Method for optimizing coal mine micro-seismic positioning based on deep learning
Through deep learning, the LocFormer model combined with convolution and Transformer architecture, the problem of insufficient station selection and accuracy in traditional coal mine microseismic positioning methods is solved, and efficient and high-precision microseismic event positioning is achieved, which is significantly better than the traditional methods.
Patent Information
- Application Number
- CN202410960754.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-17
- Publication Date
- 2024-11-15
AI Technical Summary
Traditional coal mine microseismic positioning methods are difficult to meet the needs of high efficiency and high accuracy at the same time, especially when choosing the best seismic station combination, resulting in insufficient positioning accuracy.
The LocFormer model based on deep learning is adopted, combined with convolutional layer and Transformer architecture, and through data preprocessing, feature extraction and position coding, effective station selection and source position prediction are achieved, improving positioning accuracy and efficiency.
It significantly improves the positioning accuracy and efficiency of coal mine microseismic events, realizes direct positioning errors of meter-level, and is superior to traditional methods in terms of calculation time and accuracy.
Smart Images

Figure CN118962780A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of mine earthquake detection, and specifically provides a method for optimizing coal mine microseismic positioning based on deep learning. 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] Linking dynamic disasters with coal mine microseismic positioning: Coal mine microseismic positioning 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] Connecting deep learning with coal mine microseismic positioning: Traditional positioning methods are difficult to meet the needs of both high efficiency and high precision at the same time. Deep learning can effectively identify event areas in coal mine microseismic positioning through automated feature extraction and cluster analysis, thereby improving positioning speed and accuracy, and helping to more accurately predict potential dangerous areas in mines.
[0005] Traditional earthquake 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, GrowClust, etc. Traditional microseismic location methods face many challenges, especially in effectively utilizing the data of numerous seismic stations to achieve precise 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, selecting the best station combination without prior knowledge of the source area is still the main bottleneck for improving the accuracy of microseismic location in coal mines. Summary of the invention
[0006] The technical task of the present invention is to provide a method for optimizing the microseismic positioning of coal mines based on deep learning in response to the above-mentioned problems. An enhanced Transformer model, combined with a convolutional layer, named LocFormer, is designed to improve the accuracy of the positioning of microseismic events in coal mines. The traditional positioning method used in the mining area is grid search, which divides the entire mining area into grid ranges and uses all stations in the mining area for positioning. The present invention introduces an effective station selection strategy based on the predicted area, which solves the station selection problem in the traditional method and improves the accuracy and efficiency of positioning.
[0007] A method for optimizing coal mine microseismic location based on deep learning, the method combines the Transformer architecture with a convolutional neural network (CNN), uses the convolutional neural network (CNN) to extract spatial features, uses the Transformer architecture to extract temporal features, creates a data set, builds and trains a model, and predicts the source location area to build a LocFormer model for coal mine microseismic event location;
[0008] The implementation steps of the method include:
[0009] Step 1: Create a dataset;
[0010] Step 2: Model construction and training;
[0011] Step 3: Predict the earthquake source location area.
[0012] Furthermore, the dataset creation process of the LocFormer model includes:
[0013] The microseismic waveform data of each monitoring station for each earthquake event in a period are uniformly denoised and normalized, and then the processed data set is divided into a training set and a test set in a ratio of 7:3;
[0014] Furthermore, the construction and training process of the LocFormer model includes:
[0015] Using the K-means clustering algorithm, the microseismic waveform data were divided into 3, 4, 6, 8, 10, 20, 30 and 50 categories according to the Euclidean distance between the epicenters for subsequent training calculations, and the LocFormer model was constructed and trained. The initial part of the LocFormer model architecture consists of 5 convolutional layers, which are used to preprocess the input microseismic signals.
[0016] Each layer of the LocFormer model sets specialized parameters to extract and abstract spatial features from the raw data, where:
[0017] The first convolutional layer uses 64 3x3 filters with a stride of 1 and ReLU as the activation function.
[0018] Convolutional layers 2-5: Due to the increasing complexity of subsequent layers, 128 filters are used in the second and third convolutional layers, 256 filters are used in the fourth convolutional layer, and 512 filters are used in the fifth convolutional layer, while keeping the convolution kernel size of 3x3 and the stride of 1 unchanged;
[0019] The activation function still uses ReLU.
