Method for quickly identifying mine earthquake event based on short-time small-sample micro-earthquake monitoring data

Abstract

The invention provides a method for quickly identifying a mine earthquake event based on short-time small-sample micro-earthquake monitoring data. The method comprises the following steps: collecting the short-time small-sample micro-earthquake monitoring data through a micro-earthquake monitoring system arranged in an underground coal mine; preprocessing the collected micro-seismic monitoring data to form an original data set for mine earthquake event identification; training the original data set by using a given improved generative adversarial neural network (CTTS-GAN), and generating high-quality synthetic micro-seismic data to expand the original data set; based on the expanded micro-seismic monitoring data set, the mine earthquake event can be quickly identified through a CNN-RNN earthquake detector (CRED) model. According to the method, the problem of small samples faced by micro-seismic data of an underground coal mine is effectively solved, the judgment accuracy of the CRED model is greatly improved, and the method is suitable for rapid judgment of mine earthquake events.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of mine earthquake data processing, and particularly relates to a short-time small sample microseismic monitoring data rapid identification of mine earthquake event method based on combination of an improved generative adversarial network (CTTS-GAN) and a CNN-RNN earthquake detector (CRED). BACKGROUND

[0002] In the process of coal mine underground mine earthquake monitoring, the quality of microseismic monitoring data is easily affected by instrument noise and environmental interference and other factors, and is limited by the sparse distribution of underground monitoring points, resulting in insufficient number of effective samples. This situation makes the traditional generative adversarial network model face problems such as poor stability and low feature extraction efficiency in the training process, thereby causing phenomena such as model overfitting and low generalization performance, making it difficult to achieve efficient identification of mine earthquake events. In view of the above problems, an improved generative adversarial network is needed, which takes into account the data generation quality and efficient feature learning ability, to expand the short-time small sample data set of microseismic monitoring, and then realize the mine earthquake identification algorithm of short-time small sample microseismic monitoring data through the CNN-RNN earthquake detector (CRED). SUMMARY

[0003] The present application aims to provide a short-time small sample microseismic monitoring data rapid identification of mine earthquake event method, to solve the problems of difficult microseismic data acquisition and low generated data quality in existing algorithms, to improve data processing efficiency and algorithm model performance, and to further enable rapid identification of mine earthquake events through short-time small samples.

[0004] To achieve the above purpose, the following technical solution is adopted: a short-time small sample microseismic monitoring data rapid identification of mine earthquake event method based on combination of an improved generative adversarial network (CTTS-GAN) and a CNN-RNN earthquake detector (CRED), specifically comprising the following steps:

[0005] Step 1: Data acquisition and equipment deployment

[0006] Determine the monitoring area, deploy the microseismic monitoring system, and ensure that the coverage meets the microseismic signal capture requirements. Real-time acquisition of microseismic monitoring data, including microseismic waveform, timestamp, source location and magnitude, is synchronized to the database and implemented with redundant backup and regular maintenance. After completing the installation of the equipment, perform signal stability test and calibration to eliminate the noise interference of the equipment itself;

[0007] Step 2: Data preprocessing

[0008] Denoise, missing value processing, normalization and feature extraction are performed on the collected microseismic monitoring data to ensure data quality. The preprocessed data is then divided into training set, validation set and test set for model training, hyperparameter tuning and performance evaluation, respectively;

[0009] Step 3: Improved architecture design of the generative adversarial network (CTTS-GAN)

[0010] The CTTS-GAN model generator proposed in this paper accepts random noise as input, first performs dimension conversion and adds position encoding information through linear layers and position embedding layers, and then reduces the internal covariate shift through layer normalization. Then, the feature representation of microseismic data is gradually learned through global-local fusion module (GLFM) and recurrent calculation of feedforward neural network. Finally, residual connection is used to realize effective fusion of features of each module, and convolutional layer is used to adjust the output to the same dimension as the real sample.

