Coal mine micro-seismic source positioning method and device, medium and product

Abstract

The invention discloses a coal mine micro-seismic source positioning method and device, a medium and a product, and relates to the field of micro-seismic positioning, and the method comprises the steps: obtaining the micro-seismic data of a target coal mine; the micro-seismic data comprises micro-seismic waveform data of a preset number of sampling points of the micro-seismic event; according to the micro-seismic data, determining a seismic source position of the target coal mine by using a micro-seismic seismic source prediction model; wherein the micro-seismic source prediction model is obtained by training a micro-seismic source initial prediction model by using a training data set; the training data set comprises training microseismic data of the coal mine for training and corresponding labels; the micro-seismic data for training comprises micro-seismic waveform data of a preset number of sampling points of each micro-seismic event in a preset time period; the label is a three-dimensional Gaussian volume probability of the seismic source position; the micro-seismic source initial prediction model comprises a gating circulation unit, a self-attention mechanism module and a full-connection convolutional neural network which are connected in sequence. According to the invention, the accuracy of coal mine micro-seismic source positioning is improved.

Description

Technical Field

[0001] The present application relates to the field of microseismic positioning, and in particular to a method, equipment, medium and product for locating microseismic sources in coal mines. Background Art

[0002] The dynamic disasters induced by coal mining mainly include mine tremors, gas explosions and coal and gas outbursts. These disasters are usually caused by factors such as stress concentration in the coal seam, rock cracking, and gas escape. They not only threaten the lives of miners, but may also cause equipment damage and economic losses. Therefore, it is crucial to carry out scientific prediction and monitoring and implement preventive measures.

[0003] Microseismic positioning in coal mines is an important tool for preventing dynamic disaster accidents. By quickly and accurately locating microseismic events, it helps to improve the accuracy and efficiency of mine safety monitoring.

[0004] Traditional positioning methods are difficult to meet the requirements of both high efficiency and high precision at the same time, while deep learning can train specific models for coal mine conditions. By automatically extracting time and spatial features, it can effectively locate the source of earthquakes in coal mine microseismic positioning, thereby improving positioning speed and accuracy, and helping to more accurately predict potential dangerous areas in mines.

[0005] Traditional microseismic location methods usually rely on travel time or waveform analysis, such as source scanning algorithm, full waveform time inversion method, earthquake first arrival wave offset summation method, double difference location method and GrowClust, etc. Traditional microseismic location methods face many challenges, especially in effectively utilizing the data of numerous seismic stations to achieve accurate positioning. Although in theory a large number of seismic stations provide more data support for the location of microseismic events, due to the cumulative time error of each station, the increase in the number of stations does not linearly improve the positioning accuracy. Existing studies have shown that usually selecting 6 to 8 stations around the source for data analysis can obtain better positioning results. However, due to the actual distribution of stations, the current coal mine microseismic location uses traditional methods for positioning, and the error is relatively large. Summary of the invention

[0006] The purpose of this application is to provide a method, device, medium and product for locating microseismic sources in coal mines, so as to improve the positioning accuracy of microseismic sources in coal mines.

[0007] To achieve the above objectives, this application provides the following solutions:

[0008] In a first aspect, the present application provides a method for locating a microseismic source in a coal mine, comprising:

[0009] Acquire microseismic data of a target coal mine; the microseismic data includes microseismic waveform data of a preset number of sampling points of microseismic events;

[0010] According to the microseismic data, the source position of the target coal mine is determined using a microseismic source prediction model; wherein the microseismic source prediction model is obtained by training an initial microseismic source prediction model using a training data set; the training data set includes training microseismic data and corresponding labels of training coal mines; the training microseismic data includes microseismic waveform data of a preset number of sampling points of each microseismic event within a preset time period; the label is a three-dimensional Gaussian probability of the source position; the initial microseismic source prediction model includes a gated recurrent unit, a self-attention mechanism module, and a fully connected convolutional neural network connected in sequence.

[0011] Optionally, according to the microseismic data, using a microseismic source prediction model, determining the source location of the target coal mine specifically includes:

[0012] According to the microseismic data, using a microseismic source prediction model, determining the three-dimensional Gaussian volume probability of each sampling point;

[0013] The maximum three-dimensional Gaussian volume probability is solved to obtain the earthquake source position of the target coal mine.