[0020] Furthermore, pooling layers are arranged after the first, second and fifth convolutional layers, wherein the pooling layer after the first convolutional layer reduces the spatial dimension through a 2x2 maximum pooling operation to enhance the robustness of the features; the pooling layer after the second and fifth convolutional layers uses a 2x2 maximum pooling operation to further compress the feature representation.
[0021] The CNN part contains pooling layers to reduce the dimensionality of features. This dimensionality reduction not only reduces the computational burden of subsequent processes, but also improves the generalization ability of the model by making it more robust to changes in the input data. Pooling layers help to aggregate the most important features extracted by the convolutional layers and further refine the feature representation.
[0022] Furthermore, the convolutional layer applies a set of learnable filters and then introduces nonlinearity through the ReLU activation function, allowing the LocFormer model to capture complex spatial relationships in the data, where:
[0023] For the input earthquake sequence X, the convolution operation of the lth layer is defined as follows:
[0024] X′ l+1 =f(X′ l *W′ l +b′ l ) (1)
[0025] In the proposed model, X′ l Indicates the first th The input of the layer, and W′ l and b′ l Respectively represent the first th The weights and biases of the convolutional layer of the layer, f represents the nonlinear activation function, that is, the ReLU activation function, which means performing the operation f(x)=max(0,x);
[0026] X′ l+1 It is the first th Output feature map of the layer;
[0027] The convolution operation is used to gradually extract and abstract spatial features from the original seismic data in preparation for time series processing.
[0028] Furthermore, after the method extracts spatial features from the original seismic data through a convolution operation in a convolution layer, the LocFormer model implements feature mapping of the data by introducing a fully connected layer;
[0029] The fully connected layer maps the high-dimensional feature vector output by the convolutional layer to a new space to make it suitable for the subsequent Transformer encoder;
[0030] The LocFormer model maps the feature sequence enhanced by the convolutional layer to a dimension suitable for processing by the Transformer encoder through a fully connected layer.
[0031] Furthermore, the method retains the temporal information of the sequence by adding position encoding before inputting the original seismic data into the Transformer encoder;
[0032] The position encoding is implemented by adding a position vector to the embedding vector of each input element;
[0033] The position vector has the same dimension as the embedding vector and can be added directly;
[0034] The position encoding uses a combination of sine and cosine functions to generate a unique code for each dimension, where:
[0035] The position encoding of even index 2i uses the sine function Calculate, where d m Represents the total dimensions available for positional encoding, Pos represents the position of the element in the sequence, and i represents an even index;
[0036] For the position code of the odd index 2i+1, the cosine function is used to derive the position code, and the formula is as follows:
[0037]
[0038] Furthermore, the core of the LocFormer model is the Transformer encoder, which includes 6 encoder layers processing space, involving multi-head self-attention components and feedforward networks, residual connections and layer normalization, output layers and loss functions, where:
[0039] Multi-head self-attention component: Each encoder layer contains a multi-head self-attention component with eight heads, enabling the LocFormer model to focus on different parts of the earthquake sequence;
[0040] Feed-forward network: followed by a feed-forward network (FFN) with an internal dimension of 2048 to further process the attention-conditioned data; wherein the feed-forward network (FFN) applies ReLU activation to introduce non-linearity;
[0041] Residual connections: Both the multi-head self-attention component and the feed-forward network (FFN) are enhanced with residual connections to facilitate the flow of gradients during training;
[0042] Layer Normalization: The residual is followed by application of layer normalization to stabilize the learning process through feature normalization;
[0043] Output layer: The processed data passes through another fully connected layer to convert the output of the Transformer encoder into three logits corresponding to the earthquake location category;
[0044] The softmax layer in the neural network converts these logits into probabilities;
[0045] Loss function: The loss function of the LocFormer model combines cross entropy to evaluate the difference between the prediction accuracy and the actual label, and L2 regularization to prevent overfitting by penalizing large weights. This balanced approach ensures high prediction performance and strong generalization ability on different earthquake datasets.
