Coal mine area geological disaster sample enhancement and unbalanced dataset processing method based on generative adversarial network

Abstract

The present application relates to the field of coal mine safety monitoring and data processing, and discloses a coal mine area geological disaster sample enhancement and unbalanced data set processing method based on a generative adversarial network, comprising: acquiring multi-source geological monitoring data and performing spatial alignment, and extracting and fusing feature vectors through a multi-head cross-modal attention mechanism; calculating local density values of disaster samples, identifying sparse precursor samples and decision boundary samples as a high-value sample set, and adaptively determining sampling weights; taking the high-value sample set and the sampling weights as prior conditions, and constructing a generative adversarial network to generate enhanced samples; using real disaster samples to construct a disaster class high-dimensional feature space statistical envelope, performing envelope compliance testing on the enhanced samples, and removing invalid samples; and mixing the effective generated samples and the original samples to construct a balanced training set for training a geological disaster classification model. The present application effectively solves the problem of disaster sample scarcity under the condition of extremely unbalanced coal mine monitoring data, reduces the false positive rate, and improves the disaster early warning capability.

Description

Technical Field

[0001] This invention relates to the field of coal mine safety monitoring and data processing technology, specifically to a method for enhancing geological hazard samples and processing imbalanced datasets in coal mining areas based on generative adversarial networks. Background Technology

[0002] In geological disaster monitoring in coal mining areas, multi-source sensing devices such as time-series InSAR (Synthetic Aperture Radar Interferometry), stress-strain sensors, and groundwater level gauges continuously collect data on deformation and stress status in the mining area. However, these monitoring data exhibit a significant imbalance in distribution: routine, safe operation data constitute the vast majority, while deformation or stress anomaly data containing actual disaster evolution characteristics are extremely scarce.

[0003] Existing technologies typically employ direct oversampling (such as SMOTE and its variants) or basic generative adversarial networks (GANs) to handle this type of imbalanced data, but these methods suffer from the following technical bottlenecks:

[0004] (1) The fusion depth of multi-source heterogeneous data is insufficient and the spatial correspondence is ambiguous. Temporal InSAR data is a spatially continuous two-dimensional grid sequence, while stress and strain sensors are discrete single-point time series. The two differ significantly in data dimension, spatial coverage, time sampling rate and physical semantics. Existing fusion methods mostly stop at simple splicing of data layers or direct concatenation of feature layers. They fail to fully extract cross-modal correlations and lack reasonable processing of the spatial correspondence between InSAR area observations and downhole point sensors, resulting in limited characterization ability of fused features.

[0005] (2) Lack of high-value sample identification mechanism. The occurrence of geological disasters has a complex gestation process, and early atypical deformation precursors are also "sparse features" in minority class samples. Existing generation methods perform indiscriminate oversampling of minority class samples, resulting in the generated samples overfitting to typical disaster patterns and losing key early evolutionary sparse features (i.e., intra-class pattern collapse). At the same time, "boundary samples" located in the transition zone between safe and disaster states are crucial for constructing the decision boundary of the classification model, but existing methods lack systematic identification and targeted enhancement strategies for this type of sample.

[0006] (3) There are constraints between variables in multi-source geological monitoring data determined by geomechanical laws. For example, under the same geological conditions, an acceleration in the rate of surface subsidence is usually accompanied by a corresponding adjustment in the rate of change of deep stress. Existing data interpolation or random generation methods are prone to generating cross-boundary pseudo-samples that break the above physical laws, resulting in a high false alarm rate in subsequent monitoring models.

[0007] (4) Existing GAN methods suffer from technical bias and insufficient scenario adaptability. For example, CN121351926A discloses a method for processing imbalanced data based on dynamic density-guided generative adversarial networks, which achieves imbalanced data processing through kernel density estimation and VAE latent structure learning. However, it is mainly geared towards general tabular data and does not consider the spatial alignment and fusion problem of multi-source heterogeneous data from InSAR and downhole sensors in coal mine geological monitoring. There have been studies on applying GANs to 3D modeling of coal mine borehole data (e.g., CN121304958A) and prediction of water-conducting fracture zone height, but these methods focus on the generation of specific types of data and do not systematically solve the problems of collaborative enhancement of multi-source monitoring data and verification of geological consistency by category.

[0008] In summary, existing technologies have not simultaneously addressed the three core issues in coal mine geological disaster monitoring: spatial alignment of multi-source heterogeneous data, identification of sparse precursor samples, and verification of physical constraint consistency. This invention, through a collaborative mechanism of "spatial alignment fusion → weighted identification of high-value samples → generation of conditional constraints → classification envelope verification," achieves for the first time an end-to-end imbalanced sample augmentation method for multi-source InSAR and sensor data in coal mining areas. Summary of the Invention

[0009] The purpose of this invention is to provide a method for enhancing geological disaster samples and processing imbalanced datasets in coal mining areas based on generative adversarial networks. Specifically, it is a method for constructing a sample library for monitoring geological disasters (such as goaf subsidence and slope landslides) in coal mining areas. Through multi-source data fusion representation with spatial interpolation alignment, weighted identification of high-value samples based on local density statistics, and a collaborative mechanism of conditional constraint generation and classification statistical envelope verification, it achieves effective enhancement of disaster samples and improvement of classification model performance under extremely imbalanced data conditions.

[0010] To achieve the above objectives, the following technical solution is adopted:

[0011] This invention provides a method for enhancing geological hazard samples and processing imbalanced datasets in coal mining areas based on generative adversarial networks, comprising the following steps: S1: Acquiring temporal InSAR deformation data and underground stress-strain sensor temporal data of the mining area and performing spatial alignment preprocessing; extracting deep features from the aligned temporal InSAR deformation data and underground stress-strain sensor temporal data respectively, and performing feature interaction fusion through a multi-head cross-modal attention mechanism to obtain a fused feature vector; S2: Mapping the fused feature vector to a multi-dimensional feature space, calculating the local density value of each hazard sample, and identifying and labeling sparse precursor samples and decision boundary samples as a high-value sample set based on the local density value and nearest neighbor category information. S3: Using the high-value sample set and its corresponding sampling weights as prior conditions, a conditional Wasserstein generative adversarial network with gradient penalty is constructed to learn the feature distribution of each disaster type and generate enhanced samples; S4: A high-dimensional feature space statistical envelope of the disaster class is constructed using historical real disaster samples. The enhanced samples generated in step S3 are checked for compliance with the disaster class envelope, and invalid generated samples that exceed the statistical boundary of the disaster class are removed; S5: The valid generated samples retained after validity verification are mixed with the original real samples to construct a balanced training set for training the geological disaster classification model in the mining area.