[0011] In particular, the global-local fusion module (GLFM) designed in the generator fuses the convolutional self-attention layer and the multi-head self-attention layer. Through the convolutional self-attention layer, the attention calculation of the local perception domain is strengthened, which strengthens the model's learning of the local details of the input sequence. At the same time, the multi-head self-attention layer is used to calculate the correlation between elements in the sequence, providing modeling of the global semantic structure. Finally, the results of the multi-head attention layer and the convolutional self-attention are fused through the use of residual connection, so that the generator can more comprehensively and accurately capture the features of the microseismic data.

[0012] The discriminator of the CTTS-GAN model is a binary classification model used to distinguish between real sequences and synthetic sequences in the training set. The model takes the global optimization target and the single-channel optimization target as the loss function of the model, and the corresponding discriminator is composed of a global discriminator and a single-channel discriminator. The global discriminator evaluates the difference in overall distribution characteristics between the synthetic sequence and the real sequence to determine the authenticity of the sequence; the single-channel discriminator discriminates the single-variable microseismic signal in each independent channel to evaluate the synthesis quality of the generator in each channel. The losses of the two discriminators are transmitted together through backpropagation to guide the training and optimization of the generator, gradually improving the authenticity of the synthetic sequence. The parameters of the two discriminators are independent of each other and do not interfere with each other;

[0013] Step 4: Improved training and optimization of the generative adversarial network (CTTS-GAN)

[0014] The generator and the discriminator are alternately trained to optimize the generative adversarial network, generate high-quality synthetic data, and expand the original data set. The specific steps are as follows:

[0015] Step s41: Sample a noise vector from a normal distribution and input it into the generator G to obtain a generated sample. Then, the real microseismic waveform data is preprocessed and input into the discriminator D together with the generated sample to determine the authenticity, i.e., whether the data is from the real data or the data generated by the generator;

[0016] Step s42: training the discriminator with fixed generator parameters, calculating the loss function according to the discrimination result, and updating the discriminator parameters; then training the generator with fixed discriminator parameters, calculating the loss function of the generator (i.e. the probability of the generated data being judged as real data), and updating the generator parameters. Repeat the alternative training until the mine seismic waveform generated by the generator is difficult to distinguish from the real waveform;

[0017] Step s43: adjusting the hyperparameters of the network, such as learning rate, batch size, network layer number and neuron number, etc., to further optimize the performance of the model; during the parameter optimization process, the validation set is used to evaluate the effect of different hyperparameter combinations, and the best configuration is selected;

[0018] Step s44: using the test set to evaluate the quality of the generated samples;

[0019] Step s45: using the trained generator to generate new microseismic samples, and up-sampling them to form synthetic samples of the CTTS-GAN model. Then these synthetic samples are added to the real microseismic data to expand the original data set, which is used for the training of the subsequent seismic detector (CRED) model.

[0020] Step 5: using the CNN-RNN seismic detector (CRED) model to identify mine seismic events

[0021] The present application adopts a seismic detector (CRED) model, which focuses on accurately identifying and classifying features in microseismic signals, such as P-wave and S-wave arrival time information, and then identifying mine seismic events.

[0022] The model mainly consists of a convolutional neural network (CNN) and a long short-term memory network (LSTM, which belongs to a kind of RNN). First, the frequency spectrum of the three-component seismic waveform obtained by Fourier transform and its corresponding label value are used as the input of the CRED model, the local features of the microseismic waveform are extracted by the convolutional neural network, and the residual connection is used to alleviate the gradient vanishing problem of the network. After the two-dimensional convolution layer, the feature vector is unfolded into a sequence and transmitted to the residual block of two bidirectional long short-term memory networks (Bi-LSTM) to learn the time-frequency law of the main phase of the microseismic signal. The model uses Adam optimizer, mean square error as loss function, and in the training process, the learning rate adopts periodic decay strategy, through continuous iteration training, the network parameters are optimized, and the model learns the microseismic waveform features.