[0046] Furthermore, the core of the Transformer model is its attention mechanism, which allows the LocFormer model to uniquely evaluate the importance of each sequence position; through the attention mechanism, an element can measure its association with every other element, thereby enhancing the LocFormer model's ability to contextualize information. The implementation steps are as follows:
[0047] (1) Calculate query, key and value:
[0048] For an input sequence X, the query (Q att ), key (K att ) and value (V att ) is obtained through linear transformation, the formula is as follows:
[0049] Q att =Xw Q ,K att =XW K , V att =XW V (2)
[0050] Among them, WQ, W K and W V is a learnable weight matrix corresponding to query, key, and value respectively;
[0051] (2) Calculate attention weight:
[0052] The attention weight A is obtained by calculating the dot product of the query Q and the key K and then applying the softmax function. The formula is as follows:
[0053]
[0054] where dk is the dimension of the key vector; the dot product of Q and key K is Divide by to scale the size of the dot product to prevent the softmax function from entering the saturation region due to excessive input, thereby alleviating the gradient vanishing problem;
[0055] (3) Generate output:
[0056] The output of the self-attention mechanism is obtained by combining the attention weights with their corresponding values; after this operation, the feed-forward network of each encoder layer takes over and continues to process the data received from the self-attention stage;
[0057] The feedforward network consists of a pair of linear transformations, with a ReLU activation function in between to enhance the nonlinearity of the transformation process. The formula is as follows:
[0058] FFN(z)=max(0,ZW1+b1)W2+b2 (4)
[0059] Among them, W1 and W2 are weight matrices of linear transformation, b1 and b2 are bias terms, and max(0,x) represents the ReLU activation function; each sublayer in Transformer (including self-attention layer and feedforward network layer) is followed by a residual connection and layer normalization step; given the input x, the output of the sublayer operation is Sublayer(x), and the output y after residual connection and layer normalization is:
[0060] y=LayerNorm(x+Sublayer(x)) (5)
[0061] 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.
[0062] The loss function of the LocFormer model combines cross entropy to evaluate the difference between the predicted accuracy and the actual label, and L2 regularization to prevent overfitting by penalizing large weights. This balanced approach ensures high prediction performance and strong generalization ability on different earthquake datasets.
[0063] Furthermore, the design of the loss function is crucial for effectively training the earthquake location model. The loss function used by the model combines cross entropy loss and L2 regularization to ensure that the LocFormer model not only strives to reduce the difference between the predicted probability and the actual label to improve classification accuracy, but also limits the size of the LocFormer model weights through regularization, thereby reducing the risk of overfitting and enhancing the generalization ability of the LocFormer model. Through a carefully designed loss function, the LocFormer model can effectively learn and predict coal mine microseismic locations, achieving high accuracy and good generalization performance. The calculation formula is as follows:
[0064]
[0065] in, represents the composite loss function, y represents the actual category label, Indicates the jth th The i-th data point th The predicted probability of the class;
[0066] The weights of the LocFormer model are denoted by θ, N is the number of data points, C represents the number of different categories, and M quantifies the total number of parameters in the LocFormer model; the regularization parameter is denoted by λ; the principal component cross entropy loss penalizes the LocFormer model when the predicted probability distribution deviates from the true distribution; a secondary component is the L2 regularization term, called weight decay, which penalizes the size of the parameters, scaled by λ, to promote learning more generalizable patterns and mitigate the risk of overfitting.
[0067] Furthermore, after the training, the best LocFormer model is applied to the prediction of the earthquake source location area to predict the earthquake source location area;
[0068] Statistics of different classifications and their corresponding accuracy rates were conducted.
[0069] According to the prediction results of different categories, data from certain seismic stations in a specific area are selected for secondary precise positioning.
[0070] Compared with the prior art, the method for optimizing coal mine microseismic positioning based on deep learning in the present invention has the following outstanding beneficial effects:
[0071] The proposed method can accurately identify the area of microseismic events and shows better prediction accuracy than other models in the same classification. By carefully selecting appropriate stations for secondary precise positioning, it significantly outperforms traditional positioning algorithms in terms of positioning accuracy and calculation time. With sufficient training samples, LocFormer has the potential to achieve high-precision classification, and its direct positioning error can reach the meter level. BRIEF DESCRIPTION OF THE DRAWINGS
[0072] Figure 1 It is a schematic flow chart of the method of the present invention;
[0073] Figure 2 This is the distribution map of microseismic stations in Dongtan Coal Mine;
[0074] Figure 3 It is the distribution of microseismic events and the three classification diagrams of microseismic events;
[0075] Figure 4 This is the LocFormer model architecture diagram;
[0076] Figure 5 This is a graph showing the change in accuracy and loss function over the number of training rounds in three-category training;
[0077] Figure 6 It is the correct prediction graph of the probability distribution of three-classification events;
[0078] Figure 7 It is the probability distribution error prediction diagram of three-classification events. DETAILED DESCRIPTION
[0079] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments.