[0012] In a preferred embodiment, step S1, the spatial alignment preprocessing specifically includes: using inverse distance weighted interpolation, taking each InSAR observation point as the center, calculating a weighted value based on the reciprocal of the square of the distance between the downhole stress-strain sensor measurement points within a preset initial search radius and the InSAR observation point, to obtain the sensor aggregated feature value of the InSAR observation point; when there are no effective sensors within the initial search radius, compensation is performed using nearest neighbor sensors or by gradually expanding the search radius according to a preset step size; for each sensor sequence, extracting multiple statistical features within the time window corresponding to the InSAR deformation observation period, the statistical features including mean, standard deviation, rate of change, maximum difference value, and first-order autocorrelation coefficient.

[0013] In a preferred embodiment, step S1, which involves extracting deep features from aligned temporal InSAR deformation data and downhole stress-strain sensor temporal data respectively, and performing feature interaction fusion through a multi-head cross-modal attention mechanism, specifically includes: using a first encoder to extract a first deep feature vector from the aligned temporal InSAR deformation data, wherein the first encoder is a multi-layer two-dimensional convolutional neural network; using a second encoder to extract a second deep feature vector from the aligned downhole stress-strain sensor temporal data, wherein the second encoder is a multi-layer fully connected network; using the first deep feature vector as a query and the second deep feature vector as a key and value, inputting them into the multi-head cross-modal attention module for interactive fusion, calculating scaled dot product attention, and concatenating the results output by multiple attention heads and performing a linear transformation to obtain the fused feature vector.

[0014] In a preferred embodiment, in step S1, the two independent encoders used to extract deep features are optimized for parameters using a supervised contrastive loss function during the pre-training stage. The supervised contrastive loss function uses class labels to construct positive and negative sample pairs, so that sample features belonging to the same class in the same training batch are close to each other, and sample features belonging to different classes are far apart.

[0015] In a preferred embodiment, step S2, calculating the local density value of each disaster sample specifically includes: determining the nearest neighbor parameter k according to the total number of disaster samples based on a preset empirical rule; for each disaster sample, calculating the Euclidean distance between it and the k-th nearest neighbor sample, and using this distance as an adaptive bandwidth parameter; and using a Gaussian kernel function to calculate the local density value within the k-nearest neighbor set of the disaster sample, wherein the k-nearest neighbor set includes the disaster sample itself.

[0016] In a preferred embodiment, step S2, which involves identifying and labeling sparse precursor samples and decision boundary samples as a high-value sample set and adaptively determining the sampling weight of each disaster sample, specifically includes: labeling samples whose local density is lower than a preset quantile of the density distribution of all disaster samples as sparse precursor samples based on the local density value of each disaster sample; labeling samples whose safe sample ratio in their k nearest neighbors exceeds a preset threshold as decision boundary samples based on the nearest neighbor category information of each disaster sample; designating disaster samples that are not labeled as sparse precursor samples or decision boundary samples as typical disaster samples; and calculating the sampling weight of each disaster sample based on the degree of deviation between the local density value of each disaster sample and the median density of all disaster samples, and whether the sample is a decision boundary sample, wherein sparse precursor samples and decision boundary samples are assigned a higher sampling weight than typical disaster samples.

[0017] In a preferred embodiment, in step S3, the conditional Wasserstein generative adversarial network with gradient penalty includes a generator and a discriminator; the generator's input is a concatenation vector of random noise and a conditional vector, and its output is a generated enhanced sample feature vector; the discriminator's input is a concatenation of a real or generated feature vector and a conditional vector, and spectral normalization is applied after each fully connected layer; the discriminator's loss function includes a gradient penalty term; the generator's loss function includes a feature matching loss term, which is the expected distance between the generated sample and a real sample of the same weight type in the discriminator's intermediate layer features; the number of generated enhanced samples is adaptively determined according to the target balance ratio, which is the ratio of the total number of disaster samples expected to be enhanced to the original number of safe samples.

[0018] In a preferred embodiment, step S4, constructing the high-dimensional feature space statistical envelope of the disaster class, specifically includes: using only the fused feature vectors of historical real disaster samples, calculating their covariance matrix and mean vector; using a shrinkage estimation method to regularize the covariance matrix to obtain a regularized covariance matrix; performing eigenvalue decomposition on the regularized covariance matrix, extracting principal components whose cumulative variance contribution rate exceeds a preset threshold, and constructing a low-dimensional effective subspace; calculating the Mahalanobis distance of each real disaster sample in the low-dimensional effective subspace, and setting the preset high quantile of the Mahalanobis distance of all real disaster samples as the disaster class constraint radius.

[0019] In a preferred embodiment, step S4 specifically includes the following steps for verifying the compliance of the disaster class envelope: for each generated pseudo-disaster sample, project it into the low-dimensional effective subspace and calculate its Mahalanobis distance; compare the calculated Mahalanobis distance with the disaster class constraint radius; if the Mahalanobis distance is less than or equal to the disaster class constraint radius, then the pseudo-disaster sample is determined to be a valid generated sample and retained; otherwise, it is determined to be an invalid generated sample and discarded.

[0020] In a preferred embodiment, in step S5, the geological hazard classification model for the mining area is a multi-layer fully connected network, and its last layer uses the Softmax activation function to perform binary classification of the input feature vector into safe and hazard states.

[0021] Compared with the prior art, the present invention has the following beneficial effects:

[0022] (1) Discrete downhole sensor data are mapped to a continuous InSAR observation grid using spatial interpolation. In areas without sensor coverage, nearest neighbor or extended radius compensation strategies are adopted to establish a spatial correspondence between surface deformation and deep stress. This correspondence is approximately valid under certain geological assumptions and can be used for statistical feature correlation modeling. Based on this, a dual-branch encoder and a multi-head cross-modal attention mechanism are used to effectively extract cross-modal correlation features between the two types of data, thereby improving the representation quality of the fused features.