[0014] Optionally, the training data set is used to train the initial microseismic source prediction model, specifically including:

[0015] Create a training dataset;

[0016] Construct an initial prediction model for microseismic sources;

[0017] Taking the training microseismic data of the training coal mine as input and the corresponding labels as output, the microseismic source initial prediction model is trained in combination with the total loss function to obtain the microseismic source prediction model; the total loss function is determined based on the Dice coefficient loss function, the cross entropy loss function and L2 regularization.

[0018] Optionally, a training data set is established, specifically including:

[0019] Acquire training microseismic data of training coal mines;

[0020] Preprocessing the training microseismic data to obtain processed training microseismic data; the preprocessing includes denoising and waveform normalization;

[0021] Dividing the processed training microseismic data into an initial training data set and a test data set according to a set ratio;

[0022] The initial training data set is oversampled and labels are added to obtain the training data set.

[0023] Optionally, the total loss function is expressed as:

[0024] Combined_Loss=α×Dice+β×Cross_Entropy+L2;

[0025] Among them, Combined_Loss is the total loss function value; Dice is the Dice coefficient loss function value; α is the weight of the loss term of the Dice coefficient loss function; Cross_Entropy is the cross entropy loss function value; β is the weight of the loss term of the cross entropy loss function; L2 is L2 regularization.

[0026] Optionally, the self-attention mechanism module includes a multi-head attention mechanism and a feedforward neural network connected in sequence.

[0027] Optionally, the fully connected convolutional neural network includes a first spatial feature extraction module, a second spatial feature extraction module, a third spatial feature extraction module, a fourth spatial feature extraction module, a fifth convolutional layer, a first deconvolutional layer, a second deconvolutional layer, a third deconvolutional layer, a fourth deconvolutional layer and a sixth convolutional layer connected in sequence; the first spatial feature extraction module includes a first convolutional layer, a first maximum pooling layer and a first residual module connected in sequence; the second spatial feature extraction module includes a second convolutional layer, a second maximum pooling layer and a second residual module connected in sequence; the third spatial feature extraction module includes a third convolutional layer, a third maximum pooling layer and a third residual module connected in sequence; the fourth spatial feature extraction module includes a fourth convolutional layer, a fourth maximum pooling layer and a fourth residual module connected in sequence.

[0028] In a second aspect, the present application provides a computer device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement any of the above-described methods for locating the microseismic source of a coal mine.

[0029] In a third aspect, the present application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements any of the above-described methods for locating the microseismic source of a coal mine.

[0030] In a fourth aspect, the present application provides a computer program product, including a computer program, which, when executed by a processor, implements any of the above-described methods for locating the microseismic source of a coal mine.

[0031] According to the specific embodiments provided in this application, this application has the following technical effects:

[0032] The present application provides a method, device, medium and product for locating microseismic sources in coal mines, which obtain microseismic data of a target coal mine; the microseismic data includes microseismic waveform data of a preset number of sampling points of microseismic events; according to the microseismic data, a microseismic source prediction model is used to determine the source position of the target coal mine; wherein the microseismic source prediction model is obtained by training an initial microseismic source prediction model using a training data set; the training data set includes training microseismic data and corresponding labels of the training coal mine; the training microseismic data includes microseismic waveform data of a preset number of sampling points of each microseismic event within a preset time period; the label is a three-dimensional Gaussian probability of the source position; the initial microseismic source prediction model includes a gated recurrent unit, a self-attention mechanism module and a fully connected convolutional neural network connected in sequence. The present application improves the accuracy of microseismic source positioning in coal mines. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0034] Figure 1 A schematic diagram of a flow chart of a method for locating a microseismic source in a coal mine provided in one embodiment of the present application;

[0035] Figure 2 A schematic diagram of the implementation process of the method for locating the microseismic source of a coal mine according to the present application;

[0036] Figure 3 This is a schematic diagram of the distribution of microseismic stations and microseismic events at the 6306 working face of Dongtan Coal Mine;

[0037] Figure 4 Schematic diagram of the original event distribution and oversampled event distribution of the training data set;

[0038] Figure 5 This is the GRU-ATT-ConvNet model architecture diagram;

[0039] Figure 6 This is a graph showing how the accuracy and loss function change with the number of training rounds during model training;

[0040] Figure 7 Schematic diagram of Gaussian body for predicting earthquake source location;

[0041] Figure 8 The slice diagram for predicting the earthquake source location;

[0042] Fig. 9 This is a schematic diagram of the positioning error analysis of the test data set;

[0043] Fig.10 A schematic diagram of the structure of a computer device provided in one embodiment of the present application. DETAILED DESCRIPTION

[0044] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0045] In order to make the above-mentioned objects, features and advantages of the present application more obvious and easy to understand, the present application is further described in detail below with reference to the accompanying drawings and specific implementation methods.