[0080] A method for optimizing microseismic positioning in coal mines based on deep learning, such as Figure 1 As shown in the figure, the implementation of the method is divided into three steps: creating a data set, building and training a model (convolutional neural network (CNN) extracts spatial features, Transformer extracts temporal features), and predicting the earthquake source location area:
[0081] Step 1: Create a dataset.
[0082] Dongtan Coal Mine in Shandong Province, China, is at risk of rock burst, so a microseismic monitoring system was deployed in the mine. The monitoring system consists of 27 single-component monitoring stations, which are distributed as follows: Figure 2 , 3 As shown, there are 12 stations in the category 1 area, 9 stations in the category 2 area, and 6 stations in the category 3 area. From January 21, 2023 to January 21, 2024, we collected data of a total of 14,178 microseismic events from the system to construct the dataset. All microseismic monitoring stations are single-component with a sampling rate of 500Hz. We captured 10 seconds of event waveform data, with a total of 5000 sampling points. The input data includes 5000 points of microseismic waveform data from 27 channels for each seismic event. Initially, the microseismic data underwent uniform denoising and waveform normalization, and then the dataset was divided into training and test sets in a ratio of 7:3.
[0083] During the period from January 21, 2023 to January 21, 2024, we used the K-means clustering algorithm to divide the data into 3, 4, 6, 8, 10, 20, 30 and 50 categories according to the Euclidean distance between the epicenters for subsequent training calculations. However, since there are three mining areas in operation at the same time in Dongtan Coal Mine, microseismic events are mainly distributed in and around these three mining faces. Therefore, the results of classifying all microseismic events into 3 categories will be shown. Labels 1 to 3 correspond to microseismic events in three different geographical areas, represented by blue, green and purple, respectively, as shown in Figure 1. Figure 3 The bright dots represent events with energies below 10,000 joules, and the dark dots represent events with energies above 10,000 joules.
[0084] Step 2: Model construction and training.
[0085] The model architecture of LocFormer is as follows Figure 4 As shown in Figure 2, the initial part consists of five convolutional layers, which are designed to preprocess the input microseismic signals. Each layer sets specific parameters to extract and abstract spatial features from the raw data.
[0086] Convolutional layer 1: The first layer uses 64 3x3 filters with a stride of 1 and ReLU as the activation function. The subsequent 2x2 max pooling operation reduces the spatial dimension and enhances the robustness of the features.
[0087] Convolutional layers 2-5: The complexity of subsequent layers increases, with 128 filters used in the second and third layers, 256 filters in the fourth layer, and 512 filters in the fifth layer, while keeping the kernel size of 3x3 and the stride of 1 unchanged. ReLU is still used as the activation function. The second and fifth layers also include 2x2 max pooling to further compress the feature representation.
[0088] CNN layers excel at extracting spatial features from input seismic waveforms. By applying various filters on the input data, these layers are able to identify patterns and features in seismic signals, such as edges, textures, and other spatially related information. Each convolutional layer applies a set of learnable filters and then introduces nonlinearity through the ReLU activation function, allowing the model to capture complex spatial relationships in the data. For the input earthquake sequence X, the convolution operation of layer l is defined as follows:
[0089] X′ l+1 =f(X′ l *W′ l +b′ l ) (1)
[0090] In the proposed model, X′ l Indicates the first th The input of the layer, and W′ l and b′ l Respectively represent the first th The weights and biases of the convolutional layer of the layer. f represents the nonlinear activation function, i.e., ReLU, which means performing the operation f(x) = max(0, x). X′ l+1 It is the first th Output feature map of the convolution layer. This series of convolution operations gradually extracts and abstracts spatial features from the raw seismic data in preparation for time series processing. The CNN part contains pooling layers to reduce the dimensionality of the features. This dimensionality reduction not only reduces the computational burden of subsequent processes, but also improves the generalization ability of the model by making it more robust to changes in the input data. Pooling layers help to summarize the most important features extracted by the convolution layers and further refine the feature representation.
[0091] After a series of operations in the convolutional layer to extract spatial features from the original seismic data, the model introduces a fully connected layer to realize feature mapping of the data.
[0092] Fully connected layer: Its function is to map the high-dimensional feature vector output by the convolutional layer to a new space, making it suitable for subsequent Transformer encoder processing. The model uses the fully connected layer to map the feature sequence enhanced by the convolutional layer to a dimension suitable for Transformer encoder processing.