[0023] (2) By calculating the local density based on kernel density estimation and combining the neighbor class confounding analysis, it can automatically identify sparse precursor samples within the geological disaster class and transitional samples located in the classification boundary area. By using a differentiated weighting strategy based on density distribution quantiles and neighbor proportions (the threshold can be adaptively adjusted according to the mining area data), the generator is guided to focus on enhancing these key sample areas that improve the classifier's performance, effectively alleviating the problem of intra-class pattern collapse.

[0024] (3) A classification statistical envelope verification mechanism is introduced, and covariance regularization technology is used to ensure the stability of covariance estimation under small sample conditions. Only real disaster samples are used to construct the effective region of the disaster class feature space. The pseudo samples generated by GAN are subjected to class consistency test, and erroneous samples that exceed the statistical boundary of the disaster class are automatically removed to avoid the feature patterns of safe samples being mistakenly generated as disaster samples. Furthermore, a geomechanical constraint regularization term can be optionally introduced to enhance the compatibility of the generated samples with physical laws, thereby reducing the false alarm rate of subsequent classification.

[0025] (4) This invention can be started without relying on a large number of labeled disaster samples. The feature extraction network of S1 can obtain effective representation capabilities through supervised contrastive learning on hundreds of safety samples. S2 only needs dozens of disaster samples to perform sample value assessment based on local density. The number of generated samples can be adaptively adjusted according to the actual scale of safety samples, and it has practical deployability under extremely unbalanced data conditions in coal mines.

[0026] It should be understood that the description in the Summary of the Invention is not intended to limit the key or essential features of the embodiments of the present invention, nor is it intended to restrict the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0027] The above and other features, advantages, and aspects of the various embodiments of the present invention will become more apparent from the accompanying drawings and the following detailed description. The drawings are provided for a better understanding of the invention and are not intended to limit the scope of the invention. In the drawings, the same or similar reference numerals denote the same or similar elements, wherein:

[0028] Figure 1This is a flowchart of the method for enhancing geological hazard samples and processing imbalanced datasets in coal mining areas based on generative adversarial networks, according to an embodiment of the present invention.

[0029] Figure 2 This is a schematic diagram illustrating the principle of multi-source data spatial alignment in an embodiment of the present invention;

[0030] Figure 3 This is a schematic diagram illustrating high-value sample identification based on local density according to an embodiment of the present invention;

[0031] Figure 4 This is a schematic diagram of the generative adversarial network structure according to an embodiment of the present invention;

[0032] Figure 5 This is a performance comparison chart of the classification models in embodiments of the present invention. Detailed Implementation

[0033] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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.

[0034] Furthermore, the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0035] Figure 1 This is a flowchart of a method for enhancing geological hazard samples and processing imbalanced datasets in coal mining areas based on generative adversarial networks, according to an embodiment of the present invention. For details, see [link to relevant documentation]. Figure 1 A method for enhancing geological hazard samples and processing imbalanced datasets in coal mining areas based on generative adversarial networks includes the following steps:

[0036] S1: Acquire temporal InSAR deformation data and downhole stress-strain sensor temporal data of the mining area and perform spatial alignment preprocessing; extract deep features from the aligned temporal InSAR deformation data and downhole stress-strain sensor temporal data respectively, and perform feature interaction fusion through multi-head cross-modal attention mechanism to obtain fused feature vector.

[0037] Step S1 is used to achieve spatial alignment and fusion feature extraction of multi-source geological monitoring data.

[0038] This embodiment selects 18 months of subsidence monitoring data from a coal mine (mining depth of about 450m). Among them, there are about 8,500 samples of safe status, and only 47 samples of subsidence disaster confirmed by geological engineers through on-site investigation and historical records (the ratio of positive to negative samples is about 1:180), which is a typical high imbalance dataset.

[0039] S1.1: Data acquisition and spatial alignment preprocessing.

[0040] 1. Collect the following two types of data:

[0041] (1) Temporal InSAR Deformation Data: 37 Sentinel-1 SAR images covering the mining area (spanning 18 months) were acquired. SBAS-InSAR processing was used to obtain the average deformation rate map and cumulative deformation time series of the mining area's surface line of sight. The InSAR processing parameters were as follows: a temporal baseline threshold of 120 days, a spatial baseline threshold of 200 meters, interferometric filtering using the Goldstein filtering method, stable bedrock areas as reference points, and GACOS data for atmospheric phase correction. The spatial resolution was approximately 20m × 20m. 47 known subsidence hazard locations and 8500 safe locations were selected, and a 15×15 pixel deformation rate window centered on each location was used as the InSAR input data.

[0042] (2) Time-series data of downhole stress and strain sensors: Time-series data of 12 borehole stress gauges and 6 anchor bolt force gauges deployed downhole were acquired, with a sampling frequency of 1 time / hour, for a total of 18 sensors. The location coordinates of each sensor (independent coordinate system of the mining area) are known, but they are all discrete point distributions and do not have a one-to-one correspondence with the InSAR surface observation points.

[0043] 2. Spatial alignment preprocessing:

[0044] like Figure 2 The diagram illustrates the principle of spatial alignment of multi-source data in an embodiment of the present invention. In this embodiment, the spatial alignment preprocessing includes: mapping discrete sensor data to InSAR observation points using spatial interpolation methods to construct spatially aligned multi-source data pairs; and compensating for the absence of effective sensors within the search radius using a nearest neighbor or extended radius adaptive strategy. The specific process is as follows:

[0045] To spatially match point sensor data with area InSAR observation points, inverse distance-weighted interpolation (IDW) is employed. Centered on each InSAR observation point, the aggregated sensor feature value for that point is calculated by weighting the data with the inverse square of the distance between all sensor points within a preset initial search radius (e.g., 100m), i.e., downhole stress-strain sensor points and InSAR observation points. When no effective sensor is found within the search radius, compensation is achieved using a nearest neighbor sensor strategy or by gradually expanding the radius to a maximum of 300m in preset step sizes (e.g., 50m), ensuring that each InSAR observation point has usable aggregated stress features. For each sensor sequence, multiple statistical features are extracted within a time window (e.g., a continuous 30-day time window) corresponding to the InSAR deformation observation period. Preferably, the following five statistical features are calculated: mean, standard deviation, rate of change (difference between first and last values / time span), maximum difference, and first-order autocorrelation coefficient. After IDW spatial interpolation, a 90-dimensional (18 sensors × 5 features) stress aggregation feature vector is obtained for each InSAR observation point. This vector represents the comprehensive state of deep stress within a certain range below the surface location, which is approximately valid under certain geological assumptions and is used for statistical feature correlation modeling.

