, Loss represents the loss value, N* represents the total number of working faces involved in calculating the correlation, σ represents the standard deviation of the neural network output, and R is the correlation between the actual distribution cloud map Gi of the small faults on the working face and the predicted distribution cloud map Pi.

4. The method for predicting small faults based on large-scale fault distribution information in coal seams according to claim 3, characterized in that: Calculate the correlation R between the actual distribution cloud map Gi of the working face in the training set and the predicted distribution cloud map Pi. The value range of R is [0, 1]. 0 means there is no correlation between the two sets of maps, and 1 means the two sets of maps are completely consistent.

5. The method for predicting small faults based on large-scale fault distribution information in coal seams according to claim 4, characterized in that: When calculating the correlation between Pi and Gi, the comparison includes: comparing the spatial distribution pattern of small faults, the outlines of high and low density areas, and the consistency of density gradient change trends shown by the two.

6. The method for predicting small faults based on large-scale fault distribution information in coal seams according to claim 4, characterized in that: Calculate the correlation through an algorithm. The algorithm used should have an attention mechanism or be able to filter out low-pixel areas to maintain the original color ratio of the work surface prediction cloud map; The correlation between the predicted cloud map and the actual cloud map is determined based on the consistency of the color gradient changes in the local area.

7. The method for predicting small faults based on large-scale fault distribution information in coal seams according to claim 1, characterized in that: Iteratively train the neural network prediction model, output the standard deviation each time, iteratively update the loss function, and repeat the optimization training until the loss function drops below the set threshold, or the number of iterations exceeds the maximum limit, and finally output the standard deviation parameters of each fault corresponding to the minimum loss; Combined with the overall distribution prediction cloud map of small faults in the mine, the construction of a Gaussian distribution model for predicting the density of small faults based on large-scale faults is completed.

8. The method for predicting small faults based on large-scale fault distribution information in coal seams according to claim 7, characterized in that: Based on the distribution of working faces in the actual mine, the working face area of the entire mine is marked to obtain N working faces. The N working faces are divided into a training set and a test set. The training set is used to train the neural network prediction model, and the test set is used to evaluate the performance of the test model. Based on the loss function of the actual distribution cloud map and the predicted distribution cloud map of each working surface in the test set, the performance of the entire model is evaluated to avoid overfitting and monitor the training progress. The operations include: The input data corresponding to the working surface of the test set is fed into the trained model to calculate and extract the corresponding predicted distribution cloud map Pi of the small fault; Compare the predicted cloud map with the actual small fault distribution cloud map Gi corresponding to the test set, and calculate the loss value and / or correlation R of each test working surface; The results obtained are the overall performance indicators on the test set. The performance indicators include average loss value and average correlation. The indicators quantify the generalization ability of the model on unseen data and can be used to evaluate whether the model is overfitting by comparing with the performance indicators of the training set.

9. The method for predicting small faults based on large-scale fault distribution information in coal seams according to claim 1, characterized in that: Based on the geometric characteristics and spatial distribution characteristics of large-scale faults, determine the impact range of large-scale faults on surrounding small faults and the probability distribution model that the number of small faults around the large fault obeys; The probability distribution model is one of Gaussian distribution, binomial distribution, Weibull distribution, and Poisson distribution.

10. The method for predicting small faults based on large-scale fault distribution information in coal seams according to claim 9, characterized in that: Gaussian distribution is used as the probability distribution model of the impact of large-scale faults on surrounding small faults, and a neural network prediction model is built. The mean of the Gaussian distribution is set as the position of the fault, which is expressed as a constant.

Images