Coal mine radar signal abnormal body identification method based on improved generative adversarial network

Abstract

The invention relates to the technical field of coal mine safety, and discloses a coal mine radar signal anomalous body identification method based on an improved generative adversarial network, which comprises the following steps: S1, acquiring radar signals covering the geology inside a mine in real time, and constructing a multi-band radar signal set; s2, carrying out the preprocessing of the multi-band radar signal set; s3, constructing a double-layer generative adversarial network model in which a structured feature extraction module and a double attention module are introduced, and generating abnormal body feature information data; step S4, generating exception report data based on the exception body feature information data, wherein the exception report data comprises data of the position, the type and the property of the exception body; according to the method, the double-layer generative adversarial network model is introduced, and the structured feature extraction module and the double attention module are combined, so that features of abnormal bodies can be clearer and separable through layered extraction and processing aiming at common structural anomalies and material anomalies in the coal mine environment, and the recognition accuracy and precision are improved.

Description

Technical Field

[0001] This invention relates to the field of coal mine safety technology, and more specifically, to a method for identifying anomalies in coal mine radar signals based on an improved generative adversarial network. Background Technology

[0002] As coal mining depths increase and mine environments become increasingly complex, mine safety monitoring and disaster early warning have become crucial for ensuring safe coal mine production. Radar signal technology, as an important mine monitoring method, is widely used in geological exploration and safety monitoring in coal mines due to its ability to penetrate geological layers and provide high-precision information on underground structures. However, the complex and variable coal mine environment means that radar signals are affected by various factors during propagation, such as geological conditions, equipment noise, and electromagnetic interference, leading to a decline in signal quality and consequently affecting the accuracy of anomaly identification.

[0003] The existing technology has the following drawbacks:

[0004] First, existing methods for identifying anomalies in coal mine radar signals primarily rely on traditional signal processing and statistical analysis techniques. These methods typically identify potential anomalies by preprocessing radar signals, extracting features, and performing rule-based or model-based anomaly detection. However, due to the unique characteristics of the coal mine environment, traditional methods have many shortcomings when dealing with complex and variable signals. For example, traditional filtering and denoising methods struggle to effectively eliminate the frequent non-Gaussian noise and strong interference signals present in the coal mine environment, resulting in poor signal quality after processing. Second, rule-based or model-based anomaly detection methods depend on pre-set thresholds or hypothetical models. These thresholds or models may not be effectively adapted to the complex changes in the coal mine environment in practical applications, leading to missed or false detections of anomalies.

[0005] Secondly, while deep learning-based anomaly detection methods have been applied in some fields, such as image recognition and speech recognition, directly applying these methods to coal mine radar signal processing and anomaly identification still faces challenges. Firstly, coal mine radar signals involve massive amounts of data with complex features. Existing deep learning models, such as convolutional neural networks (CNNs) or recurrent neural networks (RNNs), are prone to overfitting when processing this data, resulting in insufficient generalization ability. Furthermore, most existing deep learning models focus on extracting global features, while performing poorly in identifying hidden local anomalies in radar signals, leading to low accuracy in anomaly identification.

[0006] Third, Generative Adversarial Networks (GANs), as an important branch of deep learning, have gradually gained attention in the field of anomaly detection due to their advantages in data generation and feature extraction. Through adversarial training between the generator and discriminator, GANs enable the generator to produce samples that closely resemble the distribution of real data, thus exhibiting unique advantages in data augmentation and anomaly detection. However, traditional GAN ​​structures still face many problems when applied to coal mine radar signal processing. For example, when processing high-dimensional and complex data, the training process of the generator and discriminator in a single-layer GAN model is prone to instability, resulting in low-quality generation results. Furthermore, the feature extraction capabilities of traditional GANs are mainly concentrated on global features, while their ability to extract and separate local features is weak, making it difficult to effectively identify anomalies in radar signals.

[0007] Therefore, given the shortcomings and deficiencies of existing technologies for radar signal processing and anomaly identification in coal mines, there is an urgent need for a new technical method that can better adapt to the complexity of the coal mine environment and improve the accuracy and robustness of radar signal anomaly identification. Summary of the Invention

[0008] The purpose of this invention is to provide a method for identifying anomalies in coal mine radar signals based on an improved generative adversarial network in order to solve the above-mentioned problems.

[0009] This invention provides a method for identifying anomalous entities in coal mine radar signals based on an improved generative adversarial network, comprising the following steps:

[0010] Step S1: Acquire radar signals covering the geology inside the mine in real time and construct a multi-band radar signal set;

[0011] Step S2: Preprocess the multi-band radar signal set;

[0012] Step S3: Construct a two-layer generative adversarial network model that incorporates a structured feature extraction module and a dual attention module. The two-layer generative adversarial network model is used to process the preprocessed multi-band radar signal set and generate anomaly feature information data.

[0013] Step S4: Generate anomaly report data based on the anomaly feature information data. The anomaly report data includes data on the location, type, and nature of the anomaly.