[0046] S1.2: Construction and training of multi-source data fusion feature extraction network.

[0047] In step S1.2, the aligned temporal InSAR deformation data and downhole stress-strain sensor temporal data are processed by two independent encoder branches to extract deep features. These features are then fused using a multi-head cross-modal attention mechanism to obtain a fused feature vector. Specifically, this involves: using a first encoder to extract a first deep feature vector from the aligned temporal InSAR deformation data (the first encoder is a multi-layer two-dimensional convolutional neural network); using a second encoder to extract a second deep feature vector from the aligned downhole stress-strain sensor temporal data (the second encoder is a multi-layer fully connected network); using the first deep feature vector as a query and the second deep feature vector as a key and value, inputting them into the multi-head cross-modal attention module for interactive fusion; calculating scaled dot product attention; and concatenating the results from multiple attention heads and performing a linear transformation to obtain the fused feature vector. The details are as follows:

[0048] A fusion feature extraction network is constructed, with the following specific structure:

[0049] 1) InSAR branch encoder: A 3-layer 2D convolutional neural network (CNN) is used, with kernel sizes of 3×3, 3×3, and 3×3, and channel numbers of 32, 64, and 128 respectively. Each layer is followed by batch normalization and ReLU activation, and finally, global average pooling is used to output a 256-dimensional feature vector. .

[0050] 2) Stress branch encoder: The input is the 90-dimensional stress aggregation feature vector obtained from S1.1. A two-layer fully connected network is used, with 128 and 256 neurons respectively. Each layer is followed by Dropout (the Dropout rate is...). ReLU activation is used to output a 256-dimensional feature vector. .

[0051] 3) Multi-head cross-modal attention module: integrates InSAR features and stress characteristics The input is processed through a multi-head attention layer for interactive fusion. Let the number of attention heads be... The dimensions of each head are For the first Size, with Linear transformation as query ,by Linear transformation as a key Sum Calculate the scaled dot product attention :

[0052] ;

[0053] Will The outputs of each feature vector are concatenated, then compressed using a linear transformation and a fully connected layer to obtain the final fused feature vector. This structure allows for targeted information aggregation of InSAR features guided by stress features, while stress features provide a deep state reference for InSAR features. In this embodiment, [the following is taken]... , The fused output features are 512-dimensional and compressed to 128-dimensional.

[0054] 4) Network pre-training strategy: The two independent encoders used for extracting deep features undergo parameter optimization using a supervised contrastive loss function during the pre-training phase. Specifically:

[0055] Considering there are only 47 disaster samples in total, supervised contrastive loss is used for pre-training. Unlike SimCLR's unsupervised contrastive learning, supervised contrastive loss directly constructs positive and negative sample pairs using class labels: samples belonging to the same class within the same batch attract each other, while samples from different classes repel each other. The supervised contrastive loss function is:

[0056] ;

[0057] in, : Supervised comparison loss function value; The set of indices for all samples in the current training batch; The first in the batch One sample index; : with sample A collection of indexes of other samples belonging to the same category; : with sample The number of positive samples of the same type; : Positive sample set The sample index in; :sample The fused feature vector output by the feature extraction network; Positive samples fused feature vectors; : within the batch except for samples Any sample outside The fused feature vector, , , These are the feature vectors output by the fusion feature extraction network for the current sample, the positive sample, and other samples in the batch, respectively. Temperature coefficient, with a value of 0.1, is used to adjust the degree of concentration of feature similarity distribution. : within the batch except for samples The set of all sample indexes except those mentioned above.

[0058] The Adam optimizer was used during training, with an initial learning rate of [missing information]. The batch size is 128 (with oversampling to ensure each batch contains at least 2 disaster samples), and training is performed for 200 epochs until the loss converges. Supervised contrastive loss is more stable than SimCLR under small sample conditions because it directly constructs positive and negative pairs using class labels, without relying on random augmentation of a large number of negative samples.

[0059] S2: Map the fused feature vector to a multi-dimensional feature space, calculate the local density value of each disaster sample, identify and mark sparse precursor samples and decision boundary samples as a high-value sample set based on the local density value and neighbor category information, and adaptively determine the sampling weight of each disaster sample based on density distribution statistics and neighbor category mixing.

[0060] Step S2 is used to achieve weighted identification of high-value samples based on local density estimation. The fused feature vector obtained in step S1 is mapped to a multi-dimensional feature space, the local density value of each disaster sample is calculated, and sparse precursor samples and decision boundary samples are identified and labeled based on category label information. Sampling weights are adaptively determined based on density distribution statistics and the confounding effect of neighboring categories. The relevant threshold is determined through cross-validation or based on statistical experience using historical data from the mining area, and can be adaptively adjusted according to the data distribution of different mining areas. Figure 3The diagram illustrates the identification of high-value samples based on local density according to an embodiment of the present invention. Specifically, it includes the following steps:

[0061] S2.1: Local density calculation and adaptive parameter determination.

[0062] In step S2.1, calculating the local density value of each disaster sample specifically includes: for each disaster sample, calculating the Euclidean distance between it and its k-th nearest neighbor sample, and using this distance as an adaptive bandwidth parameter; using a Gaussian kernel function to calculate the local density value within the k-nearest neighbor set of the disaster sample. The specific process is as follows:

[0063] For each disaster sample (47 in total), calculate its local density in the fused feature space. The specific steps are as follows:

[0064] (1) k-nearest neighbor parameters Determination: Based on the total number of disaster samples In this embodiment, an appropriate number of nearest neighbors is selected based on empirical rules. =47. It is the parameter of the number of nearest neighbors, according to Based on empirical rules, the value is 7 in this embodiment.

[0065] (2) Calculate the relationship between each disaster sample and its first sample. The distance between nearest neighbors is denoted as This distance is taken as the adaptive bandwidth parameter. .

