Microseismic source positioning method, device, medium and equipment based on deep learning

Abstract

The application discloses a microseismic source positioning method and device based on deep learning, a medium and equipment, relates to the technical field of rock mass engineering safety monitoring, and can improve the positioning efficiency and positioning accuracy of the microseismic source. The method comprises the following steps: acquiring three-component microseismic waveform data collected by multiple microseismic monitoring stations; inputting the three-component microseismic waveform data into a seismic phase picking model; performing local convolution feature extraction on the three-component microseismic waveform data through an encoder; performing time sequence correlation feature extraction on the local convolution features through a BiLSTM module; performing global context feature extraction on the time sequence correlation features through a GC Block module; and performing seismic phase identification on the time sequence correlation features and the global context features through a decoder; constructing an event-station graph based on the seismic phase identification result; and inputting the event-station graph into a preset source positioning model to perform source positioning.

Description

Technical Field

[0001] This invention relates to the field of rock mass engineering safety monitoring technology, and in particular to a method, device, medium and equipment for locating microseismic sources based on deep learning. Background Technology

[0002] As rock mass engineering extends into deeper areas, complex environments such as high ground stress and high ground temperature can easily induce engineering disasters such as rockbursts, mine tremors, and landslides, directly threatening on-site construction safety. Microseismic monitoring is a core technical means for early warning of rockbursts in mines, prevention and control of geological disasters, and safety assurance of deep engineering projects. Microseismic phase pickup and source location are the two core links in the entire microseismic monitoring process, directly determining the reliability of disaster early warning and the accuracy of location.

[0003] Currently, the grid search method is commonly used for microseismic source location. However, the grid search method requires discretizing the study area and performing theoretical travel time calculations and error fitting at each candidate grid point. As the required positioning accuracy increases, the grid size must be reduced, but this leads to an exponential increase in computation time, resulting in a huge computational burden and excessive time consumption. Furthermore, the grid search method is prone to getting trapped in local optima, leading to low positioning accuracy of microseismic sources. Summary of the Invention

[0004] This invention provides a method, device, medium, and equipment for locating microseismic sources based on deep learning, which mainly improves the location efficiency and accuracy of microseismic sources.

[0005] According to a first aspect of the present invention, a deep learning-based method for locating microseismic sources is provided, comprising: In response to the microseismic source location signal for the target coal mine, three-component microseismic waveform data collected by multiple microseismic monitoring stations for the target coal mine are acquired respectively. A preset seismic phase picking model is obtained, wherein the preset seismic phase picking model includes an encoder for extracting local convolutional features, a bottleneck layer for extracting temporal features and global context information, and a decoder for extracting microseismic phases. The bottleneck layer includes a BiLSTM (Bidirectional Long Short-Term Memory) module and a GC Block (Global Context Block) module. The three-component microseismic waveform data of each microseismic monitoring station are input into the preset seismic phase picking model. The encoder extracts local convolutional features from the three-component microseismic waveform data. The BiLSTM module in the bottleneck layer extracts temporal correlation features from the local convolutional features. The GC Block module in the bottleneck layer extracts global context features from the temporal correlation features. The decoder identifies the seismic phases of the temporal correlation features and the global context features to obtain the seismic phase identification results for each microseismic monitoring station. A preset seismic source location model is obtained, wherein the preset seismic source location model includes a feature embedding layer for feature embedding, a graph feature extraction layer for graph feature extraction, a global readout layer for extracting global features, and a coordinate regression layer for seismic source location. The three-dimensional position coordinates of each of the microseismic monitoring stations are obtained, the arrival time of the seismic waves at each of the microseismic monitoring stations is determined based on the seismic phase identification results, and an event-station diagram is constructed based on the three-dimensional position coordinates and the absolute value of the seismic wave arrival time difference. The event-station map is input into the preset seismic source location model. The hidden state of high-dimensional nodes is extracted from the event-station map through the feature embedding layer. The deep node features of the hidden state of high-dimensional nodes are extracted through the graph feature extraction layer. The global representation vector of the deep node features is extracted through the global readout layer. The seismic source location result of the target coal mine is obtained by using the global representation vector through the coordinate regression layer.