[0023] The real microseismic waveform data is divided into a training set, a validation set, and a test set. The synthetic samples generated by the improved generative adversarial network (CTTS-GAN) in step four are mixed with the real microseismic waveform data in the training set according to different expansion ratios as an expanded training set, which is input into the CRED model for feature extraction and identification of mine seismic signals. Based on the microseismic signal features extracted by the CRED model, it is determined whether there is a mine seismic event based on a pre-set threshold.

[0024] Finally, the test set is used to evaluate the classification accuracy, recall rate, precision rate, and F1 value performance indicators of the CRED model when the synthetic data of the CTTS-GAN model proposed in this paper is mixed with real microseismic data at different expansion ratios, to ensure the discrimination effect of the model.

[0025] Step 6: Model iteration

[0026] The discrimination results are applied to actual microseismic monitoring to improve the reliability of the microseismic monitoring system. New microseismic data is continuously collected, and the improved generative adversarial network (CTTS-GAN) and the seismic detector (CRED) model are continuously retrained to ensure that the model maintains high precision and high performance.

[0027] Preferably, in step two, the original data is preprocessed by first denoising to filter out environmental noise and equipment noise, then checking the signal-to-noise ratio, continuity, and sampling rate of the data to remove poor-quality or incomplete data segments, and finally normalizing the data to the same dimension range for subsequent analysis and modeling. Finally, key features such as time features (P-wave and S-wave arrival time), frequency features, and amplitude features are extracted from the waveform data.

[0028] Preferably, in step three, the CTTS-GAN model uses global optimization objectives and single-channel optimization objectives as the loss function of the model, and adjusts the weight coefficients in the global optimization objectives and single-channel optimization objectives loss function by setting the value of the hyperparameter λ to balance the two objectives of "improving single-channel quality" and "maintaining inter-channel correlation" in the multi-channel synthesis task.

[0029] Preferably, in step five, the different data augmentation methods herein specifically refer to noise disturbance augmentation, TTS-GAN augmentation, CTTS-GAN-global augmentation, CTTS-GAN augmentation, and setting the augmentation ratio to 20%, 40%, respectively. In order to explore the influence of the number of data set augmentation on the model, the CTTS-GAN model proposed in this paper is used to augment the original real data set at different ratios to evaluate the influence of the generated data on the model effect. The performance evaluation indicators of the seismic detector (CRED) model in the test set can be analyzed to select the appropriate augmentation ratio to make the overall performance evaluation indicators of the model optimal.

[0030] Compared with the prior art, the beneficial effects of the present application are:

[0031] 1. The improved generative adversarial network (CTTS-GAN) is used to generate microseismic data. Unlike traditional generative adversarial networks, the model introduces a global-local fusion module (GLFM) in the generator, which realizes the collaborative modeling of local detail features and global semantic structure of microseismic waveform data. At the same time, by designing a multi-level loss function containing global optimization target and single-channel optimization target, the key features of single-variable sequence can be effectively preserved while maintaining the overall structure of multi-variable sequence. These improvements make the statistical characteristics of the synthesized mine seismic samples closer to the real monitoring samples.

[0032] 2. The expanded samples are trained and identified by the seismic detector (CRED) model, which significantly improves the discrimination accuracy of mine seismic events. Experiments show that the overall performance of the CRED model can be optimized by analyzing the performance evaluation indicators of the CRED model in the test set and selecting the appropriate expansion ratio. In addition, the model has precise discrimination ability for P-wave and S-wave characteristics in short-time small sample scenarios.

[0033] 3. The generated synthetic data is used to expand the original data set, effectively solving the small sample data problem, enhancing the generalization ability of the seismic detector (CRED) model, and significantly improving the adaptability and accuracy of the model in different microseismic monitoring scenarios.