[0066] (3) Use Gaussian kernel function to calculate samples Local density :

[0067] ;

[0068] in, : No. The feature vector of each disaster sample in the fused feature space; :sample Its first Euclidean distance between nearest neighbor samples; :sample The adaptive bandwidth parameter of the Gaussian kernel function takes the value equal to... ; :sample The local density estimate; :sample of The nearest neighbor sample set contains itself; : The first in Feature vectors of each sample; Euclidean distance norm operator.

[0069] S2.2: Classification of high-value samples and adaptive weighting.

[0070] In step S2.2, sparse precursor samples and decision boundary samples are identified and labeled as a high-value sample set, and the sampling weight of each disaster sample is adaptively determined. Specifically, this includes: based on the local density value of each disaster sample, marking samples with a local density lower than a preset quantile of the density distribution of all disaster samples as sparse precursor samples; based on the nearest neighbor category information of each disaster sample, marking samples whose proportion of safe samples in their k-nearest neighbors exceeds a preset threshold as decision boundary samples; designating disaster samples not labeled as sparse precursor samples or decision boundary samples as typical disaster samples; and calculating the sampling weight of each disaster sample based on the degree of deviation between the local density value of each disaster sample and the median density of all disaster samples, and whether the sample is a decision boundary sample, wherein sparse precursor samples and decision boundary samples are assigned a higher sampling weight than typical disaster samples. The specific process is as follows:

[0071] Based on local density values ​​and neighbor category information, disaster samples are divided into three categories:

[0072] 1) Sparse precursor samples: local density values Samples below the preset quantile of the density distribution of all disaster samples are considered to belong to a sparsely distributed "outlier" pattern in the disaster sample set, corresponding to atypical / early deformation precursor characteristics. In this embodiment, we take... Quantiles, a total of 8.

[0073] 2) Decision boundary samples: Samples whose proportion of outliers (safe samples) in their nearest neighbors exceeds a preset threshold are considered to be located in a transitional zone between safe and disaster states, and are of high value for constructing classification boundaries. In this embodiment, we take... Thresholds, a total of 12 are marked.

[0074] 3) Typical disaster samples: the remaining samples, a total of 27.

[0075] The aforementioned quantile thresholds and outlier percentage thresholds were determined through cross-validation or based on statistical experience using historical data from the mining area, and can be adaptively adjusted according to the data distribution of different mining areas. Sampling weights were assigned to the three types of samples respectively. This serves as a conditional guiding factor in the subsequent GAN generation process. Weight allocation follows the principle of high-value samples receiving high weights, employing a composite weighting based on density quantile shift and boundary confounding. sampling weights The specific calculation formula is as follows:

[0076] ;

[0077] in, The weighting coefficient of the density quantile shift term. The weighting coefficient of the boundary confusion indicator. , The data can be determined through grid search on the validation set based on the characteristics of the mining area data. In this embodiment, the data was initially verified and selected... , The sampling weights of sparse precursor samples are obtained. ≈3.0, sampling weight of decision boundary samples ≈2.5, sampling weight of typical disaster samples =1.0. :sample The sampling weights are used as conditional guiding factors for subsequent generative adversarial networks. :sample The local density value. : The median of the local density values ​​of all disaster samples. : The minimum local density value of all disaster samples. : The function to find the maximum value. : Indicator function, takes the value 1 when the condition in parentheses is true, otherwise takes the value 0. :sample The boundary determination condition, when of The condition is established when the proportion of safe samples in the nearest neighbors exceeds a preset threshold.

[0078] S3: Using the high-value sample set and its corresponding sampling weights as prior conditions, construct a conditional Wasserstein generative adversarial network with gradient penalty, learn the feature distribution of each disaster type and generate enhanced samples respectively;

[0079] Step S3 is used to generate disaster sample features under conditional constraints. Using the high-value sample set and its sampling weights obtained in S2 as prior conditions, a Conditional Wasserstein Generative Adversarial Network (CWGAN-GP) with gradient penalty is constructed to learn the feature distribution for each sub-disaster type and generate enhanced samples. Optionally, a regularization term based on geomechanical constraints is introduced during the generation process to constrain the physical relationships between variables; the number of generated samples is adaptively determined according to the target balance ratio.

[0080] S3.1: Network structure design.

[0081] like Figure 4The diagram shows a schematic of the generative adversarial network structure according to an embodiment of the present invention. A conditional Wasserstein GAN with gradient penalty (CWGAN-GP) is constructed, including a generator and a discriminator. The generator's input is a concatenation vector of random noise and a conditional vector, and its output is the generated enhanced sample feature vector. The discriminator's input is a concatenation of the real or generated feature vector and the conditional vector, and spectral normalization is applied after each fully connected layer. The discriminator's loss function includes a gradient penalty term. The generator's loss function includes a feature matching loss term, which is the expected distance between the generated sample and the real sample of the same weight type in the discriminator's intermediate layer features. The number of generated enhanced samples is adaptively determined according to the target balance ratio, which is the ratio of the total number of disaster samples to the original number of safe samples after expected enhancement. The specific structure is as follows:

[0082] 1) Generator G: Input is random noise (Scaling parameter adjusted variance) and conditional vector (Including disaster type labels and sampling weights determined in step S2) The concatenated vectors are then processed by a multi-layer fully connected network to generate feature vectors. Each layer uses batch normalization and LeakyReLU activation. In this embodiment, the noise dimension is 128, resulting in 130 dimensions after concatenation. The fully connected layer is... .

[0083] 2) Discriminator D (i.e., Critic): The input is either the real or generated feature vector and the conditional vector. The concatenation of the data is processed through a multi-layer fully connected network to output the Wasserstein distance estimate. To ensure training stability, spectral normalization is applied after each fully connected layer of the discriminator, with LeakyReLU as the activation function. In this embodiment, the fully connected layer is... .

[0084] S3.2: Loss Function and Training Strategy

[0085] Discriminator loss function:

[0086] ;

[0087] in, : Discriminator loss function value; : with conditional vector The output function of the discriminator is conditional; From the generation distribution Generate sample feature vectors from the mid-sample; From the true distribution The feature vector of the real sample sample in the middle; From interpolation distribution The feature vector of the randomly interpolated sample is obtained by randomly interpolating between the real sample and the generated sample. Generate sample feature distribution; : Distribution of features of real samples; : The random interpolation distribution between real samples and generated samples; : Conditional vector, containing disaster type labels and sampling weight information; Mathematical expectation operator; Gradient penalty coefficient, with a value of 10; The discriminator outputs the interpolated samples. The gradient; Norm operators.