[0006] Optionally, before obtaining the preset seismic phase picking model, the method further includes: Construct a pre-defined initial seismic phase picking model; Obtain a natural seismic phase dataset and a coal mine seismic phase dataset, wherein both the natural seismic phase dataset and the coal mine seismic phase dataset include sample three-component microseismic waveform data with annotation information, the annotation information being the seismic phase probability distribution curve; The natural earthquake phase dataset is divided into a natural earthquake phase training set, a natural earthquake phase validation set, and a natural earthquake phase test set. The preset initial earthquake phase picking model is trained using the natural earthquake phase training set, the trained preset initial earthquake phase picking model is validated using the natural earthquake phase validation set, and the preset initial earthquake phase picking model that meets the validation requirements is tested using the natural earthquake phase test set. Finally, the preset initial earthquake phase picking model that meets the test conditions is taken as the preset general earthquake phase picking model. The model parameters of the preset general seismic phase picking model are fine-tuned using the coal mine seismic phase dataset, and the preset general seismic phase picking model with fine-tuned parameters is used as the preset seismic phase picking model.

[0007] Optionally, before obtaining the preset seismic source location model, the method further includes: Construct a pre-defined initial earthquake source location model; Obtain natural earthquake source datasets and coal mine earthquake source datasets, wherein both the natural earthquake source datasets and the coal mine earthquake source datasets include sample event-station maps with labeled information, the labeled information being the location of the earthquake source; The natural earthquake source dataset is divided into a natural earthquake source training set, a natural earthquake source validation set, and a natural earthquake source test set. The preset initial earthquake source location model is trained using the natural earthquake source training set, validated using the natural earthquake source validation set, and tested using the natural earthquake source test set. Finally, the preset initial earthquake source location model that meets the test conditions is taken as the preset general earthquake source location model. The model parameters of the preset general seismic source location model are fine-tuned using the coal mine seismic source dataset, and the preset general seismic source location model with fine-tuned parameters is used as the preset seismic source location model.

[0008] Optionally, an event-station diagram is constructed based on the three-dimensional location coordinates and the absolute value of the seismic wave arrival time difference, including: Determine the reference microseismic monitoring stations, and determine the relative arrival time difference of the corresponding microseismic monitoring stations based on the arrival time of the seismic waves at each of the microseismic monitoring stations and the arrival time of the seismic waves at the reference microseismic monitoring stations; In each microseismic monitoring station, K nearest neighbor microseismic monitoring stations corresponding to any given microseismic monitoring station are identified. Based on the three-dimensional position coordinates, the Euclidean distance and the absolute value of the seismic wave arrival time difference between each microseismic monitoring station and its corresponding K nearest neighbor microseismic monitoring stations are determined. An event-station map is constructed based on the three-dimensional location coordinates, the Euclidean distance, the relative time difference, and the absolute value of the seismic wave time difference. The microseismic monitoring station is used as a node in the event-station map, the three-dimensional location coordinates and the relative time difference are used as node features of the corresponding node, and the Euclidean distance and the absolute value of the seismic wave time difference are used as edge features of the edges connecting the microseismic monitoring station to its K nearest neighbor microseismic monitoring stations.

[0009] Optionally, high-dimensional node hidden state extraction is performed on the event-station graph through a feature embedding layer, including: The feature embedding layer sequentially performs linear mapping, normalization, and random discarding on the node features in the event-station graph to obtain the high-dimensional hidden state of each node. The graph feature extraction layer performs deep node feature extraction on the hidden state of the high-dimensional node, including: The hidden state of the high-dimensional node is iteratively updated by multi-layer concatenated convolutional blocks in the graph feature extraction layer. For each node, during the message passing process of each layer concatenated convolutional block, the hidden state of the high-dimensional node of the neighboring node of the node is fused with the edge features corresponding to the edge between the node and the neighboring node. The fusion result is then transformed by a multilayer perceptron to obtain the deep node features of the node. The global readout layer extracts global representation vectors from the deep node features, including: The global readout layer uses a global average pooling function to aggregate the deep node features of each node to obtain a global representation vector. The source location result of the target coal mine is obtained by performing source location on the global representation vector through the coordinate regression layer, including: The global representation vector is mapped and regressed through the linear output layer in the coordinate regression layer to obtain the source location result containing the mean of the three-dimensional coordinate prediction of the source.