[0093] Before the data is fed into the Transformer encoder, position encoding is added to preserve the temporal information of the sequence. Position encoding is usually implemented by adding a position vector to the embedding vector of each input element. This position vector has the same dimension as the embedding vector and can be added directly. Position encoding uses a combination of sine and cosine functions to generate a unique encoding for each dimension.
[0094] Positional encoding: The positional encoding of even index 2i uses a sine function Calculate, where d m Represents the total dimensions available for positional encoding, Pos represents the position of the element in the sequence, and i represents an even index.
[0095] For odd index 2i+1, the positional encoding is derived using the cosine function
[0096] The core of LocFormer is the Transformer encoder, which processes spatially and temporally rich data through 6 encoder layers:
[0097] Multi-head attention and feed-forward network: Each encoder layer contains a multi-head self-attention mechanism with eight heads, which enables the model to focus on different parts of the earthquake sequence. It is followed by a feed-forward network (FFN) with an internal dimension of 2048 to further process the attention-conditioned data. ReLU activation is applied in FFN to introduce non-linearity.
[0098] Residual connections and layer normalization: Both the attention component and FFN are enhanced with residual connections to facilitate the flow of gradients during training. Layer normalization is applied after the residual to stabilize the learning process through feature normalization.
[0099] At the heart of the Transformer model is its attention mechanism, which allows the model to uniquely assess the importance of each sequence position. Through the attention mechanism, an element can weigh its relevance to every other element, thereby enhancing the model's ability to contextualize information.
[0100] This is achieved by following these steps:
[0101] Compute queries, keys, and values:
[0102] For an input sequence X, the query (Q att ), key (K att ) and value (V att ) is obtained by linear transformation:
[0103] Q att =XW Q , K att =XW K , V att =XW V (2)
[0104] Among them, W Q , W K and W V are learnable weight matrices corresponding to queries, keys, and values, respectively.
[0105] Calculate the attention weights:
[0106] The attention weight A is obtained by calculating the dot product of the query Q and the key K and then applying the softmax function:
[0107]
[0108] Here, d k is the dimension of the key vector. The dot product is Dividing by ℓ to scale the size of the dot product prevents the softmax function from entering the saturation region due to excessive input, thereby alleviating the gradient vanishing problem.
[0109] Build output:
[0110] The output of the self-attention mechanism is obtained by combining the attention weights with their corresponding values. After this operation, the feed-forward network of each encoder layer takes over and continues to process the data received from the self-attention stage. The network consists of a pair of linear transformations interspersed with a ReLU activation function to enhance the nonlinearity of the transformation process:
[0111] FFN(z)=max(0,ZW1+b1)W2+b2 (4)
[0112] Among them, W1 and W2 are the weight matrices of the linear transformation, b1 and b2 are bias terms, and max(0,x) represents the ReLU activation function. Each sublayer in the Transformer (including the self-attention layer and the feedforward network layer) is followed by a residual connection and layer normalization step. Given an input x, the output of the sublayer operation is Sublayer(x), and the output y after residual connection and layer normalization is:
[0113] y=LayerNorm(x+Sublayer(x))(5)
[0114] 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.
[0115] Output layer and loss function: The processed data passes through another fully connected layer to convert the output of the Transformer encoder into three logits corresponding to the earthquake location category. The softmax layer converts these logits into probabilities. The model's loss function combines cross entropy to evaluate the difference between the predicted accuracy and the actual label, and L2 regularization to prevent overfitting by penalizing large weights. This balanced approach ensures high prediction performance and strong generalization ability on different earthquake datasets.
[0116] The design of the loss function is crucial for effectively training the earthquake location model. The loss function used by the model combines cross entropy loss and L2 regularization. This combination ensures that the model not only strives to reduce the difference between the predicted probability and the actual label to improve classification accuracy, but also limits the size of the model weights through regularization, thereby reducing the risk of overfitting and enhancing the generalization ability of the model. Through a carefully designed loss function, the LocFormer model can effectively learn and predict coal mine microseismic locations, achieving high accuracy and good generalization performance.
[0117]
[0118] In formula (6), represents the composite loss function, where y represents the actual category label, Indicates the jth th The i-th data point th The predicted probability of a class. The weights of the model are denoted by θ, N is the number of data points, C represents the number of different classes, and M quantifies the total number of parameters in the model. The regularization parameter is denoted by λ. The principal component cross entropy loss penalizes the model when the predicted probability distribution deviates from the true distribution. A secondary component is the L2 regularization term, often called weight decay, which penalizes the size of the parameters, scaled by λ, to encourage learning more generalizable patterns and mitigate the risk of overfitting.