[0088] Generator loss function (combining feature matching terms guided by conditional weights, with optional physical constraint regularization terms):

[0089] ;

[0090] in, : Generator loss function value; Feature matching loss term; Geomechanical constraint regularization term; : The weight coefficient of the feature matching loss term, which is 0.1 in this embodiment; : The weighting coefficient of the geomechanical constraint regularization term, which is 0.05 in this embodiment.

[0091] Among them, feature matching loss Defined as the expected L2 distance between generated samples and real samples of the same weight type in the intermediate layer features of the discriminator, it is used to stabilize the generation quality and make the generation distribution converge towards the target type. The weights are used for feature matching loss.

[0092] For geomechanical constraint regularization terms, the specific definition is:

[0093] ;

[0094] : The weighting coefficient within the geomechanical constraint regularization term; in this embodiment, the value is 1. : Generate the feature components corresponding to the surface subsidence rate in the feature vector; : Generate the characteristic components corresponding to the rate of change of stress in the characteristic vector; Geological constants determined by geomechanical experiments in the mining area; in this embodiment, the value is taken as 1.2. : Absolute value operator.

[0095] It should be noted that when reliable prior knowledge of geomechanics is lacking in the mining area, it can be... When set to 0, the geomechanical constraint regularization term is disabled, and the generator's loss function only includes adversarial loss and feature matching loss. Those skilled in the art can implement this variant without creative effort based on the above description. In this embodiment, the existing geological report of the mining area is used... , and set , .

[0096] Training uses the Adam optimizer, and the generator learning rate is... Discriminator learning rate The discriminator is updated every 5 times, and the generator is updated once. A total of 5000 iterations are trained.

[0097] S3.3: Determine the adaptive generation quantity.

[0098] For the three classes of high-value samples marked in step S2, apply the corresponding weight conditional vectors. Generate augmented samples. The number generated is adaptively determined based on the target balance ratio: let the original number of safe samples be... The original disaster sample size was The expected total number of disaster samples after enhancement is 100%. ,in, The target balance ratio. In this embodiment... , ,Pick Number of target disaster samples The required number of samples is 850-47=803. The generation quota is allocated according to the original quantity ratio of the three types of samples (8:12:27): 150 sparse precursor samples, 220 boundary samples, and 433 typical samples.

[0099] S4: Construct a high-dimensional feature space statistical envelope for disaster classes using historical real disaster samples, perform a disaster class envelope compliance check on the enhanced samples generated in step S3, and remove invalid generated samples that exceed the statistical boundaries of disaster classes;

[0100] Step S4 is used to verify the effectiveness of features based on the categorical statistical envelope. High-dimensional feature space statistical envelopes for disaster classes are constructed using historical real disaster samples. To improve the stability of covariance estimation, regularization is performed using shrinkage estimation or diagonal loading methods. The enhanced samples generated in S3 (as pseudo-disaster samples) are then subjected to a disaster class envelope compliance check, and invalid generated samples exceeding the statistical boundaries of disaster classes are removed. Specifically, the steps are as follows:

[0101] S4.1: Construction of categorical statistical envelope.

[0102] In step S4.1, constructing the high-dimensional feature space statistical envelope of disaster classes using historical real disaster samples specifically includes: calculating the covariance matrix and mean vector using only the fused feature vectors of historical real disaster samples; regularizing the covariance matrix using a contraction estimation method to obtain a regularized covariance matrix; performing eigenvalue decomposition on the regularized covariance matrix to extract principal components whose cumulative variance contribution rate exceeds a preset threshold, and constructing a low-dimensional effective subspace; calculating the Mahalanobis distance of each real disaster sample in the low-dimensional effective subspace, and setting the preset high quantile of the Mahalanobis distances of all real disaster samples as the disaster class constraint radius. The specific process is as follows:

[0103] Since safety samples and disaster samples have different distribution areas in the feature space, constructing a constraint envelope by mixing them will blur the statistical structure boundary between the two classes. Therefore, this invention adopts a class-based statistical envelope construction strategy.

[0104] (1) Construction of disaster envelope: Only the fused feature vectors of real disaster samples are used to calculate their covariance matrix. and mean vector To improve the stability of covariance estimation, a shrinkage estimation method is used to regularize the covariance matrix.

[0105] ;

[0106] in, The original covariance matrix of the fused feature vectors of real disaster samples; : The mean vector of the fused feature vectors of real disaster samples; : The covariance matrix after shrinkage estimation and regularization; Shrinkage strength parameter, with a value range of: ; The average of the diagonal elements of the original covariance matrix, i.e., the average characteristic variance; :and An identity matrix of the same dimension.

[0107] in, For shrinkage strength, The average characteristic variance, It is an identity matrix. Shrinkage strength. The following formula can be used to estimate:

[0108] ;

[0109] Frobenius norm operator; Covariance matrix The Middle Line number Column elements; elements of the covariance matrix The sample variance estimate; : Position of all elements in the covariance matrix The summation operator.

[0110] Eigenvalue decomposition is performed on the regularized covariance matrix, and principal components with cumulative variance contribution rates exceeding a preset threshold are extracted to construct a low-dimensional effective subspace. The Mahalanobis distance of each real disaster sample in the principal component subspace is calculated.

[0111] ;

[0112] : The feature vector of the sample for which the Mahalanobis distance is to be calculated; :sample Mahalanobis distance relative to the statistical envelope of disaster categories; The transpose operator for vectors or matrices; Regularized covariance matrix The inverse matrix.

[0113] The preset high quantile (95th quantile in this embodiment) of the Mahalanobis distance of all real disaster samples is taken as the disaster class constraint radius. .

[0114] (2) Construction of security class envelope (optional): Using the same method, construct the security class statistical envelope using security samples to obtain the security class mean. and radius This is used for subsequent false alarm analysis.

[0115] S4.2: Generate sample verification.