[0014] As a further optimization of the present invention, step S1 includes the following specific steps:

[0015] Step S11: Deploy multi-band radar equipment at multiple locations within the mine, and use the multi-band radar equipment to collect multi-band radar signals. Set the collection time to T, and within time T, collect signals at a sampling frequency... Signals were collected from different locations inside the mine to form a multi-band radar signal set. Where n represents the number of radar signals collected, and m represents the number of collection points within the mine. This represents the i-th multi-band radar signal acquired at the j-th acquisition point;

[0016] Step S12: For multi-band radar signals from different acquisition points Record the frequency distribution of the signal ,in, Representing frequency, constructing a frequency distribution set for multi-band radar signals. ;

[0017] Step S13: Based on the collected multi-band radar signal set and frequency distribution set This forms the final multi-band radar signal set. , .

[0018] As a further optimization of the present invention, step S2 includes the following specific steps:

[0019] Step S21: For the final multi-band radar signal set... Denoising processing is performed on the multi-band radar signal set using an adaptive denoising algorithm. Noise is identified and removed from the signal, and the denoised multi-band radar signal set is recorded as follows: ;

[0020] Step S22: Process the denoised multi-band radar signal set Normalization is performed using a linear normalization method to scale the signal amplitude range to the [0,1] interval. The normalized multi-band radar signal set is denoted as... ;

[0021] Step S23: Process the normalized multi-band radar signal Multi-stage filtering is performed, using bandpass filters to eliminate low-frequency and high-frequency interference components, resulting in the intermediate frequency signal of the multi-band radar signal. An adaptive filter is used to remove residual noise and irrelevant signal interference to obtain a preprocessed multi-band radar signal set. .

[0022] As a further optimization of the present invention, the two-layer generative adversarial network model in step S3 includes a first-layer generative adversarial network and a second-layer generative adversarial network.

[0023] The first layer of generative adversarial network is used to extract the preprocessed multi-band radar signal set. Global features in the dataset, and a global feature vector set generated based on these global features. and the global feature vector set The input is fed into the second layer of the generative adversarial network, where p represents the number of feature vectors. This represents the k-th global eigenvector;

[0024] The second layer of the generative adversarial network is used to process the global feature vector set. Refinement and separation processes are performed to generate anomaly feature vector sets. Where q represents the number of feature vectors of the anomaly. This represents the feature vector of the l-th anomaly.

[0025] As a further optimization of the present invention, the first-layer generative adversarial network includes a first-layer generator. First-layer discriminator and embedding the first-layer generator The encoder section and the first-layer discriminator The feature extraction layer contains a dual attention module, which includes a temporal attention module and a frequency domain attention module;

[0026] The calculation formula for the temporal attention module is as follows:

[0027] ;

[0028] in, Represents the query matrix. Represents the key matrix. Represents a value matrix, It is a key vector Dimensions It is a feature representation that has been weighted by attention weights;

[0029] The calculation formula for the frequency domain attention module is as follows:

[0030] ;

[0031] in, This represents the preprocessed multi-band radar signal set. , This indicates the preprocessed multi-band radar signal set. Perform a Fast Fourier Transform. This represents a multilayer perceptron. express function, It represents the Hadamah accumulation. This represents the spectral characteristics after frequency domain weighting enhancement.

[0032] As a further optimization of the present invention, the second-layer generative adversarial network includes a second-layer generator. Second-layer discriminator Integrated into the second-layer generator Second-layer discriminator Internal hierarchical attention refining module and introduction of a second-layer generator The structured feature extraction module in the middle, the hierarchical attention refinement module includes a cascaded feature selection attention network and an anomaly decoupling attention network;

[0033] The calculation formula for the feature selection attention network is as follows:

[0034] ;

[0035] ;

[0036] in, Represents the attention weight vector. and It consists of a learnable weight matrix and a bias vector. Represents a linear transformation. express function, It represents the Hadamah accumulation. Indicates weighted Modulated characteristics;

[0037] The calculation formula for the anomaly-decoupled attention network is as follows:

[0038] ;

[0039] ;

[0040] in, These represent the query matrix, key matrix, and value matrix, respectively. Let represent the projection matrices of the i-th head, respectively. This represents the learnable output projection matrix;

[0041] The calculation formula for the structured feature extraction module is as follows:

[0042] ;

[0043] in, Representing the implicit code c and the generated features Mutual information between them This represents the expected log-likelihood. It approximates the true posterior distribution neural networks, This represents the entropy of the implicit code c. This represents a hyperparameter that controls the weights of mutual information terms.

[0044] As a further optimization of the present invention, the output of the second-layer generative adversarial network includes a set of decoupled and classified anomaly feature vectors. and its corresponding classification labels derived from implicitly encoded c. .

[0045] As a further optimization of the present invention, it also includes:

[0046] Discriminator in the second layer of the generative adversarial network Next, an interpretability analysis module is integrated to generate structured interpretive data for each identified anomaly feature.

[0047] As a further optimization of the present invention, the discriminator in the second-layer generative adversarial network... Next, an interpretability analysis module is integrated, which includes the following steps:

[0048] Step 100: For the discriminator The feature map output by the last convolutional layer Calculate its gradient with respect to the final classification decision. ;

[0049] Step 200: Obtain the weights using gradient global average pooling. And generate class activation heatmap ;

[0050] ;

[0051] ;

[0052] Where Z is the feature map The total number of positions in the time dimension. It is the score of category c. For feature maps The gradient at position i, Indicates the overall importance weight. express function;

[0053] Step 300: Activate the class heatmap With preprocessed multi-band radar signal set Alignment in the time / frequency domain;

[0054] Step 400: Calculate the statistical characteristics of the original signals corresponding to the significant regions in the heatmap and generate a structured interpretation report, including anomaly location, anomaly type confidence, key evidence features, and feature contribution ranking.