[0046] The present application provides a method for locating the microseismic source of a coal mine, which is used for locating microseismic events in coal mines. The method proposes a gated attention convolutional neural network (Gated Recurrent Unit-Attention-Fully Connected Neural Network, GRU-ATT-ConvNet) model, which combines the gated recurrent unit (Gated Recurrent Unit, GRU) architecture with the fully connected convolutional neural network (Fully Connected Neural Network, FCNN), and adds a self-attention mechanism (Transformer Encoder) to enhance feature extraction, which can accurately locate the location of the microseismic event (i.e., the source location), and show a prediction accuracy rate that is better than other models in similar positioning algorithms. The present application is significantly superior to traditional positioning algorithms in terms of positioning accuracy and calculation time. Based on the training model of existing samples, a new microseismic event can be located and predicted within 0.1s, and its absolute error is within 20 meters. With sufficient training samples, GRU-ATT-ConvNet has the potential to achieve high-precision coal mine microseismic positioning, and its direct positioning accuracy can reach the meter level.

[0047] In an exemplary embodiment, Figure 1 As shown, a method for locating a microseismic source in a coal mine is provided, comprising the following steps:

[0048] S1: Acquire microseismic data of a target coal mine; the microseismic data includes microseismic waveform data of a preset number of sampling points of microseismic events.

[0049] S2: According to the microseismic data, the source position of the target coal mine is determined using a microseismic source prediction model; wherein the microseismic source prediction model is obtained by training an initial microseismic source prediction model using a training data set; the training data set includes training microseismic data and corresponding labels of training coal mines; the training microseismic data includes microseismic waveform data of a preset number of sampling points of each microseismic event within a preset time period; the label is a three-dimensional Gaussian probability of the source position; the initial microseismic source prediction model includes a gated recurrent unit, a self-attention mechanism module, and a fully connected convolutional neural network connected in sequence.

[0050] As an optional implementation, step S2 specifically includes:

[0051] According to the microseismic data, the three-dimensional Gaussian volume probability of each sampling point is determined using a microseismic source prediction model.

[0052] The maximum three-dimensional Gaussian volume probability is solved to obtain the earthquake source position of the target coal mine.

[0053] This application proposes an enhanced fully connected convolutional model, combined with a recursive audit network, named GRU-ATT-ConvNet, to improve the accuracy of coal mine microseismic event location. The model introduces a fast location strategy that only requires 6 stations, and trains the model for coal mine conditions, requiring only one prior information, the event waveform, which solves the problem of requiring a large amount of prior information in traditional methods and improves the convenience and efficiency of location.

[0054] The traditional positioning method used in coal mines is grid search, which divides the entire coal mine into grids and uses all the stations in the coal mine for positioning. Previous studies have shown that selecting 6 to 8 stations around the earthquake source for data analysis can usually obtain better positioning results. The GRU-ATT-ConvNet model first selects historical microseismic data from 6 stations in a specific area to build a data set for model training in a specific area to better suit microseismic prediction in coal mines. A large number of experiments have verified the superior performance of the GRU-ATT-ConvNet model in microseismic positioning tasks, opening up a new path for the application of deep learning technology in the field of microseismic positioning in coal mines. Figure 2 As shown, this application is generally divided into three steps: creating a data set, building and training a model (Gated Recurrent Unit (GRU) extracts temporal features, Transformer Encoder captures long-distance dependencies in time series, enhances feature information, and Convolutional Neural Network (CNN) extracts spatial features), and predicting the earthquake source location.

[0055] As an optional implementation method, the microseismic source initial prediction model is trained using the training data set, specifically including:

[0056] Step 1: Create a training dataset.

[0057] As an optional implementation, step 1 specifically includes:

[0058] Acquire training microseismic data of a training coal mine.

[0059] The training microseismic data is preprocessed to obtain processed training microseismic data; the preprocessing includes denoising processing and waveform normalization processing.

[0060] The processed training microseismic data are divided into an initial training data set and a test data set according to a set ratio.

[0061] The initial training data set is oversampled and labels are added to obtain the training data set.