[0116] In step S4.2, the compliance check of the disaster envelope specifically includes: for each generated pseudo-disaster sample, projecting it into a low-dimensional effective subspace and calculating its Mahalanobis distance; comparing the calculated Mahalanobis distance with the disaster class constraint radius; if the Mahalanobis distance is less than or equal to the disaster class constraint radius, the pseudo-disaster sample is determined to be a valid generated sample and retained; otherwise, it is determined to be an invalid generated sample and discarded. Specifically:

[0117] For each candidate disaster feature sample generated in step S3 Calculate its Mahalanobis distance relative to the disaster class envelope. Retain satisfaction Samples that meet the statistical distribution rules of disaster categories are considered valid augmented samples; samples that exceed the boundaries are considered invalid samples that deviate from the disaster category characteristic area during the generation process and are removed.

[0118] In this embodiment, after verification by S4, 708 out of 803 candidate samples were retained (retention rate 88.2%). The 95 samples that were removed exhibited characteristics closer to the safety class envelope region in the feature space, meaning they generated cross-boundary samples that violated the feature distribution of the disaster class. This step effectively prevented the risk of mistakenly generating disaster samples from the feature patterns of safety class samples.

[0119] S5: Mix the valid generated samples retained after validity verification with the original real samples to construct a balanced training set for training the geological hazard classification model in the mining area.

[0120] Step S5 is used to construct a balanced sample library and train the classification model. The valid generated samples verified in step S4 are mixed with the original real samples to construct a balanced training set, which is used to train the geological hazard classification model for the mining area, and outputs engineering evaluation indicators including false alarm rate and early warning time.

[0121] The 708 valid generated samples retained in step S4 are mixed with the original training samples (8500 safe samples and 47 real disaster samples) to construct an enhanced training set.

[0122] A geological hazard classification model for mining areas is constructed using a multi-layer fully connected network, with the last layer employing the Softmax activation function. Preferably, in this embodiment, a 3-layer fully connected network (128→64→2) with Softmax output is used to perform a safety / hazard binary classification on the input feature vector. Classifiers with the same structure are trained before and after enhancement, maintaining consistent training parameters (Adam optimizer, learning rate...). (Batch size 64, early stop number 20).

[0123] Test set evaluation: The test set includes independent validation samples from the original data that were not used in the training. These samples were independently verified and labeled by geological engineers and include 500 safe samples and 12 hazardous samples. To evaluate the stability of the method, five replicate experiments were conducted (each time the training / validation / test sets were randomly divided to maintain a consistent distribution of hazardous samples across all sets). The mean index and standard deviation are reported. Engineering evaluation metrics are also supplemented.

[0124] It should be noted that the specific parameter values ​​(such as k value, quantile threshold, weight coefficient, etc.) in the above embodiments are all determined based on the data distribution characteristics of this embodiment. In practical applications, those skilled in the art can adaptively adjust the above parameters through cross-validation or empirical rules according to the data distribution characteristics of different mining areas. These adjustments do not depart from the protection scope of this invention.

[0125] (1) Performance comparison of classification models Figure 5 As shown in Table 1:

[0126] Table 1. Performance Comparison Results of Classification Models

[0127]

[0128] Figure 5 (a) shows the comparison of precision, recall, F1 score, AUC, and false alarm rate before and after sample augmentation; Figure 5 (b) shows the comparison of the early warning time (days) before and after sample enhancement.

[0129] The early warning time indicator is defined as follows: taking the date of disaster occurrence confirmed by geological engineers as the baseline, the time interval between the date on which the model output probability first exceeds the early warning threshold and the date of disaster occurrence is calculated. In this embodiment, the average early warning time of the enhanced model is 3.6 days better than that of the unenhanced model, indicating that the enhanced sample library better covers early deformation characteristics, enabling the model to identify disaster evolution trends earlier.

[0130] (2) Ablation experiment:

[0131] To verify the effectiveness of each core module, ablation experiments were conducted under the same data partitioning, and the results are shown in Table 2 below:

[0132] Table 2 Ablation Experiment Results

[0133]

[0134] The results show that the high-value sample weighting module (S2) made the most significant contribution to the improvement of recall, verifying the necessity of targeted enhancement of sparse precursors and boundary samples; the removal of the classification envelope verification (step S4) caused the precision to decrease from 0.84 to 0.74 and the false positive rate to increase from 16% to 26%, directly proving the key role of this module in reducing the false positive rate; the removal of the cross-modal attention mechanism caused the F1 score to decrease by 0.04, indicating that deep fusion of multi-source data has a practical contribution to improving the quality of feature representation.

[0135] In summary, the present invention provides a method for enhancing geological hazard samples and processing imbalanced datasets in coal mining areas based on generative adversarial networks. By spatial alignment and deep fusion of multi-source monitoring data, identification of high-value samples based on statistical distribution, feature generation under conditional constraints, and verification of classification geological constraints, the method effectively enhances hazard samples and improves the performance of classification models under extremely imbalanced data conditions.

[0136] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. 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.

[0137] It should also be noted that, in the embodiments of this application, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0138] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined in the embodiments of this application may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown in this application, but is to be accorded the widest scope consistent with the principles and novel features disclosed in the embodiments of this application.

Claims

1. A method for enhancing geological hazard samples and processing imbalanced datasets in coal mining areas based on generative adversarial networks, characterized in that, Includes the following steps: S1: Acquire temporal InSAR deformation data and downhole stress-strain sensor temporal data of the mining area and perform spatial alignment preprocessing; Deep features were extracted from aligned time-series InSAR deformation data and downhole stress-strain sensor time-series data, respectively, and feature interaction fusion was performed through a multi-head cross-modal attention mechanism to obtain a fused feature vector. S2: Map the fused feature vector to a multi-dimensional feature space, calculate the local density value of each disaster sample, identify and mark sparse precursor samples and decision boundary samples as a high-value sample set based on the local density value and neighbor category information, and adaptively determine the sampling weight of each disaster sample based on density distribution statistics and neighbor category mixing. S3: Using the high-value sample set and its corresponding sampling weights as prior conditions, construct a conditional Wasserstein generative adversarial network with gradient penalty, learn the feature distribution of each disaster type and generate enhanced samples respectively; S4: Construct a high-dimensional feature space statistical envelope for disaster classes using historical real disaster samples, perform a disaster class envelope compliance check on the enhanced samples generated in step S3, and remove invalid generated samples that exceed the statistical boundaries of disaster classes; S5: Mix the valid generated samples retained after validity verification with the original real samples to construct a balanced training set for training the geological hazard classification model in the mining area.