[0010] Optionally, after acquiring three-component microseismic waveform data collected from multiple microseismic monitoring stations for the target coal mine, the method further includes: For the three-component microseismic waveform data collected by each microseismic monitoring station, determine the covariance matrix that is common to the three-component microseismic waveform data. The covariance matrix is ​​decomposed into eigenvalues ​​to obtain multiple eigenvalues ​​and eigenvectors corresponding to each eigenvalue. The main polarization direction corresponding to the three-component microseismic waveform data is determined based on each eigenvector. Each component of the three-component microseismic waveform data is projected onto the main polarization direction to obtain the main component microseismic waveform data. The main component microseismic waveform data is then subjected to forward filtering, signal inversion, reverse filtering, and signal inversion again in sequence to obtain the preprocessed main component microseismic waveform data. Based on the main polarization direction, the preprocessed main component microseismic waveform data is decomposed into preprocessed three component microseismic waveform data. The three-component microseismic waveform data of each of the aforementioned microseismic monitoring stations are input into the preset seismic phase picking model, including: The preprocessed three-component microseismic waveform data corresponding to each of the microseismic monitoring stations are input into the preset seismic phase picking model.

[0011] Optionally, the three-component microseismic waveform data of each of the microseismic monitoring stations are input into the preset phase picking model, including: For each single-component microseismic waveform data in the three-component microseismic waveform data, the waveform complexity and signal-to-noise ratio of each single-component microseismic waveform data are determined respectively; the weighting coefficient of the corresponding single-component microseismic waveform data is determined based on the waveform complexity and the signal-to-noise ratio; and the single-component microseismic waveform data are weighted and fused based on the weighting coefficient to obtain the microseismic waveform fusion feature. The microseismic waveform fusion features of each of the microseismic monitoring stations are input into the preset seismic phase picking model.

[0012] According to a second aspect of the present invention, a microseismic source localization device based on deep learning is provided, comprising: The waveform acquisition unit is used to acquire three-component microseismic waveform data collected by multiple microseismic monitoring stations for the target coal mine in response to the microseismic source location signal for the target coal mine. The phase model acquisition unit is used to acquire a preset phase picking model, wherein the preset phase picking model includes an encoder for extracting local convolutional features, a bottleneck layer for extracting temporal features and global context information, and a decoder for extracting microseismic phases. The bottleneck layer includes a BiLSTM (Bidirectional Long Short-Term Memory) module and a GC Block (Global Context Block) module. The phase identification unit is used to input the three-component microseismic waveform data of each microseismic monitoring station into the preset phase picking model, extract local convolutional features from the three-component microseismic waveform data through the encoder, extract temporal correlation features from the local convolutional features through the BiLSTM module in the bottleneck layer, extract global context features from the temporal correlation features through the GC Block module in the bottleneck layer, and identify the phase through the temporal correlation features and the global context features through the decoder to obtain the phase identification result corresponding to each microseismic monitoring station. The positioning model acquisition unit is used to acquire a preset seismic source positioning model, wherein the preset seismic source positioning model includes a feature embedding layer for feature embedding, a graph feature extraction layer for graph feature extraction, a global readout layer for extracting global features, and a coordinate regression layer for seismic source positioning. A construction unit is used to obtain the three-dimensional position coordinates of each of the microseismic monitoring stations, determine the arrival time of the seismic waves at each of the microseismic monitoring stations based on the seismic phase identification results, and construct an event-station diagram based on the three-dimensional position coordinates and the absolute value of the seismic wave arrival time difference. The seismic source location unit is used to input the event-station map into the preset seismic source location model, extract the hidden state of high-dimensional nodes from the event-station map through a feature embedding layer, extract deep node features from the hidden state of high-dimensional nodes through a graph feature extraction layer, extract global representation vectors from the deep node features through a global readout layer, and perform seismic source location on the global representation vectors through a coordinate regression layer to obtain the seismic source location result of the target coal mine.

[0013] According to a third aspect of the present invention, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the above-described deep learning-based microseismic source localization method.

[0014] According to a fourth aspect of the present invention, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the above-described deep learning-based microseismic source localization method.