[0034] 4. Supports real-time identification of short-time microseismic signals, ensuring that the model maintains high-precision microseismic event identification during continuous updating of short-time small sample data. BRIEF DESCRIPTION OF DRAWINGS

[0035] Figure 1 The overall flowchart of the short-time small sample microseismic monitoring data rapid identification mine seismic event method provided by the present application is shown in the figure;

[0036] Figure 2A generator structure diagram of an improved generative adversarial network (CTTS-GAN) model of the application;

[0037] Figure 3 a, 3b and 3c are respectively a structure flow diagram of a global-local fusion module (GLFM) in a generator of an improved generative adversarial network model of the application, a diagram of a convolution attention calculation method, and a diagram of a multi-head attention calculation method;

[0038] Figure 4 A discriminator structure diagram of an improved generative adversarial network (CTTS-GAN) model of the application;

[0039] Figure 5 A time-frequency comparison diagram of a part of real microseismic samples and synthetic samples of each model of an embodiment of the application;

[0040] Figure 6 A CNN-RNN seismic detector (CRED) model architecture diagram of an embodiment of the application. DETAILED DESCRIPTION

[0041] The application will be further described in detail below with reference to the drawings.

[0042] As shown in Figure 1 , the application provides a short-time small-sample microseismic monitoring data rapid identification of mine earthquake event method. The method comprises the following steps:

[0043] Step 1: data acquisition and equipment deployment

[0044] Determine the monitoring area and target, lay out a high-sensitivity seismograph, and ensure that the coverage meets the weak signal capture requirements. Real-time acquisition of microseismic monitoring data, including microseismic waveform, timestamp, source location and magnitude, is synchronized to the database, and redundant backup and regular maintenance are implemented. After completing the equipment installation, perform signal stability test and calibration to eliminate the device's own noise interference;

[0045] Step 2: data preprocessing

[0046] The collected microseismic monitoring data is denoised, missing value processed, normalized and feature extracted. According to the preset threshold, in this embodiment, the threshold is selected as 3.75Hz, and the data segments with smaller amplitude or smaller waveform change are filtered out. At the same time, the signal-to-noise ratio, continuity and sampling rate of the data are checked, and the data segments with poor quality or incompleteness are removed to ensure the data quality. Then the preprocessed data is divided into training set, validation set and test set, which are respectively used for model training, hyperparameter tuning and performance evaluation;

[0047] Step 3: architecture design of improved generative adversarial network (CTTS-GAN)

[0048] As Figure 2 shown, the generator structure diagram in the improved generative adversarial network (CTTS-GAN) model proposed in this paper. The generator first converts the input noise z through linear layer and position embedding layer and adds position encoding information, and then reduces the network internal covariant offset through layer normalization (LN). Then, the global-local fusion module (GLFM) and the feedforward neural network are used to learn the feature representation of microseismic data step by step. Finally, the residual connection is used to realize the effective fusion of the features of each module, and the convolution layer is used to output the synthesized data with the same dimension as the real samples.

[0049] In particular, a global-local fusion module (GLFM) based on attention mechanism is designed in the generator, as shown in Figure 3 The module realizes the collaborative optimization of local features and global features by fusing convolution self-attention layer and multi-head self-attention layer. Through the dynamic convolution process of the convolution self-attention layer, the attention information of the local data is extracted combined with the context features, so as to strengthen the learning of the local detail features of the input sequence. At the same time, the multi-head self-attention layer is used to calculate the global correlation between elements in the sequence in parallel, where each attention head independently calculates a set of attention weights, and finally the multiple sets of calculation results are spliced and fused to capture the feature dependency relationship in different dimensions. Finally, the output results of the multi-head attention layer and the convolution self-attention layer are fused through the use of residual connection, so that the generator can more comprehensively and accurately capture the features of microseismic data.