2. The method for enhancing geological hazard samples and processing imbalanced datasets in coal mining areas based on generative adversarial networks according to claim 1, characterized in that, In step S1, the spatial alignment preprocessing specifically includes: The inverse distance weighted interpolation method is adopted. Taking each InSAR observation point as the center, the weighted value is calculated based on the reciprocal of the square of the distance between the downhole stress and strain sensor measuring points within the preset initial search radius and the InSAR observation point, so as to obtain the sensor aggregated feature value of the InSAR observation point. When there is no effective sensor within the initial search radius, compensation is performed by using the nearest neighbor sensor or by gradually expanding the search radius according to a preset step size. For each sensor sequence, multiple statistical features are extracted within the time window corresponding to the InSAR deformation observation period. These statistical features include mean, standard deviation, rate of change, maximum difference, and first-order autocorrelation coefficient.

3. The method for enhancing geological hazard samples and processing imbalanced datasets in coal mining areas based on generative adversarial networks according to claim 1 or 2, characterized in that, In step S1, the extraction of deep features from the aligned time-series InSAR deformation data and the downhole stress-strain sensor time-series data, and the feature interaction and fusion through a multi-head cross-modal attention mechanism, specifically includes: A first deep feature vector is extracted from aligned temporal InSAR deformation data using a first encoder, which is a multi-layer two-dimensional convolutional neural network. A second deep feature vector is extracted from the aligned downhole stress-strain sensor time series data using a second encoder, which is a multi-layer fully connected network. The first deep feature vector is used as the query, and the second deep feature vector is used as the key and value. They are input into the multi-head cross-modal attention module for interactive fusion, scaling dot product attention is calculated, and the results output by multiple attention heads are concatenated and linearly transformed to obtain the fused feature vector.

4. The method for enhancing geological hazard samples and processing imbalanced datasets in coal mining areas based on generative adversarial networks according to claim 3, characterized in that, In step S1, the two independent encoders used to extract deep features are optimized for parameters using a supervised contrastive loss function during the pre-training phase; The supervised contrastive loss function uses class labels to construct positive and negative sample pairs, so that sample features belonging to the same class in the same training batch are close to each other, and sample features belonging to different classes are far apart.

5. The method for enhancing geological hazard samples and processing imbalanced datasets in coal mining areas based on generative adversarial networks according to claim 1, characterized in that, In step S2, calculating the local density value of each disaster sample specifically includes: The nearest neighbor parameter k is determined according to the total number of disaster samples and a preset empirical rule. For each disaster sample, the Euclidean distance between it and the k-th nearest neighbor sample is calculated and used as an adaptive bandwidth parameter. A Gaussian kernel function is used to calculate the local density value in the k-nearest neighbor set of the disaster sample, where the k-nearest neighbor set includes the disaster sample itself.

6. The method for enhancing geological hazard samples and processing imbalanced datasets in coal mining areas based on generative adversarial networks according to claim 1, characterized in that, In step S2, identifying and labeling sparse precursor samples and decision boundary samples as a high-value sample set, and adaptively determining the sampling weight of each disaster sample, specifically includes: Based on the local density values ​​of each disaster sample, samples with local densities lower than the preset quantile of the density distribution of all disaster samples are marked as sparse precursor samples; Based on the nearest neighbor category information of each disaster sample, samples whose proportion of safe samples in their k nearest neighbors exceeds a preset threshold are marked as decision boundary samples; Disaster samples that are not labeled as sparse precursor samples and decision boundary samples are used as typical disaster samples; Based on the degree of deviation between the local density value of each disaster sample and the median density of all disaster samples, and whether the sample is a decision boundary sample, the sampling weight of each disaster sample is calculated in combination, where sparse precursor samples and decision boundary samples are assigned higher sampling weights than typical disaster samples.

7. The method for enhancing geological hazard samples and processing imbalanced datasets in coal mining areas based on generative adversarial networks according to claim 1, characterized in that, In step S3, the conditional Wasserstein generative adversarial network with gradient penalty includes a generator and a discriminator; The generator takes as input a concatenated vector of random noise and a conditional vector, and outputs as a generated enhanced sample feature vector. The input to the discriminator is the concatenation of real or generated feature vectors and condition vectors, and spectral normalization is used after each fully connected layer. The loss function of the discriminator includes a gradient penalty term; The loss function of the generator includes a feature matching loss term, which is the expected distance between the generated sample and the real sample of the same weight type in the intermediate layer features of the discriminator; The number of enhanced samples generated is adaptively determined based on the target balance ratio, which is the ratio of the total number of disaster samples to the original number of safe samples after enhancement.

8. The method for enhancing geological hazard samples and processing imbalanced datasets in coal mining areas based on generative adversarial networks according to claim 1, characterized in that, In step S4, constructing the high-dimensional feature space statistical envelope of the disaster class specifically includes: The covariance matrix and mean vector are calculated using only the fused feature vectors of historical real disaster samples; The covariance matrix is ​​regularized using a shrinkage estimation method to obtain a regularized covariance matrix. Eigenvalue decomposition is performed on the regularized covariance matrix to extract principal components whose cumulative variance contribution rate exceeds a preset threshold, and a low-dimensional effective subspace is constructed. In the low-dimensional effective subspace, the Mahalanobis distance of each real disaster sample is calculated, and the preset high quantile of the Mahalanobis distance of all real disaster samples is set as the disaster class constraint radius.

9. The method for enhancing geological hazard samples and processing imbalanced datasets in coal mining areas based on generative adversarial networks according to claim 8, characterized in that, In step S4, the compliance check of the disaster-type envelope specifically includes: For each generated pseudo-disaster sample, project it onto the low-dimensional effective subspace and calculate its Mahalanobis distance; Compare the calculated Mahalanobis distance with the hazard class constraint radius; If the Mahalanobis distance is less than or equal to the disaster class constraint radius, the pseudo-disaster sample is determined to be a valid generated sample and retained; otherwise, it is determined to be an invalid generated sample and discarded.

10. The method for enhancing geological hazard samples and processing imbalanced datasets in coal mining areas based on generative adversarial networks according to claim 1, characterized in that, In step S5, the geological hazard classification model for the mining area is a multi-layer fully connected network, and its last layer uses the Softmax activation function to perform binary classification of the input feature vector into safe and hazard states.

Images