[0015] This invention provides a method, apparatus, medium, and device for microseismic source localization based on deep learning. Compared with current methods for microseismic source localization using grid search, this invention simultaneously utilizes waveform information in three directions—east-west, north-south, and vertical—to capture seismic wave polarization characteristics more completely than single-component picking, thus improving the accuracy of seismic phase identification. By reading waveforms twice—in both forward and reverse directions—using BiLSTM, the discrete convolutional features output by the encoder are concatenated into continuous and complete temporal features. This solves the problem that pure convolutional networks can only see local windows and cannot understand the context of waveforms, significantly improving the accuracy of seismic wave arrival time picking. Based on ordered temporal features, GC... The Block module calculates global attention weights and extracts global contextual features, highlighting effective phase components and suppressing noise. This cascaded design, which first models the temporal sequence and then globally weights the signals, allows the attention mechanism to function based on ordered features, making it more accurate than directly applying attention to the original waveform. By uniformly encoding the distribution of stations in physical space and wave propagation time constraints into a graph structure, dependence on wave velocity models is avoided, thereby improving the accuracy of source location. The graph feature extraction layer in the preset source location model can comprehensively analyze node and edge features, thus improving the accuracy of source location. The above-mentioned source location process of this invention can be completed in one step without global optimization, and avoids getting trapped in local optima, thus improving the efficiency and accuracy of source location. Attached Figure Description

[0016] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings: Figure 1 A flowchart of a deep learning-based microseismic source localization method provided by an embodiment of the present invention is shown. Figure 2 A flowchart of another microseismic source localization method based on deep learning provided by an embodiment of the present invention is shown; Figure 3 A schematic diagram of a microseismic source localization device based on deep learning provided in an embodiment of the present invention is shown. Figure 4 A schematic diagram of another microseismic source localization device based on deep learning provided in an embodiment of the present invention is shown. Figure 5 A schematic diagram of the physical structure of a computer device provided in an embodiment of the present invention is shown. Detailed Implementation

[0017] The present invention will be described in detail below with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in the present application can be combined with each other.

[0018] Currently, the grid search method for locating microseismic sources is time-consuming and labor-intensive. Furthermore, the grid search method is prone to getting trapped in local optima, resulting in low accuracy in locating microseismic sources.

[0019] To address the aforementioned problems, embodiments of the present invention provide a microseismic source localization method based on deep learning, such as... Figure 1 As shown, the method includes: 101. In response to the microseismic source location signal for the target coal mine, acquire three-component microseismic waveform data collected by multiple microseismic monitoring stations for the target coal mine.

[0020] In this embodiment of the invention, N microseismic monitoring stations are pre-deployed underground and on the surface of the target coal mine. Each station is equipped with a three-component accelerometer to collect microseismic waveform data in the east-west (X component), north-south (Y component), and vertical (Z component) directions, respectively, as three-component microseismic waveform data. For example, the sampling rate can be set to 1000Hz, each waveform segment to be 3 seconds long, for a total of 3000 sampling points. By simultaneously utilizing waveform information in the east-west, north-south, and vertical directions, this embodiment of the invention captures the polarization characteristics of seismic waves more completely than single-component acquisition, thus improving the accuracy of phase identification.

[0021] 102. Obtain the preset seismic phase picking model, which includes an encoder for extracting local convolutional features, a bottleneck layer for extracting temporal features and global context information, and a decoder for extracting microseismic phases. The bottleneck layer includes a BiLSTM (Bidirectional Long Short-Term Memory) module and a GC Block (Global Context Block) module.

[0022] In this embodiment of the invention, the preset phase picking model is based on U-Net, and the encoder consists of three concatenated convolutional blocks. Each convolutional block contains a 7×1 convolutional kernel, a ReLU activation function, and a max pooling layer. A cascaded fusion structure of BiLSTM and GC Block is designed at the bottleneck layer, that is, BiLSTM and GC Block are sequentially connected in series at the bottleneck layer.

[0023] 103. Input the three-component microseismic waveform data of each microseismic monitoring station into the preset phase picking model. The encoder extracts local convolutional features from the three-component microseismic waveform data. The BiLSTM module in the bottleneck layer extracts temporal correlation features from the local convolutional features. The GC Block module in the bottleneck layer extracts global context features from the temporal correlation features. The decoder identifies the phases based on the temporal correlation features and global context features to obtain the phase identification results for each microseismic monitoring station.