[0119] The batch size was set to 64, the Adam optimizer was used, and the initial learning rate was set to 1e-4. As the training progressed, we implemented a learning rate warm-up strategy, gradually increasing the learning rate to 1e-4 in the initial epoch, and then fine-tuning the training process by gradually reducing the learning rate based on the performance feedback of the validation set. The loss function combines cross entropy loss and L2 regularization to promote the generalization ability of the model. A total of 350 epochs were trained, and if the model's performance on the validation set exceeded all previous epochs, the model was saved as the current best model. This ensures that we always have the best performing version of the model.
[0120] Figure 5 The figure shows the changes in accuracy and loss function during training, where the green curve represents the loss function value and the red curve represents the accuracy. It can be observed that in the early stages of training, the loss function drops rapidly, indicating that the model is able to learn quickly and effectively extract features from the data, thereby quickly reducing the prediction error. The red curve represents the accuracy, which also rises rapidly in the early stages of training as the loss function decreases. However, after about 150 cycles, the rising trend of the accuracy tends to stabilize, indicating that the model's learning of the training data has been close to saturation. Throughout the training process, the accuracy showed an overall upward trend, with some fluctuations in the second half. This may be due to changes in the model's adaptability to specific data batches in certain cycles, or temporary instability in the optimization process caused by the learning rate adjustment. In the end, the best performing version of the model achieved an accuracy of 0.96, reaching an ideal state.
[0121] Step 3: Predict the earthquake source location area.
[0122] After the training is completed, the best model is applied to the prediction of the earthquake source location area. The following figure lists the probability distribution of correct and incorrect prediction results, such as Figure 5 The white triangle indicates the epicenter. Figure 6 Shown is a plot of correct predictions, where the probabilities are mostly distributed on the correct classification, with the probability of the correct classification being higher than 99%; Figure 7 The results of misclassification are shown, where the probability distribution is more scattered, with the probabilities of the three regions being 43%, 28%, and 29% respectively.
[0123] The statistics of different classifications and their corresponding accuracy are shown in Table 1. It can be observed that as the number of classification categories increases, the prediction accuracy of the model decreases: when the events are divided into 3 categories, the accuracy of randomly selecting an area is 33.3%, while our model prediction accuracy is 96.7%, which is significantly higher than the random accuracy. When the events are divided into 50 categories, although the model prediction accuracy drops to 37.3%, the positioning resolution is greatly improved and is still much higher than the random accuracy of 2%. This performance degradation is mainly due to the reduction in the number of training samples for each category, resulting in insufficient training. In theory, if there is a sufficient number of training samples, our model can achieve high-resolution direct positioning.
[0124] Table 1 Prediction accuracy of different categories
[0125] category 3 4 6 8 10 20 30 50 Accuracy 96.7% 94.7% 90.7% 82.4% 69.4% 53.8% 43.1% 37.3% Random Accuracy 33.3% 25% 16.7% 12.5% 10% 5% 3.33% 2%
[0126] Finally, based on the prediction results of the three categories, we selected the data of certain seismic stations in a specific area for secondary precise positioning. In areas 1 and 2, the six stations closest to the mining face were selected for secondary positioning, while all stations were selected in area 3. For the entire mining area, a layered velocity model based on CRUST1.0 was adopted, and PhaseNet was used for P-wave arrival time annotation. We compared the accuracy and calculation time of direct positioning by grid search of all stations with the grid search of only some stations after optimization by LocFormer. As shown in Table 2, the method of the present invention is significantly better than the traditional positioning method.
[0127] Table 2 Comparison of positioning accuracy and efficiency between LocFormer and traditional methods
[0128]
[0129] The embodiments described above are only preferred specific implementations of the present invention. Common changes and substitutions made by those skilled in the art within the scope of the technical solution of the present invention should be included in the protection scope of the present invention.
Claims
1. A method for optimizing microseismic positioning in coal mines based on deep learning, It is characterized in that The method combines the Transformer architecture with a convolutional neural network, uses the convolutional neural network to extract spatial features, uses the Transformer architecture to extract temporal features, and constructs a LocFormer model for locating microseismic events in coal mines by creating a data set, building and training a model, and predicting the source location area. The implementation steps of the method include: Step 1: Create a dataset; Step 2: Model construction and training; Step 3: Predict the earthquake source location area.