[0055] As a further optimization of the present invention, the following steps are also included:

[0056] During the training of the two-layer generative adversarial network, the network structure of the generator and discriminator is optimized layer by layer, and the calculation method of mutual information is optimized.

[0057] The beneficial effects of this invention are as follows:

[0058] 1. This invention introduces a two-layer generative adversarial network (GAN) model, adding a second layer to the original GAN ​​model and combining it with the InfoGAN structured feature extraction module. The first-layer GAN focuses on extracting global features of multi-band radar signals, effectively capturing macroscopic feature information in radar signals. The second-layer GAN further refines and separates the local features of anomalies. For common structural and material anomalies in coal mining environments, the layered extraction and processing can make the features of anomalies clearer and more separable, thereby improving the accuracy and precision of identification.

[0059] 2. This invention optimizes the training process of the generator and discriminator through the collaborative work of a two-layer generative adversarial network. This includes improving the loss function design of the discriminator and adjusting the network layer structure of the generator. Through layer-by-layer optimization, it can effectively solve the problem that traditional generative adversarial network models are prone to falling into an unstable state during training, avoid model collapse or poor generation quality, enable the generator to generate samples that are closer to real radar signals, and the discriminator has a stronger ability to identify abnormal objects, further improving the model's generalization ability and robustness in complex environments.

[0060] 3. The two-layer generative adversarial network of the present invention can process and optimize multi-band radar signals. By extracting and separating the features of signals in different frequency bands, it improves the ability to identify complex anomalies in the coal mine environment. The first layer network extracts global features and performs preliminary classification, while the second layer network further refines the feature information of different types of anomalies, including the ability to distinguish between structural anomalies and material anomalies. This multi-level and multi-angle feature extraction and separation technology ensures the accuracy of the identification results and adaptability to diverse anomalies.

[0061] 4. This invention embeds a hierarchical attention mechanism in the generator and discriminator, enabling the model to dynamically focus on key time segments, frequency bands or spatial regions in the signal, enhancing the ability to capture subtle abnormal features. An interpretability module is added after the output layer of the two-layer GAN to visualize the basis for anomaly identification, provide feature importance analysis, and enhance the transparency and acceptability of the model. Attached Figure Description

[0062] Figure 1 This is a flowchart of the coal mine radar signal anomaly identification method based on improved generative adversarial networks according to the present invention.

[0063] Figure 2 This is a data processing flowchart of the coal mine radar signal anomaly identification method based on improved generative adversarial networks according to the present invention. Detailed Implementation

[0064] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed merely to enable those skilled in the art to better understand and implement the subject matter described herein. Furthermore, features described in some examples may be combined in other examples.

[0065] like Figure 1 As shown, the method for identifying anomalies in coal mine radar signals based on improved generative adversarial networks includes the following steps:

[0066] Step S1: Acquire radar signals covering the geology inside the mine in real time and construct a multi-band radar signal set;

[0067] In an optional embodiment of the present invention, step S1 includes the following specific steps:

[0068] Step S11: Deploy multi-band radar equipment at multiple locations within the mine, and use the multi-band radar equipment to collect multi-band radar signals. Set the collection time to T, and within time T, collect signals at a sampling frequency... Signals were collected from different locations inside the mine to form a multi-band radar signal set. Where n represents the number of radar signals collected, and m represents the number of collection points within the mine. This represents the i-th multi-band radar signal acquired at the j-th acquisition point;

[0069] Step S12: For multi-band radar signals from different acquisition points Record the frequency distribution of the signal ,in, Representing frequency, constructing a frequency distribution set for multi-band radar signals. ;

[0070] Step S13: Based on the collected multi-band radar signal set and frequency distribution set This forms the final multi-band radar signal set. , This is to comprehensively cover the geological and structural information inside the mine.

[0071] Step S2: Preprocess the multi-band radar signal set, including noise reduction, normalization and filtering operations;

[0072] In an optional embodiment of the present invention, step S2 includes the following specific steps:

[0073] Step S21: For the final multi-band radar signal set... Denoising processing is performed on the multi-band radar signal set using an adaptive denoising algorithm. Noise is identified and removed from the signal, and the denoised multi-band radar signal set is recorded as follows: ;

[0074] Step S22: Process the denoised multi-band radar signal set Normalization is performed using a linear normalization method to scale the signal amplitude range to the [0,1] interval. The normalized multi-band radar signal set is denoted as... ;

[0075] Step S23: Process the normalized multi-band radar signal Multi-stage filtering is performed, using bandpass filters to eliminate low-frequency and high-frequency interference components, resulting in the intermediate frequency signal of the multi-band radar signal. An adaptive filter is used to remove residual noise and irrelevant signal interference to obtain a preprocessed multi-band radar signal set. .