[0050] The discriminator structure of the CTTS-GAN model is shown in Figure 4 The discriminator is a binary classification model used to distinguish between real sequences and synthesized sequences in the training set. The model takes the global optimization target and the single-channel optimization target as the loss function of the model, and the corresponding discriminator is composed of a global discriminator and a single-channel discriminator, which are used to evaluate the difference between the overall distribution characteristics of the synthesized sequence and the real sequence and the synthesis quality of the single-variable microseismic signal in each independent channel. The losses of the two discriminators are transmitted together through backpropagation to guide the training and optimization of the generator, gradually improving the authenticity of the synthesized sequence. The parameters of the two discriminators are independent of each other and do not interfere with each other;

[0051] Step 4: Training and optimization of improved generative adversarial network (CTTS-GAN)

[0052] The generator and the discriminator are alternately trained to optimize the generative adversarial network, generate high-quality synthesized data, and expand the original data set. The specific steps are as follows:

[0053] Step s41: Sample noise vectors from a normal distribution and input them into the generator G to obtain generated samples. Preprocess the real microseismic waveform data and input it together with the generated samples into the discriminator D to determine whether the data is real or generated by the generator;

[0054] Step s42: Train the discriminator with fixed generator parameters, calculate the loss function according to the discrimination result, and update the discriminator parameters. Then train the generator with fixed discriminator parameters, calculate the loss function of the generator (i.e. the probability of the generated data being judged as real data), and update the generator parameters. Repeat the alternating training until the generated microseismic waveform is difficult to distinguish from the real waveform;

[0055] Step s43: Adjust the hyperparameters of the network, such as learning rate, batch size, network layer number and neuron number, to further optimize the performance of the model. During the parameter optimization process, use the validation set to evaluate the effect of different hyperparameter combinations and select the best configuration;

[0056] Step s44: Use the test set to evaluate the quality of the generated samples. In this paper, the average cosine similarity (avg_cos_sim) and the average Jensen-Shannon distance (avg_jen_dis) are used as evaluation indicators to measure the fidelity of the synthesized samples and the consistency of the probability distribution with the real microseismic data in waveform morphology, so as to compare the performance of different models in the task of microseismic sequence synthesis. The results are shown in the following table:

[0057] Table 1 The avg_cos_ s imscore and the avg_ jen_dis distance on the seismic dataset

[0058]

[0059] From Table 1, it can be seen that CTTS-GAN achieves the best results on both evaluation criteria, indicating that the proposed model can better synthesize statistical features similar to real samples and generate more realistic seismic waveform data.

[0060] As Figure 5The time-frequency comparison diagram of the part of the real microseismic sample of the embodiment and the model synthetic sample is shown. Specifically, 500 time points are intercepted from the P-wave starting point for each microseismic event, ensuring that each sample completely contains the arrival time information and key waveform features of P waves and S waves. As can be seen from the figure, compared with the seismic waveform samples generated by other models, the CTTS-GAN model considers the similarity between local blocks in the sequence and the mutual influence between multiple channels and the single-channel quality trade-off relationship, and the P wave and S wave waveform form generated by the CTTS-GAN model is more realistic and can better reflect the evolution rule of the waveform on the time axis.

[0061] Step s45: using the trained generator to generate new microseismic samples, and upsampling the new microseismic samples to form synthetic samples of the CTTS-GAN model. Then, the synthetic samples are added to the real microseismic data to expand the original data set, which is used for training of a subsequent seismic detector (CRED) model;

[0062] Step 5: using the seismic detector (CRED) model to identify mine earthquake events

[0063] The application adopts a CNN-RNN seismic detector (CRED) model, as shown in the figure. The model focuses on accurately identifying and classifying features in the microseismic signal, such as the arrival time information of P waves and S waves, and then determining whether the microseismic signal is a mine earthquake. Figure 6

[0064] The model mainly consists of a convolutional neural network (CNN) and a long short-term memory network (LSTM, which belongs to a kind of RNN). First, the frequency spectrum diagram of the three-component seismic waveform obtained through Fourier transform and the corresponding label value are used as the input of the CRED model, the local features of the microseismic waveform are extracted through the convolutional neural network, and the residual connection is used to alleviate the gradient disappearance problem of the network. After the two-dimensional convolution layer, the feature vector is expanded into a sequence and transmitted to the residual block of two bidirectional long short-term memory networks (Bi-LSTM) to learn the time-frequency rule of the main phase of the microseismic signal. The model uses the Adam optimizer, the mean square error as the loss function, and the learning rate adopts a periodic decay strategy during the training process. Through continuous iterative training, the network parameters are optimized, and the model learns the microseismic waveform features.