[0062] In actual application, Dongtan Coal Mine in Shandong Province, China has rock burst risk, so a microseismic monitoring system is deployed in the mine. The microseismic monitoring system consists of 12 single-component monitoring stations (stations). As the working face is mined, 4 stations have been removed, and 2 of the remaining 8 stations have poor waveform quality. The remaining 6 normal operating stations are distributed as follows: Figure 3 As shown, from August 24, 2023 to September 4, 2023, a total of 153 clear waveform microseismic events (including low energy, high energy and blasting events) were collected from the system to construct a data set. All stations are single-component with a sampling rate of 500Hz. This application captures the waveform data 5 seconds before and 10 seconds after the earthquake, for a total of 7,500 sampling points. The input data includes microseismic waveform data of 7,500 sampling points at 6 stations for each microseismic event. Initially, the microseismic data was uniformly denoised and waveform normalized, and then the data set was shuffled and divided into an initial training data set and a test data set in a ratio of 8:2.

[0063] Figure 4 (a) is the original distribution of microseismic events in the training set. It can be found that the data distribution is unbalanced. In the partition of the working face, the maximum number of events in some areas can reach 19, and the minimum is 0. Training with this training data set will cause the model to predict the majority class and ignore the minority class samples, and the generalization ability will be weakened. Therefore, after the data set is shuffled, it is first divided into an initial training data set and a test data set, and then the initial training data set is oversampled. Based on the maximum number of events in the area of ​​19 as the benchmark, some areas with less than 19 events are sampled, except for the areas with no events originally, such as Figure 4 As shown in (b) in the figure, after oversampling the training set, the total number of samples reaches 553 events.

[0064] The scope of the study area (coal mine for training) is 600m×500m×200m, which is divided into 60×50×40 grids, and the grid step is 10m×10m×5m. The training data set constructed in this application is 553 events, 6 stations for each event, and each station records 7500 amplitude data points. The corresponding label set is the three-dimensional Gaussian probability of the source location XYZ, with a size of (60, 50, 40).

[0065] Step 2: Construct an initial microseismic source prediction model.

[0066] In practical applications, the architecture of the GRU-ATT-ConvNet model is as follows Figure 5 As shown in the figure, it consists of a gated recurrent unit GRU, a self-attention mechanism module TransformerEncoder, and a fully connected convolutional neural network FCNN.

[0067] The first part of the model is a three-layer gated recurrent unit (GRU). GRU is an improved recurrent neural network (RNN) for processing sequence data. Its structure is more complex than the standard RNN, and it introduces update gates and reset gates to better capture long-term dependencies in the sequence and effectively alleviate the gradient vanishing problem.

[0068] Update gate z t Control the hidden state h of the previous moment t-1 How to be brought into the current hidden state h t :

[0069] z t =σ(W z ·x t +U z ·h t-1 +b z )(1)

[0070] Among them, W z and U z is the weight matrix of the update gate, b z is the bias term of the update gate, x t is the current input, h t-1 is the hidden state of the previous moment, and σ is the sigmoid function.

[0071] Reset Gate t Control the hidden state h of the previous moment t-1 How much information needs to be forgotten:

[0072] r t =σ(W r ·x t +U r ·ht-1 +b r )(2)

[0073] Among them, W r , U r is the weight matrix of the reset gate; b r is the bias term for the reset gate.

[0074] Candidate hidden state h~ t The calculation is as follows, which uses the reset gate r t To determine how the information from the previous moment is combined with the current input information:

[0075]

[0076] Among them, W r and U r is the weight matrix of the reset gate; b h is the bias of the candidate hidden state, not a parameter of a specific gate; Represents element-wise multiplication.

[0077] The final hidden state h t It is the weighted sum of the candidate hidden state and the hidden state at the previous moment, and the weight is determined by the update gate:

[0078]

[0079] The advantage of GRU is that it can effectively handle dependencies in long time series by deciding how much past information to retain and how much new information to introduce through the gating mechanism. In the GRU-ATT-ConvNet model:

[0080] The first layer of GRU has 64 units, the input shape is (6, 7500), and the output shape is (6, 64).

[0081] The second layer of GRU has 128 units, the input shape is (6, 64), and the output shape is (6, 128).

[0082] The third layer of GRU has 256 units, the input shape is (6, 128), and the output shape is (6, 256).

[0083] The role of these GRU layers is to extract the temporal features of the input sequence, encode these temporal features into higher-dimensional representations, and provide richer features for the subsequent TransformerEncoder.