[0076] Step S3: Construct a two-layer generative adversarial network (GAN) model incorporating a structured feature extraction module and a dual attention module. This GAN model processes the preprocessed multi-band radar signal set and generates anomaly feature information data, including structural and material anomalies, such as... Figure 2 As shown;

[0077] In an optional embodiment of the present invention, the two-layer generative adversarial network model in step S3 includes a first-layer generative adversarial network and a second-layer generative adversarial network.

[0078] The first layer of the generative adversarial network is used to extract the preprocessed multi-band radar signal set. Global features in the dataset, and a global feature vector set generated based on these global features. and the global feature vector set The input is fed into the second layer of the generative adversarial network, where p represents the number of feature vectors. This represents the k-th global eigenvector;

[0079] The second-layer generative adversarial network is used to process the global feature vector set. Refinement and separation processes are performed to generate anomaly feature vector sets. Where q represents the number of feature vectors of the anomaly. This represents the feature vector of the l-th anomaly.

[0080] The first-layer generative adversarial network includes a first-layer generator. First-layer discriminator and embedding the first-layer generator The encoder section and the first-layer discriminator The dual attention module in the feature extraction layer, the first layer generator and the first-layer discriminator A one-dimensional convolutional neural network architecture is adopted to adapt to radar signal sequence data. The dual attention module includes a temporal attention module and a frequency domain attention module. The temporal attention module is used to focus on key segments of the signal in the time dimension, while the frequency domain attention module calculates attention weights in the frequency domain dimension after performing a fast Fourier transform on the signal to enhance the focus on key frequency components.

[0081] The calculation formula for the temporal attention module is as follows:

[0082] ;

[0083] in, Represents the query matrix. Represents the key matrix. The representation matrices are all obtained by performing different linear transformations on the input feature map. They map the input signal to different representation subspaces to facilitate the calculation of correlations. It is a key vector Dimensions This is used to scale the dot product result, preventing the gradient from becoming too small in the Softmax function. It is a feature representation that has been weighted by attention weights;

[0084] The calculation formula for the frequency domain attention module is as follows:

[0085] ;

[0086] in, This represents the preprocessed multi-band radar signal set. , This indicates the preprocessed multi-band radar signal set. Perform a Fast Fourier Transform to obtain its frequency domain representation. This represents a multilayer perceptron used to learn the importance score for each frequency component. express A function is used to compress these scores into the (0,1) interval as attention weights. It represents the Hadamah accumulation. This represents the spectral characteristics after frequency domain weighting enhancement.

[0087] It should be noted that the dual attention module can filter interference from both the time (depth) and frequency dimensions, focusing on key information. For example, when a minor delamination (separation) occurs in the roof strata, a weak additional echo will be mixed in the reflected signal. However, this echo is easily drowned out by the periodic vibration noise and random electromagnetic noise generated by the operation of the coal mining machine. The time-series attention module learns that the normal strata reflection wave and mechanical vibration noise have a relatively fixed pattern (such as periodicity) on the time axis. When a small echo that does not conform to the strata model prediction and does not follow the vibration period appears, the attention weight will increase significantly at this time point, indicating "an anomaly here". The frequency domain attention module analyzes in the frequency domain and finds that the frequency components of this weak echo (e.g., high-frequency scattering caused by the characteristics of the delamination interface) are different from common electromagnetic interference frequency bands (such as power frequency harmonics). The frequency domain attention will strengthen this characteristic frequency band and suppress the interference frequency band, achieving early and accurate anomaly location and providing valuable early warning time for support.

[0088] Alternatively, radar signals may generate strong reflected echoes when encountering metallic I-beams (debris) or water-rich fissures. Relying solely on echo intensity can easily lead to misjudgment. The frequency domain attention module can identify that metal reflects radar waves with wide bandwidth and low loss characteristics, while water strongly absorbs certain high-frequency components, resulting in characteristic dips in the reflection spectrum. The frequency domain attention module automatically amplifies these discriminative spectral differences, while the time-series attention module, combined with the signal's two-way travel time, can roughly determine the target depth. The module integrates depth information and spectral characteristics to allocate attention. By extracting spectral fingerprints, the module can provide a tendency judgment at the initial stage of alarm—"strong reflection originates from a target with metallic spectral characteristics" or "strong reflection is accompanied by high-frequency attenuation, consistent with water characteristics." This greatly improves the specificity and operability of the early warning information.

[0089] The second-layer generative adversarial network includes a second-layer generator. Second-layer discriminator Integrated into the second-layer generator Second-layer discriminator Internal hierarchical attention refining module and introduction of a second-layer generator The structured feature extraction module and hierarchical attention refinement module include a cascaded feature selection attention network and an anomaly decoupling attention network. The feature selection attention network is used to process the global features of the input. Importance scoring is performed, redundant information is filtered, and anomaly decoupling is achieved. Based on feature selection, the attention network uses a multi-head attention mechanism to learn different semantic anomaly features (such as structural anomalies and material anomalies) in parallel.