2. A method for optimizing coal mine microseismic positioning based on deep learning according to claim 1, It is characterized in that The dataset creation process of the LocFormer model includes: The microseismic waveform data of each monitoring station for each earthquake event in a period are uniformly denoised and waveform normalized, and then the processed data set is divided into a training set and a test set in a ratio of 7:
3.
3. A method for optimizing coal mine microseismic positioning based on deep learning according to claim 1 or 2, It is characterized in that The construction and training process of the LocFormer model includes: Using the K-means clustering algorithm, the microseismic waveform data were divided into 3, 4, 6, 8, 10, 20, 30 and 50 categories according to the Euclidean distance between the epicenters for subsequent training calculations, and the LocFormer model was constructed and trained. The initial part of the LocFormer model architecture consists of 5 convolutional layers, which are used to preprocess the input microseismic signals. Each layer of the LocFormer model sets parameters to extract and abstract spatial features from the raw data, where: The first convolutional layer uses 64 3x3 filters with a stride of 1 and ReLU as the activation function. The second and third convolutional layers use 128 filters, the fourth convolutional layer uses 256 filters, and the fifth convolutional layer uses 512 filters, while keeping the convolution kernel size of 3x3 and the stride of 1 unchanged; The activation function uses ReLU.
4. A method for optimizing coal mine microseismic positioning based on deep learning according to claim 3, It is characterized in that The first, second and fifth convolutional layers are followed by pooling layers, wherein the pooling layer after the first convolutional layer reduces the spatial dimension through a 2x2 maximum pooling operation to enhance the robustness of the features; the pooling layer after the second and fifth convolutional layers uses a 2x2 maximum pooling operation to further compress the feature representation.
5. A method for optimizing coal mine microseismic positioning based on deep learning according to claim 4, It is characterized in that The convolutional layers apply a set of learnable filters and then introduce nonlinearity through the ReLU activation function, allowing the LocFormer model to capture complex spatial relationships in the data, where: For the input earthquake sequence X, the convolution operation of the lth layer is defined as follows: X′ l+1 =f(X′ l *W′ l +b′ l ) (1) In the proposed model, X′ l Indicates the first th The input of the layer, and W′ l and b′ l Respectively represent the first th The weights and biases of the convolutional layer of the layer, f represents the nonlinear activation function, which is the ReLU activation function, indicating the execution of the operation f(x) = max(0,x); X′ l+1 It is the first th Output feature map of the layer; Extracting and abstracting spatial features from raw seismic data through the convolution operation to prepare for time series processing; The method extracts spatial features from raw seismic data through convolution operations in a convolutional layer, and the LocFormer model implements feature mapping of the data by introducing a fully connected layer; The fully connected layer maps the high-dimensional feature vector output by the convolutional layer to a new space to make it suitable for the subsequent Transformer encoder; The LocFormer model maps the feature sequence enhanced by the convolutional layer to a dimension suitable for processing by the Transformer encoder through a fully connected layer.
6. A method for optimizing coal mine microseismic positioning based on deep learning according to claim 5, It is characterized in that The method adds position encoding to retain the temporal information of the sequence before inputting the original seismic data into the Transformer encoder; The position encoding is implemented by adding a position vector to the embedding vector of each input element; The position vector has the same dimension as the embedding vector and can be added directly; The position encoding uses a combination of sine and cosine functions to generate a unique code for each dimension, where: The position encoding of even index 2i uses the sine function Calculate, where d m Represents the total dimensions available for positional encoding, Pos represents the position of the element in the sequence, and i represents an even index; For the position code of the odd index 2i+1, the position code is derived using the cosine function, and the formula is as follows:
7. A method for optimizing coal mine microseismic positioning based on deep learning according to claim 5, It is characterized in that The core of the LocFormer model is the Transformer encoder, which includes 6 encoder layers processing space, involving multi-head self-attention components and feedforward networks, residual connections and layer normalization, output layers and loss functions, where: Multi-head self-attention component: Each encoder layer contains a multi-head self-attention component with eight heads, enabling the LocFormer model to focus on different parts of the earthquake sequence; Feed-forward network: followed by a feed-forward network with an internal dimension of 2048 to further process the attention-conditioned data; wherein the feed-forward network introduces non-linearity by applying ReLU activations; Residual connections: Both the multi-head self-attention component and the feed-forward network are enhanced with residual connections to facilitate the flow of gradients during training; Layer Normalization: The residual is followed by application of layer normalization to stabilize the learning process through feature normalization; Output layer: The processed data passes through another fully connected layer to convert the output of the Transformer encoder into three logits corresponding to the earthquake location category; The softmax layer in the neural network converts these logits into probabilities; Loss Function: The loss function of the LocFormer model combines cross entropy to evaluate the difference between the predicted accuracy and the actual label, and L2 regularization to prevent overfitting by penalizing large weights.