[0084] The output of GRU is fed into a Transformer Encoder. Transformer Encoder mainly consists of a Multi-Head Attention mechanism and a Feedforward Neural Network (FNN). Transformer Encoder is known for its efficiency and accuracy in processing long sequences, especially its attention mechanism allows the model to focus on different parts of the input sequence and consider the relationship between these parts when encoding.

[0085] Multi-Head Attention: There are 8 heads, and the key dimension of each head is 256. The role of this layer is to calculate the correlation between each position in the input sequence and adjust the representation of each position through these correlations.

[0086] Feedforward Neural Network (FNN): There are two layers, and the dimensions of the fully connected layers are 512 and 256. This network performs further nonlinear transformations on the output of the multi-head attention layer.

[0087] The formula of Transformer Encoder can be expressed as:

[0088]

[0089] Among them, Q, K, and V represent the query matrix, key matrix, and value matrix respectively. k is the dimension of the key vector.

[0090] As an optional implementation, the fully connected convolutional neural network includes a first spatial feature extraction module, a second spatial feature extraction module, a third spatial feature extraction module, a fourth spatial feature extraction module, a fifth convolutional layer, a first deconvolutional layer, a second deconvolutional layer, a third deconvolutional layer, a fourth deconvolutional layer and a sixth convolutional layer connected in sequence; the first spatial feature extraction module includes a first convolutional layer, a first maximum pooling layer and a first residual module connected in sequence; the second spatial feature extraction module includes a second convolutional layer, a second maximum pooling layer and a second residual module connected in sequence; the third spatial feature extraction module includes a third convolutional layer, a third maximum pooling layer and a third residual module connected in sequence; the fourth spatial feature extraction module includes a fourth convolutional layer, a fourth maximum pooling layer and a fourth residual module connected in sequence.

[0091] The features (6, 256, 1) encoded by the Transformer Encoder are passed into a fully connected convolutional neural network (FCNN). This FCNN network consists of a convolutional layer, a pooling layer, and a residual module (ResNet) to further extract features and increase the nonlinear expression ability of the model.

[0092] The convolution layer in the fully connected convolutional neural network FCNN extracts features from the input data through convolution operations. The convolution operation of each layer is completed by a filter (or convolution kernel), which slides on the input data to generate a feature map.

[0093] The input data of the convolutional layer can be two-dimensional (such as the height and width of an image) or multi-dimensional (such as the height, width, and number of channels of an image).

[0094] The filter (convolution kernel) W is a small matrix (such as 3×3 or 5×5) that slides over the input data.

[0095] Convolution operation:

[0096]

[0097] Among them, Y[i,j] is the value of the output feature map, X[i+m,j+n] is the local area of ​​the input data, W[m,n] is the weight of the filter, and b is the bias term.

[0098] The GRU-ATT-ConvNet model consists of multiple convolutional layers, pooling layers, residual modules, and deconvolution, as follows:

[0099] Convolutional layer 1 (the first convolutional layer) has 128 filters, a convolution kernel size of 3×3, tanh as the activation function, and an output shape of (6, 256, 128).

[0100] Max pooling layer 1 (the first max pooling layer), the convolution kernel size is 2×2, and the output shape is (3, 128, 128).

[0101] Residual module 1 (the first residual module), with a convolution kernel size of 3×3 and an activation function of tanh, directly adds the input to the convolution result and then activates it through ReLU.

[0102] Convolutional layer 2 (the second convolutional layer), with 64 filters, a convolution kernel size of 3×3, and an output shape of (3, 128, 64).

[0103] Max pooling layer 2 (second max pooling layer), the convolution kernel size is 3×2, and the output shape is (1, 64, 64).

[0104] Residual module 2 (second residual module) is similar to residual module 1, but the number of convolution kernels is changed to 64.

[0105] Convolutional layer 3 (the third convolutional layer), with 32 filters, a convolution kernel size of 3×3, and an output shape of (1, 64, 32).

[0106] Max pooling layer 3 (the third max pooling layer), the convolution kernel size is 1×16, and the output shape is (1, 4, 64).

[0107] Residual module 3 (the third residual module) is similar to residual module 2, but the number of convolution kernels is changed to 32.

[0108] Convolutional layer 4 (the fourth convolutional layer), 16 filters, convolution kernel size is 3×3, and output shape is (1, 4, 16).

[0109] Max pooling layer 4 (the fourth max pooling layer), the convolution kernel size is 1×4, and the output shape is (1, 1, 16).

[0110] Residual module 4 (the fourth residual module) is similar to residual module 1, but the number of convolution kernels is changed to 16.