[0065] In this embodiment, 10,000 real microseismic waveform data are selected and divided into a training set, a validation set and a test set according to a ratio of 7:2:1. The synthetic samples of the CTTS-GAN model in the above step four are mixed with the real microseismic data in the training set according to different expansion ratios to form an expanded training set, which is input into the CRED model to pick up and identify the features of the mine seismic signal. According to the features of the microseismic signal picked up by the CRED model, such as the arrival time of P waves and S waves and the amplitude, whether there is a mine earthquake event is determined based on a preset threshold. ​

[0066] Finally, the performance of the CRED model when the synthetic data generated by different data augmentation methods and the CTTS-GAN model proposed in this paper are mixed with real microseismic data at different expansion ratios is evaluated using the test set. The evaluation indicators should include classification accuracy, recall rate, precision, and F1 value, etc. to ensure the discriminant effect of the model.

[0067] Accuracy: Measures the proportion of samples correctly predicted by the model (true positives and true negatives) to the total number of samples. It focuses on the model's ability to identify positive examples, but for data distribution imbalance, accuracy may be too rough. The formula is:

[0068]

[0069] Where TP represents true positives (actual positive, predicted positive), TN represents true negatives (actual negative, predicted negative), FP represents false positives (actual negative, predicted positive), and FN represents false negatives (actual positive, predicted negative).

[0070] Recall: Measures the proportion of positive samples detected by the model. It focuses on the model's ability to identify positive examples. A good model should find more positive examples, resulting in a higher recall rate. However, high recall rate may result in more false positives. The formula is:

[0071]

[0072] Precision: Measures the proportion of actual positive samples among the samples predicted as positive by the model. It examines the reliability of the model's prediction results. High precision means that the model's positive example judgment is more accurate, but it may miss some positive examples. The formula is:

[0073]

[0074] F1 value: The harmonic mean of precision and recall, used to consider precision and recall comprehensively, especially suitable for situations where there is a trade-off between the two. The formula is:

[0075]

[0076] In the experiments in this section, Accuracy, Recall, Precision, and F1 value are used as performance evaluation indicators for the model.

[0077] Table 2 The results of CRED model under different expansion methods

[0078]

[0079] From Table 2, under different expansion modes, the expansion effect of the CTTS-GAN model is generally better than that of other models, and when the expansion ratio is 40%, the overall indicators of the CTTS-GAN expansion are better.

[0080] Table 3 The results of CRED model under different expansion ratios

[0081]

[0082] From Table 3, with the increase of the proportion of synthetic samples in real samples, the accuracy, precision, recall rate and F1 value of the seismic detector (CRED) model show a trend of first rising and then falling. When the synthetic data is mixed with real data at an expansion ratio of 60%, most indicators reach a high level.

[0083] Through the above detailed process, high-quality synthetic microseismic data can be effectively generated using CTTS-GAN to expand the original data set, and combined with the CRED model to discriminate and monitor the mine seismic situation, thereby improving the overall level of seismic research and monitoring;

[0084] Step 6: Model iteration

[0085] The trained model is applied to actual mine microseismic monitoring, and real-time monitoring of microseismic data is performed. In the actual application process, new microseismic data is continuously collected, and the improved generative adversarial network (CTTS-GAN) and seismic detector (CRED) model are periodically retrained to ensure that the model maintains high precision and high efficiency, and continuously provides reliable support for mine microseismic monitoring.