8. A method for optimizing coal mine microseismic positioning based on deep learning according to claim 7, It is characterized in that The core of the Transformer model is its attention mechanism, which allows the LocFormer model to uniquely evaluate the importance of each sequence position; through the attention mechanism, an element can measure its association with every other element, thereby enhancing the LocFormer model's ability to contextualize information. The implementation steps are as follows: (1) Calculate query, key and value: For an input sequence X, the query (Q att ), key (K att ) and value (V att ) is obtained through linear transformation, the formula is as follows: Q att =XW Q , K att =XW K , V att =XW V (2) Among them, W Q , W K and W V is a learnable weight matrix corresponding to query, key, and value respectively; (2) Calculate attention weight: The attention weight A is obtained by calculating the dot product of the query Q and the key K and then applying the softmax function. The formula is as follows: Among them, d k is the dimension of the key vector; the dot product of Q and the key K is Divide by to scale the size of the dot product to prevent the softmax function from entering the saturation region due to excessive input, thereby alleviating the gradient vanishing problem; (3) Generate output: The output of the self-attention mechanism is obtained by combining the attention weights with their corresponding values; after this operation, the feed-forward network of each encoder layer takes over and continues to process the data received from the self-attention stage; The feedforward network consists of a pair of linear transformations, with a ReLU activation function in between to enhance the nonlinearity of the transformation process. The formula is as follows: FFN(z)=max(0,ZW 1 +b 1 )W 2 +b 2 (4) Among them, W 1 and W 2 is the weight matrix of the linear transformation, b 1 and b 2 is the bias term, max(0,x) represents the ReLU activation function; each sublayer in the Transformer is followed by a residual connection and layer normalization step; given the input x, the output of the sublayer operation is Sublayer(x), and the output y after residual connection and layer normalization is: y=LayerNorm(x+Sublayer(x)) (5) Layer Normalization (LayerNorm) is usually applied after the residual connection to normalize all features of each sample.
9. The method for optimizing coal mine microseismic positioning based on deep learning according to claim 7, It is characterized in that The loss function combines cross entropy loss and L2 regularization to ensure that the LocFormer model not only aims to reduce the difference between the predicted probability and the actual label to improve classification accuracy, but also limits the size of the LocFormer model weights through regularization, thereby reducing the risk of overfitting and enhancing the generalization ability of the LocFormer model; the calculation formula is as follows: in, represents the composite loss function, y represents the actual category label, Indicates the jth th The i-th data point th The predicted probability of the class; The weights of the LocFormer model are denoted by θ, N is the number of data points, C represents the number of different categories, and M quantifies the total number of parameters in the LocFormer model; the regularization parameter is denoted by λ; the principal component cross entropy loss penalizes the LocFormer model when the predicted probability distribution deviates from the true distribution; the secondary component is the L2 regularization term, which is weight decay and penalizes the size of the parameters, scaled by λ, to promote learning more generalizable patterns and mitigate the risk of overfitting.
10. A method for optimizing coal mine microseismic positioning based on deep learning according to claim 9, It is characterized in that After the training, the best LocFormer model is applied to the prediction of the earthquake source location area to predict the earthquake source location area; Statistics of different classifications and their corresponding accuracy rates were conducted. According to the prediction results of different categories, data from certain seismic stations in a specific area are selected for secondary precise positioning.
Citation Information
Cited By
Well-ground combined enhanced micro-seismic positioning method, device, medium and product
CN119620193A
Coal mine earthquake time sequence feature prediction method based on multi-source data space-time diagram convolutional network
CN120011792A
Method for automatically inverting seismic slip distribution based on deep learning
CN121500386A
Mine micro-seismic multivariable time sequence prediction method based on double-flow structure
CN121679672A
Microseismic event positioning method and system based on gating feature fusion and graph attention hybrid network
CN122410605A