[0024] In this embodiment of the invention, the three-component microseismic waveform data is first convolved sequentially by multiple convolutional blocks in the encoder to obtain the local convolutional features output by the last convolutional block. Then, based on these local convolutional features, the temporal correlation of the waveform is extracted bidirectionally using BiLSTM, integrating the local convolutional features obtained from the encoder into continuous and complete temporal correlation features. Next, the GC Block constructs global context information based on these temporal correlation features, calculates attention weights, and highlights effective seismic phase components to obtain global context features. This design of first temporal modeling and then global weighting allows global attention to function based on ordered temporal features, better suppressing noise and highlighting seismic phase signals. Finally, the temporal correlation features and the global context features are decoded by the decoder to obtain the seismic phase identification result. This seismic phase identification result is a three-component probability distribution, corresponding to the predicted probability values ​​of P-wave, S-wave, and background noise, respectively.

[0025] Specifically, in the downsampling encoding stage of the left half of the preset phase picking model, the encoder convolution kernel size can be set to 7, the stride to 1, and the max pooling layer kernel size and stride to 4. As the network deepens, the feature size gradually decreases. In the bottleneck layer, BiLSTM and GC Block are sequentially connected, where the temporal resolution of the features is the lowest and the receptive field is the largest. BiLSTM concatenates the hidden states of the two LSTM layers and outputs them. GC Block uses 1×1 convolutions and the Softmax function to generate attention weights, which are then weighted and summed on the input features to obtain a global context vector. This vector is then non-linearly mapped through two 1×1 convolutions containing LayerNorm (layer normalization), and finally, the transformed global features are fused back into the original feature stream through an addition operation. The right half of the network is the upsampling decoding stage. The decoder first performs upsampling through deconvolution layers, allowing the temporal correlation features and the global context feature size to be gradually restored. The upsampled temporal correlation features and the global context features are concatenated with the features of the corresponding layer of the encoder in the channel dimension, and then processed again through a one-dimensional convolutional block. The convolutional kernel parameters are consistent with those in the encoding stage, such as a size of 7. The output layer of the model is a 1×1 convolutional layer, which maps the number of channels of the features to 3. Finally, in the inference stage, the output is processed by the Softmax activation function to normalize the values ​​to the [0,1] interval, and the final seismic phase probability distribution curve is obtained as the seismic phase recognition result.

[0026] This invention uses BiLSTM to read waveforms twice, in both forward and reverse directions, to concatenate the discrete convolutional features output by the encoder into continuous and complete temporal features. This solves the problem that pure convolutional networks can only see local windows and cannot understand the context of the waveform, significantly improving the accuracy of seismic arrival. Based on ordered temporal features, the GC Block module calculates global attention weights and extracts global contextual features, highlighting effective phase components and suppressing noise. This cascaded design of first modeling the temporal sequence and then globally weighting allows the attention mechanism to function based on ordered features, which is more accurate than directly adding attention to the original waveform.

[0027] 104. Obtain a preset source location model, wherein the preset source location model includes a feature embedding layer for feature embedding, a graph feature extraction layer for graph feature extraction, a global readout layer for extracting global features, and a coordinate regression layer for source location.

[0028] In this embodiment of the invention, the feature embedding layer includes a linear layer, a batch normalization layer, and a Dropout layer; the graph feature extraction layer consists of four cascaded GINEConv (Graph Isomorphism Network with Edge featuresConv) convolutional blocks, each of which integrates a two-layer MLP (Multi-Layer Perceptron) as an update function. The MLP contains a "linear layer-batch normalization-ReLU-linear layer" structure, which aims to enhance the expressive power of feature transformation; the global average pooling function serves as the global readout layer; and the coordinate regression layer contains two fully connected network layers.

[0029] 105. Obtain the three-dimensional position coordinates of each microseismic monitoring station, determine the arrival time of the seismic waves at each microseismic monitoring station based on the seismic phase identification results, and construct an event-station diagram based on the three-dimensional position coordinates and the arrival time of the seismic waves.

[0030] In this embodiment of the invention, the three-dimensional spatial coordinates of each station are accurately measured using a total station or positioning device. The phase identification results include the P-wave arrival time (the time when the P-wave first arrives at station i) and the S-wave arrival time (the time when the S-wave first arrives at station j). The arrival times of all stations are organized into an arrival time vector. For example, if the phase identification results are a phase probability distribution curve, the time step corresponding to the maximum probability of the P-wave is determined as the P-wave arrival time, and the time step corresponding to the maximum probability of the S-wave is determined as the S-wave arrival time. The S-wave arrival time and the P-wave arrival time are collectively referred to as the seismic wave arrival time. Then, the stations are used as graph nodes, the distance between stations is used as the edge between nodes, and the absolute value of the seismic wave arrival time difference between two nodes is used as the feature of the edge between nodes to construct an event-station graph. This embodiment of the invention improves the location accuracy of the seismic source by uniformly encoding the distribution of stations in physical space and the wave propagation time constraints into a graph structure, avoiding dependence on wave velocity models.