[0086] Finally, it should be noted that the above examples are only used to illustrate the principles of the present patent application and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A method for quickly identifying mine earthquake events from short-time small-sample microseismic monitoring data, characterized in that The method comprises the following steps: Step 1: Determine the monitoring area, deploy microseismic sensors, plan the equipment layout density to cover the target area, complete the equipment installation, calibration and signal stability test; Step 2: Real-time acquisition of microseismic monitoring data, providing information about microseismic, such as microseismic waveform characteristics, time stamp, source location and magnitude, etc., storing to local or central database and performing regular backup; Step 3: Preprocessing of the collected microseismic monitoring data, including denoising, missing value processing, normalization and feature extraction; Step 4: Dividing the preprocessed data into training set, validation set and test set, respectively for model training, hyperparameter tuning and performance evaluation; Step 5: Constructing an improved generative adversarial neural network (CTTS-GAN), including generator and discriminator: The generator accepts random noise as input and generates synthetic samples similar to real microseismic data; The design principle is: based on the Transformer architecture, a global-local fusion module (GLFM) is designed in the generator, combining multi-head attention mechanism and convolutional self-attention to capture the global dependence and local similarity of microseismic data; The discriminator accepts real microseismic data and samples generated by the generator, and distinguishes their authenticity; The design principle is: the global optimization target and single-channel optimization target are taken as the loss function of the model, and the corresponding discriminator is composed of global discriminator and single-channel discriminator, and the loss of the two discriminators is transmitted to the generator through back propagation to guide the training optimization of the generator; Step 6: Training the improved generative adversarial neural network (CTTS-GAN) through an alternating optimization strategy, specifically including: Fixing the generator parameters, inputting real samples and generated samples to train the discriminator, and updating the discriminator parameters; Fixing the discriminator parameters, inputting the generated samples into the discriminator, and updating the generator parameters based on the discrimination results; Repeat the alternating training until the discrimination degree between the generated samples and the real samples is lower than the preset threshold; Step 7: Adjusting the hyperparameters of the improved generative adversarial neural network (CTTS-GAN), including learning rate, batch size and network layer number, optimizing the model performance through the validation set; Step 8: Using the test set to evaluate the quality of the generated samples, in order to quantitatively compare the effects of different models in the microseismic sequence synthesis task, the average cosine similarity (avg_cos_sim) and the average Jensen-Shannon distance (avg_jen_dis) of the synthesized samples are calculated, to reflect the fidelity of the synthesized samples in waveform form and the consistency with the real microseismic data in probability distribution; Step 9: Dividing the real microseismic waveform data into training set, validation set and test set according to the proportion, and then using the trained improved generative adversarial network (CTTS-GAN) model to synthesize microseismic data, mixing the synthesized microseismic data with the real microseismic waveform data in the training set according to different expansion ratios as the expanded training set, which is used for subsequent training of the seismic detector (CRED) model; Step 10: The seismic detector (CRED) model focuses on accurately identifying and classifying features in microseismic signals, such as P waves and S waves, to distinguish mine seismic events. The expanded data set is input into the CRED model for training, and the network parameters are optimized. Step 11: The expanded data set to be detected is input into the trained seismic detector (CRED) model to extract P wave and S wave features, including P wave and S wave arrival time, amplitude, and to distinguish mine seismic events based on a pre-set threshold. Step 12: The test set is used to evaluate the classification accuracy, recall rate, precision, and F1 value performance indicators of the seismic detector (CRED) model when the synthetic data generated by different data expansion methods and the improved generative adversarial network (CTTS-GAN) proposed in this paper are mixed with real microseismic data at different expansion ratios. Step 13: The trained seismic detector (CRED) model is used for real-time analysis of microseismic data, and new data is continuously collected to iteratively optimize the model.

2. The method according to claim 1, wherein the method is characterized in that: In step 3, during the preprocessing of the original data, noise removal is performed to filter out environmental and equipment noise. The signal-to-noise ratio, continuity, and sampling rate of the data are then checked, and poor-quality or incomplete data segments are removed. Normalization is then performed on the data to ensure that it is within the same dimension range. Finally, key features such as time features (P wave and S wave arrival time), frequency features, and amplitude features are extracted from the waveform data.