[0090] The calculation formula for the feature selection attention network is as follows:

[0091] ;

[0092] ;

[0093] in, Represents the attention weight vector. and It consists of a learnable weight matrix and a bias vector. Represents a linear transformation. express The function maps output values ​​to the range (0,1) as importance scores for each feature dimension. It represents the Hadamah accumulation. Indicates weighted In the modulated features, less important features are suppressed;

[0094] The calculation formula for the anomaly-decoupled attention network is as follows:

[0095] ;

[0096] ;

[0097] in, These represent the query matrix, key matrix, and value matrix, respectively, derived from the input features. Obtained through different linear projections These represent the projection matrices of the i-th head, used to project the input into different subspaces. Each head independently performs scaled dot product attention (e.g., ...). (functions) generate characteristic representations of the subspace. The operation concatenates the outputs of all heads and finally projects them through a learnable output projection matrix. The features are then fused to obtain the final refined feature representation;

[0098] The calculation formula for the structured feature extraction module is as follows:

[0099] ;

[0100] in, Representing the implicit code c and the generated features Mutual information between them measures the information contained in the implicit code c about... The amount of information is such that direct calculation of mutual information is difficult; therefore, a variational lower bound is used for approximation. This represents the expected log-likelihood. It approximates the true posterior distribution A neural network (auxiliary classifier) ​​that bases its data on features. Predict the hidden coding point c. The entropy of the implicit code c is given by the prior distribution of c. When the distribution is fixed (such as a uniform categorical distribution or a standard normal distribution), this term is a constant. The hyperparameter represents the weights of the mutual information terms. Maximizing this loss associates the hidden code c with meaningful, interpretable feature dimensions, such as anomaly types.

[0101] In an optional embodiment of the invention, the output of the second-layer generative adversarial network includes a decoupled and classified set of anomaly feature vectors. and its corresponding classification labels derived from implicitly encoded c. .

[0102] In an optional embodiment of the present invention, it further includes:

[0103] Discriminator in the second layer of the generative adversarial network Subsequently, an interpretability analysis module is integrated to generate structured explanatory data for each identified anomaly feature, providing visual and numerical interpretations to improve the transparency and credibility of model decisions.

[0104] Discriminator in the second layer of the generative adversarial network Next, an interpretability analysis module is integrated, which includes the following steps:

[0105] Step 100: For the discriminator The feature map output by the last convolutional layer Calculate its gradient with respect to the final classification decision. ;

[0106] Step 200: Obtain the weights using gradient global average pooling. And generate class activation heatmap ;

[0107] ;

[0108] ;

[0109] Where Z is the feature map The total number of positions in the time dimension. It is the score of category c. For feature maps The gradient at position i reflects the importance of that position's feature to the class decision. By averaging the gradients at all positions i, the overall importance weight of channel k is obtained. , Indicates the overall importance weight. express The function converts the feature maps of each channel. Use the corresponding weights A linear combination is performed, and the ReLU function is applied to the result, retaining only the feature regions that positively contribute to class c. The final result is a significance heatmap aligned with the time dimension of the input signal. ;

[0110] Step 300: Activate the class heatmap With preprocessed multi-band radar signal set Alignment in the time / frequency domain;

[0111] Step 400: Calculate the significant regions in the heatmap (e.g., regions above a threshold). The statistical characteristics (such as energy, peak value, and spectral centroid) of the original signal corresponding to the part of the signal are analyzed, and a structured interpretation report is generated, including anomaly location, anomaly type confidence, key evidence features, and feature contribution ranking.

[0112] Anomaly location: the approximate time / depth location where the anomaly occurred;

[0113] Anomaly type confidence: The probability of the model classifying structural anomalies / material anomalies;

[0114] Key evidence features: Description of the most critical signal features that lead to this judgment (e.g., "a high-frequency energy surge occurs during the t1-t2 time period");

[0115] Feature contribution ranking: The top N feature dimensions and their contribution values, sorted by contribution.

[0116] Step S4: Generate anomaly report data based on the anomaly feature information data. The anomaly report data includes data on the location, type, and nature of the anomaly.

[0117] The automatic generation of anomaly report data includes the following sections:

[0118] Abstract: This paper summarizes the total number of anomalies, main types, and highest risk items discovered during this monitoring.

[0119] Detailed list: List each anomaly in tabular form its unique ID, location (collection point j, time index), type (structure / substance), risk level (high / medium / low), and model confidence level.

[0120] Visualization analysis: Includes the original waveforms and spectrograms of key anomalous signals, as well as a comparison chart overlaid with Grad-CAM heatmaps.

[0121] Evidence and Explanation: Citing the analysis results of S4a2, describe the basis for the model's judgment in natural language.

[0122] Recommended measures: Based on the anomaly type and historical data, provide preliminary handling recommendations (such as "It is recommended to conduct intensive scanning of the area" or "It meets the characteristics of historical gas accumulation, and ventilation inspection is recommended").

[0123] It should be noted that by utilizing the synergistic effect of two-layer generative adversarial networks, the first-layer generative adversarial network identifies and initially classifies anomalies, while the second-layer generative adversarial network refines the classification. Through optimized discriminators and generators, different types of anomalies are accurately identified, effectively capturing macroscopic feature information in radar signals. The second-layer generative adversarial network further refines and separates the local features of anomalies. For common structural and material anomalies in coal mining environments, the features of anomalies can be made clearer and more separable through layered extraction and processing, thereby improving the accuracy and precision of identification.