[0111] Layer Normalization (LayerNorm) is usually applied after the residual connection to normalize all features of each sample, which helps stabilize the training of deep models.

[0112] Convolution layer 5 (the fifth convolution layer), 2 filters, convolution kernel size is 1×1, and output shape is (1, 1, 2).

[0113] The role of these convolutional layers is to gradually extract spatial features and ensure effective information transfer through residual connections. The pooling layer is used to reduce the spatial dimension of the feature map to prevent overfitting. In order to restore the spatial size of the feature map, the model uses multiple deconvolutional layers (or transposed convolutional layers). The main parameters and output shape of the deconvolutional layer are as follows:

[0114] Deconvolution layer 1 (the first deconvolution layer), with 16 filters, a kernel size of 2×2, a stride of 2×5, and an output shape of (2, 5, 16).

[0115] Deconvolution layer 2 (the second deconvolution layer), with 32 filters, a convolution kernel size of 3×3, a stride of 3×5, and an output shape of (6, 25, 32).

[0116] Deconvolution layer 3 (the third deconvolution layer), with 64 filters, a convolution kernel size of 3×3, a stride of 5×2, and an output shape of (30, 50, 64).

[0117] Deconvolution layer 4 (the fourth deconvolution layer), with 64 filters, a convolution kernel size of 2×2, a stride of 2×1, and an output shape of (60, 50, 64).

[0118] The output layer, i.e., convolution layer 6 (the sixth convolution layer), has 40 filters, a convolution kernel size of 1×1, and an output shape of (60, 50, 40).

[0119] Step 3: Taking the training microseismic data of the training coal mine as input and the corresponding labels as output, combined with the total loss function, the initial microseismic source prediction model is trained to obtain the microseismic source prediction model; the total loss function is determined based on the Dice coefficient loss function, the cross entropy loss function and L2 regularization.

[0120] The total loss function of the initial microseismic source prediction model is defined as:

[0121] Combined_Loss=α×Dice+β×Cross_Entropy+L2(7)

[0122] Among them, Combined_Loss is the total loss function value; Dice is the Dice coefficient loss function value; α is the weight of the loss term of the Dice coefficient loss function; Cross_Entropy is the cross entropy loss function value; β is the weight of the loss term of the cross entropy loss function; L2 is L2 regularization.

[0123] The total loss function used by the GRU-ATT-ConvNet model combines Dice coefficient loss, cross entropy loss, and L2 regularization. The introduction of each part has its specific purpose, which is to improve the performance of the model in different aspects and ensure that the model performs better in specific tasks.

[0124] Dice coefficient loss is mainly used to measure the degree of overlap between two Gaussian bodies (such as predicted values ​​and true values). The formula is:

[0125]

[0126] Cross Entropy Loss is a widely used classification loss function, especially suitable for multi-class classification problems. It measures the difference between the predicted distribution and the true distribution. The formula is:

[0127] Cross_Entropy_Loss=-∑y true log(y pred )(9)

[0128] L2 regularization suppresses overfitting of the model by introducing the sum of squares of model parameters into the loss function. It penalizes large weight values, prompting the model to learn smoother functions, and helps improve the generalization ability of the model and avoid overfitting on the training data. The formula is:

[0129] L2=λΣω 2 (10)

[0130] Among them, λ is the regularization coefficient; ω is the weight parameter.

[0131] By combining Dice coefficient loss and cross entropy loss, the model can focus on accurate classification between categories (driven by cross entropy loss) while maintaining sensitivity to minority classes in imbalanced data (driven by Dice coefficient loss). This combination ensures balanced performance of the model on different categories, especially in the case of imbalanced data. The introduction of L2 regularization helps the model avoid overfitting problems. Since deep learning models usually have a large number of trainable parameters, it is easy to perform too well on training data but perform poorly on test data. Through L2 regularization, the model is forced to learn more general features, reducing the risk of overfitting.

[0132] In the GRU-ATT-ConvNet model, a training strategy based on the Adam optimizer was adopted, the initial learning rate was set to 0.0001, and the mini-batch gradient descent method with a batch size of 32 was used for model training. To optimize the model performance, the ReduceLROnPlateau callback mechanism was applied. When the validation loss did not improve significantly within 5 cycles, the learning rate would be automatically halved. In addition, to prevent the model from overfitting, the Early Stopping strategy was adopted. When the validation loss did not improve for 20 consecutive cycles, the training was stopped and the optimal weights were restored. This comprehensive strategy effectively balanced the speed and stability of model training. A total of 200 cycles of training were conducted. If the model's performance on the validation set exceeded all previous cycles, the model was saved as the current best model. This ensures that the best performing version of the model is always available.