[0031] 106. Input the event-station map into the preset seismic source location model, extract the hidden state of high-dimensional nodes from the event-station map through the feature embedding layer, extract the deep node features from the hidden state of high-dimensional nodes through the graph feature extraction layer, extract the global representation vector from the deep node features through the global readout layer, and obtain the seismic source location result of the target coal mine through the coordinate regression layer.

[0032] In this embodiment of the invention, the feature embedding layer first maps the node features in the event-station graph to a high dimension through a linear layer. Then, it sequentially passes through a one-dimensional batch normalization, ReLU activation function, and a Dropout layer with a dropout rate of 0.2 to complete the initial high-dimensional mapping of features and obtain the high-dimensional node hidden state. Next, the high-dimensional node hidden state is input to the graph feature extraction layer. The high-dimensional node hidden state is then processed sequentially through the linear layer-batch normalization-ReLU-linear layer results contained within the MLP in each convolutional layer of the graph feature extraction layer to obtain deep node features. This process enhances the expressive power of the feature transformation. In each convolutional layer of the graph feature extraction layer... During information transmission, the hidden states of neighboring nodes and their corresponding edge features in the event-station graph are first fused using element-wise addition, then aggregated after ReLU activation. Each convolutional output uses the ReLU activation function to ensure non-linear representation during deep feature extraction. Next, a global average pooling function is used as the global readout layer, aggregating the deep node features of all nodes into a high-dimensional global representation vector. Finally, a coordinate regression layer is used as the linear output layer, mapping this global representation vector to a low-dimensional output (source location result). The low-dimensional output corresponds to multi-dimensional Gaussian distribution parameters, where the first three dimensions are the predicted mean of the three-dimensional coordinates of the source, and the last three dimensions are the corresponding logarithmic variance. This embodiment of the invention, through the graph feature extraction layer in the preset source location model, can comprehensively analyze node and edge features, thereby improving the source location accuracy. The above-mentioned source location process of this invention can be completed in one step, without global optimization, and avoids getting trapped in local optima, thus improving the source location efficiency and accuracy.

[0033] This invention provides a deep learning-based microseismic source localization method. Compared with current methods using grid search, this invention simultaneously utilizes waveform information from three directions—east-west, north-south, and vertical—to capture seismic wave polarization features more completely than single-component picking, thus improving the accuracy of phase identification. By reading waveforms twice—in both forward and reverse directions—using BiLSTM, the discrete convolutional features output by the encoder are concatenated into continuous and complete temporal features. This solves the problem that pure convolutional networks can only see local windows and cannot understand the context of waveforms, significantly improving the accuracy of seismic wave arrival time picking. Based on the ordered temporal features, GC... The Block module calculates global attention weights and extracts global contextual features, highlighting effective phase components and suppressing noise. This cascaded design, which first models the temporal sequence and then globally weights the signals, allows the attention mechanism to function based on ordered features, making it more accurate than directly applying attention to the original waveform. By uniformly encoding the distribution of stations in physical space and wave propagation time constraints into a graph structure, dependence on wave velocity models is avoided, thereby improving the accuracy of source location. The graph feature extraction layer in the preset source location model can comprehensively analyze node and edge features, thus improving the accuracy of source location. The above-mentioned source location process of this invention can be completed in one step without global optimization, and avoids getting trapped in local optima, thus improving the efficiency and accuracy of source location.

[0034] Furthermore, to better illustrate the process of locating microseismic sources described above, and as a refinement and extension of the above embodiments, this invention provides another method for locating microseismic sources based on deep learning, such as... Figure 2 As shown, the method includes: 201. In response to the microseismic source location signal for the target coal mine, acquire three-component microseismic waveform data collected by multiple microseismic monitoring stations for the target coal mine.