[0124] A two-layer generative adversarial network can process and optimize multi-band radar signals. By extracting and separating the features of signals from different frequency bands, it improves the ability to identify complex anomalies in the coal mine environment. The first layer of the network extracts global features and performs preliminary classification, while the second layer further refines the feature information of different types of anomalies, including the ability to distinguish between structural anomalies and material anomalies. This multi-level, multi-angle feature extraction and separation technology ensures the accuracy of the identification results and adaptability to diverse anomalies.

[0125] For example, in areas with disordered rock strata, small faults (structural anomalies) and sudden changes in lithology (such as a change from sandstone to mudstone, material anomalies) may appear similar on radar profiles, both showing disordered phase axes or amplitude variations. However, the former carries a much higher risk than the latter.

[0126] Feature selection attention: First, the module examines global features, ignoring background lithological noise that is prevalent throughout the region, and focuses on local abrupt changes.

[0127] Anomaly decoupling attention (multi-head mechanism):

[0128] Head A (Structural Anomaly Head): Specifically designed to find patterns of phase discontinuities and in-phase axis breaks in signals. For small faults, it will find a clear, traceable fault line.

[0129] Head B (Material Anomaly Head): Specifically analyzes the amplitude decay rate and spectral variation patterns of signals. For lithological variations, it finds that amplitude and spectral changes are more diffuse rather than along a clear interface.

[0130] Maximizing mutual information: forcing the network to combine the features extracted from the A-head with the implicit coding of "structural anomalies". Strong correlation, linking the features extracted from the B-head with the implicit coding of "material anomalies". Strong correlation.

[0131] Compared to traditional methods and ordinary GANs: traditional methods may only report "abnormal regions" and cannot classify them. Ordinary GANs may extract mixed features, leading to ambiguous classification.

[0132] The advantage of this module is that it can output clear and interpretable judgments: "A structural anomaly characterized by phase discontinuity exists at coordinates (X,Y) (92% confidence level)"; "A material anomaly characterized by spectral dispersion variation exists at coordinates (M,N) (85% confidence level)". This provides geological engineers with a basis for decision-making, allowing them to prioritize drilling verification of faults.

[0133] The following issues also exist: both old air zones (filled with air) and gas (CH4) accumulation zones exhibit low-resistivity, high-reflection "low-resistivity" anomalies, but their causes and hazards are different.

[0134] Feature selection attention: Focus on regions in radar signals that exhibit significant waveform stretching (low speed) and strong amplitude reflection.

[0135] Anomaly decoupling attention:

[0136] Head A (Structural Void Head): Learn the radar response model of a typical void, i.e., the interface reflection is extremely sharp, and the internal multiple reflection signals are chaotic but have certain regularity.

[0137] Head B (Gas Anomalous Head): The model of the effect of gas on electromagnetic waves is studied. Its reflection interface may not be as sharp as that of a cavity, and due to the scattering of gas at specific frequencies, it may leave weak resonance characteristics in the spectrum.

[0138] Mutual information optimization: Through training with a large number of samples, the network learns to decouple subtle differences in waveform shape and spectral texture to different feature channels.

[0139] Compared with traditional methods: Traditional geophysical exploration methods have difficulty distinguishing between the two, and they are often collectively referred to as "low-resistivity anomalies", which require drilling for verification, resulting in high costs and risks.

[0140] This module's advantages include: through hierarchical refinement, the network can provide trend reports such as: "The waveform characteristics of Region 1 are more consistent with the cavity model; it is recommended to verify the goaf map"; "The reflection spectrum of Region 2 shows atypical resonance characteristics, and the possibility of gas enrichment cannot be ruled out; it is recommended to combine it with gas monitoring." This achieves risk-level classification and early warning.

[0141] In summary, a two-layer generative adversarial network, which introduces a dual attention module and a hierarchical attention refinement module, can elevate the analysis of coal mine radar signals from the traditional "signal alarm" level to a new level of "intelligent diagnosis".

[0142] In an optional embodiment of the present invention, the following steps are further included:

[0143] During the training of the two-layer generative adversarial network, the network structure of the generator and discriminator is optimized layer by layer, and the calculation method of mutual information is optimized.

[0144] Specifically, the following steps are included:

[0145] Network initialization and parameter settings:

[0146] Initialize generator , and discriminator , The weights are initialized using the Xavier method to ensure gradient stability during the initial training phase.

[0147] Optimizer settings: The Adam optimizer is used, where the learning rate of the first layer network is... The learning rate of the second layer network Momentum parameters , .

[0148] Set the batch size B=32, and the dimension of the implicit code c is 10, where the first 5 dimensions are used to represent discrete anomaly types (such as structural anomalies, material anomalies, etc.), and the last 5 dimensions are used to represent continuous anomaly attributes (such as scale, intensity).