[0133] Figure 6 The changing trends of loss function and accuracy during model training are shown in Figure 2. The blue curve represents training loss, the green dotted line represents validation loss, the red dotted line represents training accuracy, and the orange dotted line represents validation accuracy.

[0134] from Figure 6It can be observed that the training loss of the model drops rapidly in the early training stage, which indicates that the model can learn from the data and reduce the error quickly in the early stage. Similarly, the validation loss also drops rapidly in the early stage, but after about 40 cycles, the validation loss curve gradually flattens out and fluctuates slightly, indicating that the model may have approached the limit of its generalization ability. At the same time, the training accuracy (red dotted line) rises rapidly in the early stages of training and gradually stabilizes in the later stages. The validation accuracy (orange dotted line) also shows a similar trend, approaching a stable state after about 50 cycles. In the end, the best performing version of the model achieved an accuracy of 0.82 in the 150th cycle, reaching a relatively ideal state.

[0135] After training, the best model is applied to predict the earthquake source location. Figure 7 A prediction example in the test set is shown. The maximum probability of the predicted Gaussian body can reach 95.3%. The coordinates of the point with the highest probability calculated by the three-dimensional Gaussian body solution are (23, 16, 8), indicating that this model can predict a source location with a probability of 95.3%.

[0136] Figure 8 yes Figure 7 A 2D slice of YX, YZ, XZ at the point of maximum probability, Figure 8 The red star represents the true earthquake source location, and the slice color represents the probability of the predicted location. The lighter the color, the higher the probability. Figure 8 It can be seen that the predicted location of the example event is accurate and basically consistent with the actual location, and the GRU-ATT-ConvNet model only takes 0.1s to predict an event.

[0137] This application defines the error judgment criteria, including horizontal error (HE), depth error (DE), and absolute error (AE):

[0138]

[0139] DE=|z pred -z true |(12)

[0140]

[0141] The localization error of the test set is Fig. 9As shown in the figure, the horizontal error is low in most cases, mainly concentrated between 0 and 20 meters, indicating that the model has good positioning accuracy in the horizontal direction. However, there are some large error values, which may be caused by a few extreme events or noise data. The depth errors are almost all concentrated between 0 and 10 meters, and a few errors are beyond this range. Compared with the horizontal error, the depth error has a higher concentration, showing good depth prediction ability. The absolute error values ​​are mostly concentrated between 0 and 20 meters. There are some large errors, but the overall error level is still low, showing the stability of the model in overall positioning. As can be seen from the box plot, the medians of the three errors are all low, indicating that the model performs well in most cases. However, there are many outliers in the horizontal error and absolute error, indicating that there are large errors in some cases. The depth error is relatively concentrated and performs the most stable.

[0142] This application can accurately predict the location of microseismic events, and has extremely fast computing efficiency and prediction accuracy. It only takes 0.1s to predict an event, and the accuracy can reach more than 90%. With sufficient training samples, GRU-ATT-ConvNetr has the potential to achieve fast and high-precision positioning, and its direct positioning error can reach the meter level. Improving this capability will be the focus of our future work.

[0143] In an exemplary embodiment, a computer device is provided. The computer device may be a server or a terminal. The internal structure diagram thereof may be as follows: Fig.10 As shown. The computer device includes a processor, a memory, an input / output interface (Input / Output, referred to as I / O) and a communication interface. Among them, the processor, the memory and the input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The input / output interface of the computer device is used to exchange information between the processor and an external device. The communication interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a method for locating the microseismic source of a coal mine is implemented.

[0144] Those skilled in the art will understand that Fig.10The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components. In an exemplary embodiment, a computer device is provided, including a memory and a processor, wherein a computer program is stored in the memory, and the processor implements the above-mentioned method for locating the microseismic source of a coal mine when executing the computer program.

[0145] In an exemplary embodiment, a computer-readable storage medium is provided, storing a computer program, which, when executed by a processor, implements the above-mentioned method for locating the microseismic source of a coal mine.

[0146] In an exemplary embodiment, a computer program product is provided, including a computer program, which implements the above-mentioned method for locating the microseismic source of a coal mine when executed by a processor.

[0147] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with relevant regulations.