3. The method according to claim 1, characterized in that: In step 5, the CTTS-GAN model uses global optimization objectives and single-channel optimization objectives as loss functions. The loss function expression for the generator is: The loss function expression for the discriminator is: The loss function expressions for the global discriminator and the single-channel discriminator are as follows: In the above formula, G(z) represents the synthetic sequence, D0 represents the global discriminator, D i represents the single-channel discriminator, G(z) i represents the i-th variable of the synthetic sequence generated by the generator, p z (z) represents the distribution to which the noise z is subjected, p data represents the joint distribution to which the real sequence x is subjected, p i data represents the distribution of the i-th variable of the real sequence, E represents the expected value, a and b respectively represent the label value of the real sequence and the label value of the synthetic sequence in the training process, ε represents the parameter of label smoothing, generally a = 1, ε = 0.1, and λ is a hyperparameter, by changing the value of λ, the model can balance the two goals of "improving single-channel quality" and "maintaining inter-channel correlation" in the multi-channel synthesis task.

4. The method according to claim 1, wherein the method is characterized in that: The average cosine similarity (avg_cos_sim) in step 8 is an effective method to calculate the statistical feature similarity of the feature vector of the real sequence f i and the feature vector of the synthesized sequence g i This paper considers a more extensive statistical feature, i.e. the median, mean, standard deviation, variance, sum, root mean square, maximum, minimum, absolute maximum in the real sequence and the synthesized sequence. The cosine similarity calculation formula on a single channel is: where j represents the jth statistical feature value in the sequence feature vector, j ∈ (1, 2, … 9); The average cosine similarity is the average value of the corresponding feature vector pairs of the real sequence and the synthetic sequence on all channels, and the expression is as follows: In the above formula, d represents the number of variables in a single multivariate time series, k represents the kth variable, n represents the data set size of the time series, and avg_cos_sim is closer to 1, indicating that the statistical features of the synthetic sequence are closer to those of the real sequence. The average Jensen-Shannon distance (avg_jen_dis) represents the average value of the Jensen-Shannon distance between each statistical feature of the real sequence and the synthetic sequence on all channels, and the probability distribution of the jth statistical feature value f nj of the real sequence feature vector can be represented as f j_real , and the probability distribution of the jth statistical feature value g nj of the synthetic sequence feature vector can be represented as f j_syn . The calculation formula of the Jensen-Shannon distance is as follows: The calculation formula of the average Jensen-Shannon distance is: The smaller the value of avg_jen_dis, the closer the distribution of the synthetic sequence and the real sequence in each statistical feature dimension, and the closer the synthetic sequence to the real sequence.

5. The method according to claim 1, wherein the method is characterized in that: In step 10, the CNN layer of the seismic detector (CRED) model uses multi-scale convolution kernels to extract local features of the waveform, and the RNN layer uses long short-term memory networks (LSTM) to solve the problem of gradient vanishing or gradient explosion in long-term data, thereby better handling the long-term dependency problem of microseismic data.

6. The method according to claim 1, wherein the method is characterized in that: In step 12, different data augmentation methods in this paper refer to noise disturbance augmentation, TTS-GAN augmentation, CTTS-GAN-global augmentation, CTTS-GAN augmentation, and setting the augmentation ratio to 20%, 40% respectively; in order to explore the influence of the number of data set augmentation on the model, the CTTS-GAN model proposed in this paper is used to synthesize data to augment the original real data set at different ratios, to evaluate the influence of the generated data on the model effect, and the overall performance of the model can be optimized by analyzing the performance evaluation indexes of the CRED model in the test set and selecting the appropriate augmentation ratio.

7. The method according to claim 1, wherein the method is characterized by: In step 13, the improved generative adversarial network (CTTS-GAN) and the seismic detector (CRED) model are incrementally trained based on newly collected microseismic data, and the model parameters are updated regularly.

Images