[0149] Layered pre-training:

[0150] First layer network training: Fix the parameters of the second layer network, and use the preprocessed dataset. Training the first layer of GAN ( , Alternate updates and :

[0151] renew :maximize ;

[0152] renew : minimize ;

[0153] Training until The loss is stable and It can generate models with a reasonable global structure;

[0154] Second layer network training: Fix the parameters of the first layer network, and train the global features extracted by the first layer. As input, train the second layer GAN ( , ) and mutual information module;

[0155] renew :maximize ;

[0156] renew With auxiliary classifier Q: Minimize the total generation loss ;

[0157] This stage focuses on learning the mapping from global features to refined anomalous features and the association between implicit coding and features.

[0158] End-to-end joint fine-tuning:

[0159] Unfreeze all network layer parameters.

[0160] An alternating optimization strategy is adopted, with updates performed sequentially in each iteration:

[0161] Update all discriminators: and Maximize their respective discrimination losses and

[0162] Update all generators and related modules: , And Q, minimize the global total loss ;

[0163] in, It is the attention regularization term, applied to all attention weight matrices in the network. The Frobenius norm (i.e., the square root of the sum of squares of all elements) is used to penalize, preventing the attention mechanism from over-focusing on a very small number of neurons and encouraging diversity in feature utilization. It is the regularization coefficient.

[0164]

[0165] in, Encouragement discriminator Correctly identify the real preprocessed signal , Encourage generator Generate signals that are so realistic they can fool the discriminator. It is a random noise vector that follows a certain distribution (such as the standard normal distribution). It is a reconstruction loss that constrains the encoding-decoding process, ensuring that important information is not lost during feature extraction. Denotes the square of the L2 norm. It is a hyperparameter used to balance adversarial loss and reconstruction loss;

[0166]

[0167] in, Encouragement discriminator Correctly identify the real preprocessed signal , Encourage generator Generate signals that are so realistic they can fool the discriminator. It is a random noise vector that follows a certain distribution (such as the standard normal distribution). It is a reconstruction loss that constrains the encoding-decoding process, ensuring that important information is not lost during feature extraction. Denotes the square of the L2 norm. It is a hyperparameter used to balance adversarial loss and reconstruction loss;

[0168] To prevent pattern collapse, one-sided label smoothing (smoothing the real data labels from 1 to 0.9) is occasionally added during the training of the discriminator.

[0169] Training monitoring and convergence judgment:

[0170] Monitoring metrics: In addition to monitoring combat losses, also monitor reconstruction losses. Mutual information loss The numerical change.

[0171] Introduce a validation set and calculate the precision, recall, and F1 score for anomaly classification on the validation set.

[0172] Convergence condition: When the F1 score on the validation set no longer improves for P=10 consecutive training epochs, and the total loss... Training is stopped when the fluctuation amplitude is less than the threshold ϵ=1e−4.

[0173] Use an early stopping strategy to save the model parameters when the validation set performance is optimal.

[0174] It should be noted that the training process of the generator and discriminator is optimized through the collaborative work of the two-layer generative adversarial network. This includes improving the loss function design of the discriminator and adjusting the network layer structure of the generator. Through layer-by-layer optimization, the problem of traditional generative adversarial network models easily falling into an unstable state during training can be solved, avoiding model collapse or poor quality of generated results. This enables the generator to generate samples that are closer to real radar signals, while the discriminator has a stronger ability to identify abnormal objects, further improving the model's generalization ability and robustness in complex environments.

[0175] The above description of this embodiment is not limited to the specific implementation described above. The specific implementation described above is merely illustrative and not restrictive. Those skilled in the art can make many other forms based on the guidance of this embodiment, all of which are within the protection scope of this embodiment.

Claims

1. A method for identifying anomalies in coal mine radar signals based on an improved generative adversarial network, characterized in that, Includes the following steps: Step S1: Acquire radar signals covering the geology inside the mine in real time and construct a multi-band radar signal set; Step S2: Preprocess the multi-band radar signal set; Step S3: Construct a two-layer generative adversarial network model that incorporates a structured feature extraction module and a dual attention module. The two-layer generative adversarial network model is used to process the preprocessed multi-band radar signal set and generate anomaly feature information data. Step S4: Generate anomaly report data based on the anomaly feature information data. The anomaly report data includes data on the location, type, and nature of the anomaly.

2. The method for identifying anomalies in coal mine radar signals based on an improved generative adversarial network according to claim 1, characterized in that, Step S1 includes the following specific steps: Step S11: Deploy multi-band radar equipment at multiple locations within the mine, and use the multi-band radar equipment to collect multi-band radar signals. Set the collection time to T, and within time T, collect signals at a sampling frequency... Signals were collected from different locations inside the mine to form a multi-band radar signal set. Where n represents the number of radar signals collected, and m represents the number of collection points within the mine. This represents the i-th multi-band radar signal acquired at the j-th acquisition point; Step S12: For multi-band radar signals from different acquisition points Record the frequency distribution of the signal ,in, Representing frequency, constructing a frequency distribution set for multi-band radar signals. ; Step S13: Based on the collected multi-band radar signal set and frequency distribution set This forms the final multi-band radar signal set. , .