[0148] Those of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to the memory, database or other medium used in the embodiments provided in the present application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM may be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM).

[0149] The database involved in each embodiment provided in this application may include at least one of a relational database and a non-relational database. The non-relational database may include a distributed database based on blockchain, etc., but is not limited thereto. The processor involved in each embodiment provided in this application may be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, etc., but is not limited thereto.

[0150] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0151] This article uses specific examples to illustrate the principles and implementation methods of this application. The description of the above embodiments is only used to help understand the method and core ideas of this application. At the same time, for those skilled in the art, according to the ideas of this application, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting this application.

Claims

1. A method for locating microseismic sources in coal mines, characterized in that: include: Acquire microseismic data of a target coal mine; the microseismic data includes microseismic waveform data of a preset number of sampling points of microseismic events; According to the microseismic data, the source position of the target coal mine is determined using a microseismic source prediction model; wherein the microseismic source prediction model is obtained by training an initial microseismic source prediction model using a training data set; the training data set includes training microseismic data and corresponding labels of training coal mines; the training microseismic data includes microseismic waveform data of a preset number of sampling points of each microseismic event within a preset time period; the label is a three-dimensional Gaussian probability of the source position; the initial microseismic source prediction model includes a gated recurrent unit, a self-attention mechanism module, and a fully connected convolutional neural network connected in sequence.

2. The method for locating the microseismic source of a coal mine according to claim 1, characterized in that: According to the microseismic data, the microseismic source prediction model is used to determine the source location of the target coal mine, specifically including: According to the microseismic data, using a microseismic source prediction model, determining the three-dimensional Gaussian volume probability of each sampling point; The maximum three-dimensional Gaussian volume probability is solved to obtain the earthquake source position of the target coal mine.

3. The method for locating the microseismic source of a coal mine according to claim 1, characterized in that: The training data set is used to train the initial microseismic source prediction model, including: Create a training dataset; Construct an initial prediction model for microseismic sources; Taking the training microseismic data of the training coal mine as input and the corresponding labels as output, the microseismic source initial prediction model is trained in combination with the total loss function to obtain the microseismic source prediction model; the total loss function is determined based on the Dice coefficient loss function, the cross entropy loss function and L2 regularization.

4. The method for locating the microseismic source of a coal mine according to claim 3, characterized in that: Create a training data set, including: Acquire training microseismic data of training coal mines; Preprocessing the training microseismic data to obtain processed training microseismic data; the preprocessing includes denoising and waveform normalization; Dividing the processed training microseismic data into an initial training data set and a test data set according to a set ratio; The initial training data set is oversampled and labels are added to obtain the training data set.

5. The method for locating the microseismic source of a coal mine according to claim 3, characterized in that: The expression of the total loss function is: Combined_Loss=α×Dice+β×Cross_Entropy+L2; Among them, Combined_Loss is the total loss function value; Dice is the Dice coefficient loss function value; α is the weight of the loss term of the Dice coefficient loss function; Cross_Entropy is the cross entropy loss function value; β is the weight of the loss term of the cross entropy loss function; L2 is L2 regularization.

6. The method for locating the microseismic source of a coal mine according to claim 1, characterized in that: The self-attention mechanism module includes a multi-head attention mechanism and a feedforward neural network connected in sequence.

7. The method for locating the microseismic source of a coal mine according to claim 1, characterized in that: The fully connected convolutional neural network includes a first spatial feature extraction module, a second spatial feature extraction module, a third spatial feature extraction module, a fourth spatial feature extraction module, a fifth convolutional layer, a first deconvolutional layer, a second deconvolutional layer, a third deconvolutional layer, a fourth deconvolutional layer and a sixth convolutional layer connected in sequence; the first spatial feature extraction module includes a first convolutional layer, a first maximum pooling layer and a first residual module connected in sequence; the second spatial feature extraction module includes a second convolutional layer, a second maximum pooling layer and a second residual module connected in sequence; the third spatial feature extraction module includes a third convolutional layer, a third maximum pooling layer and a third residual module connected in sequence; the fourth spatial feature extraction module includes a fourth convolutional layer, a fourth maximum pooling layer and a fourth residual module connected in sequence.

8. A computer device comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the method for locating the microseismic source of a coal mine as described in any one of claims 1 to 7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method for locating the microseismic source of a coal mine described in any one of claims 1 to 7 is implemented.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the method for locating the microseismic source of a coal mine described in any one of claims 1 to 7 is implemented.

Images