3. The method for identifying anomalies in coal mine radar signals based on an improved generative adversarial network according to claim 2, characterized in that, Step S2 includes the following specific steps: Step S21: For the final multi-band radar signal set... Denoising processing is performed on the multi-band radar signal set using an adaptive denoising algorithm. Noise is identified and removed from the signal, and the denoised multi-band radar signal set is recorded as follows: ; Step S22: Process the denoised multi-band radar signal set Normalization is performed using a linear normalization method to scale the signal amplitude range to the [0,1] interval. The normalized multi-band radar signal set is denoted as... ; Step S23: Process the normalized multi-band radar signal Multi-stage filtering is performed, using bandpass filters to eliminate low-frequency and high-frequency interference components, resulting in the intermediate frequency signal of the multi-band radar signal. An adaptive filter is used to remove residual noise and irrelevant signal interference to obtain a preprocessed multi-band radar signal set. .

4. The method for identifying anomalies in coal mine radar signals based on an improved generative adversarial network according to claim 3, characterized in that, The two-layer generative adversarial network model in step S3 includes a first-layer generative adversarial network and a second-layer generative adversarial network. The first layer of generative adversarial network is used to extract the preprocessed multi-band radar signal set. Global features in the dataset, and a global feature vector set generated based on these global features. and the global feature vector set The input is fed into the second layer of the generative adversarial network, where p represents the number of feature vectors. This represents the k-th global eigenvector; The second layer of the generative adversarial network is used to process the global feature vector set. Refinement and separation processes are performed to generate anomaly feature vector sets. Where q represents the number of feature vectors of the anomaly. This represents the feature vector of the l-th anomaly.

5. The method for identifying anomalies in coal mine radar signals based on an improved generative adversarial network according to claim 4, characterized in that, The first layer generative adversarial network includes a first layer generator. First-layer discriminator and embedding the first-layer generator The encoder section and the first-layer discriminator The feature extraction layer contains a dual attention module, which includes a temporal attention module and a frequency domain attention module; The calculation formula for the temporal attention module is as follows: ； in, Represents the query matrix. Represents the key matrix. Represents a value matrix, It is a key vector Dimensions It is a feature representation that has been weighted by attention weights; The calculation formula for the frequency domain attention module is as follows: ； in, This represents the preprocessed multi-band radar signal set. , This indicates the preprocessed multi-band radar signal set. Perform a Fast Fourier Transform. This represents a multilayer perceptron. express function, It represents the Hadamah accumulation. This represents the spectral characteristics after frequency domain weighting enhancement.

6. The method for identifying anomalies in coal mine radar signals based on an improved generative adversarial network according to claim 5, characterized in that, The second-layer generative adversarial network includes a second-layer generator. Second-layer discriminator Integrated into the second-layer generator Second-layer discriminator Internal hierarchical attention refining module and introduction of a second-layer generator The structured feature extraction module in the middle, the hierarchical attention refinement module includes a cascaded feature selection attention network and an anomaly decoupling attention network; The calculation formula for the feature selection attention network is as follows: ； ； in, Represents the attention weight vector. and It consists of a learnable weight matrix and a bias vector. Represents a linear transformation. express function, It represents the Hadamah accumulation. Indicates weighted Modulated characteristics; The calculation formula for the anomaly-decoupled attention network is as follows: ； ； in, These represent the query matrix, key matrix, and value matrix, respectively. Let represent the projection matrices of the i-th head, respectively. This represents the learnable output projection matrix; The calculation formula for the structured feature extraction module is as follows: ； in, Representing the implicit code c and the generated features Mutual information between them This represents the expected log-likelihood. It approximates the true posterior distribution neural networks, This represents the entropy of the implicit code c. This represents a hyperparameter that controls the weights of mutual information terms.

7. The method for identifying anomalies in coal mine radar signals based on an improved generative adversarial network according to claim 6, characterized in that, The output of the second-layer generative adversarial network includes a set of decoupled and classified anomaly feature vectors. and its corresponding classification labels derived from implicitly encoded c. .

8. The method for identifying anomalies in coal mine radar signals based on an improved generative adversarial network according to claim 7, characterized in that, Also includes: Discriminator in the second layer of the generative adversarial network Next, an interpretability analysis module is integrated to generate structured interpretive data for each identified anomaly feature.

9. The method for identifying anomalies in coal mine radar signals based on an improved generative adversarial network according to claim 8, characterized in that, The discriminator in the second-layer generative adversarial network Next, an interpretability analysis module is integrated, which includes the following steps: Step 100: For the discriminator The feature map output by the last convolutional layer Calculate its gradient with respect to the final classification decision. ; Step 200: Obtain the weights using gradient global average pooling. And generate class activation heatmap ; ； ； Where Z is the feature map The total number of positions in the time dimension. It is the score of category c. For feature maps The gradient at position i, Indicates the overall importance weight. express function; Step 300: Activate the class heatmap With preprocessed multi-band radar signal set Alignment in the time / frequency domain; Step 400: Calculate the statistical characteristics of the original signals corresponding to the significant regions in the heatmap and generate a structured interpretation report, including anomaly location, anomaly type confidence, key evidence features, and feature contribution ranking.

10. The method for identifying anomalies in coal mine radar signals based on an improved generative adversarial network according to claim 9, characterized in that, It also includes the following steps: During the training of the two-layer generative adversarial network, the network structure of the generator and discriminator is optimized layer by layer, and the calculation method of mutual information is optimized.

Images