[0035] In this embodiment of the invention, to improve the data quality of the three-component microseismic waveform data, it is necessary to preprocess the three-component microseismic waveform data. Based on this, the method includes: determining the covariance matrix common to the three-component microseismic waveform data collected by each microseismic monitoring station; performing eigenvalue decomposition on the covariance matrix to obtain multiple eigenvalues ​​and eigenvectors corresponding to each eigenvalue; determining the principal polarization direction common to the three-component microseismic waveform data based on each eigenvector; projecting each component of the three-component microseismic waveform data onto the principal polarization direction to obtain principal component microseismic waveform data; sequentially performing forward filtering, signal inversion, inverse filtering, and signal inversion again on the principal component microseismic waveform data to obtain preprocessed principal component microseismic waveform data; and decomposing the preprocessed principal component microseismic waveform data into preprocessed three-component microseismic waveform data based on the principal polarization direction.

[0036] Specifically, the covariance matrix corresponding to the three-component microseismic waveform data is determined, and eigenvalue decomposition is performed on this covariance matrix to obtain multiple eigenvalues ​​and eigenvectors corresponding to each eigenvalue. The direction of the eigenvector with the largest eigenvalue is selected as the principal polarization direction. The three-component microseismic waveform data is then projected onto the principal polarization direction to obtain the principal component microseismic waveform data. Further, the principal component microseismic waveform data undergoes forward phase filtering, followed by time-based signal inversion. After inversion, the principal component microseismic waveform data undergoes inverse phase filtering, and the processed principal component microseismic waveform data is then time-based signal inversion to obtain the preprocessed principal component microseismic waveform data. Finally, the processed principal component microseismic waveform data is mapped back to the three-component space along the principal polarization direction to obtain the preprocessed three-component microseismic waveform data. This invention reduces the dimensionality of the three-component data into a single principal component with optimal signal-to-noise ratio by projecting along the main polarization direction. Then, it removes noise through zero-phase filtering without introducing any phase delay. Finally, it is decomposed back into three components. This significantly improves the signal-to-noise ratio while completely preserving the time accuracy of the seismic wave arrival, fundamentally solving the problem of inaccurate phase arrival picking in strong noise environments. The preprocessed three-component microseismic waveform data is then input into a preset phase picking model for phase picking, which improves the accuracy of phase picking.

[0037] 202. Obtain the preset phase picking model, which includes an encoder for extracting local convolutional features, a bottleneck layer for extracting temporal features and global context information, and a decoder for extracting microseismic phases. The bottleneck layer includes a BiLSTM (Bidirectional Long Short-Term Memory) module and a GC Block (Global Context Block) module.

[0038] In this embodiment of the invention, to obtain a preset seismic phase picking model, it is necessary to train and construct the preset seismic phase picking model in advance. Based on this, the method includes: constructing a preset initial seismic phase picking model; obtaining a natural seismic phase dataset and a coal mine seismic phase dataset, wherein both the natural seismic phase dataset and the coal mine seismic phase dataset include sample three-component microseismic waveform data with annotation information, the annotation information being a seismic phase probability distribution curve; dividing the natural seismic phase dataset into a natural seismic phase training set, a natural seismic phase validation set, and a natural seismic phase test set; training the preset initial seismic phase picking model using the natural seismic phase training set; validating the trained preset initial seismic phase picking model using the natural seismic phase validation set; testing the preset initial seismic phase picking model that meets the validation requirements using the natural seismic phase test set; finally, using the preset initial seismic phase picking model that meets the test conditions as a preset general seismic phase picking model; fine-tuning the model parameters of the preset general seismic phase picking model using the coal mine seismic phase dataset; and using the preset general seismic phase picking model with fine-tuned parameters as the preset seismic phase picking model.

[0039] The natural seismic phase dataset includes various types of microseismic data from nature, covering a wide range and containing a large amount of data. The coal mine seismic phase dataset, on the other hand, has a relatively small sample size, containing only coal mine microseismic data. Specifically, during model training, a pre-defined initial seismic phase picking model is first constructed. Then, the natural and coal mine seismic phase datasets are acquired, ensuring all necessary files are included. The data is converted to a format understandable by the pre-defined initial seismic phase picking model. The natural seismic phase dataset is then used to train, validate, and test the model. Specifically, the dataset can be partitioned first: using randomness or a specific strategy (such as stratified sampling), the natural seismic phase dataset is divided into training, validation, and test sets. The model is then trained using the training set, validated using the validation set, and tested using the test set to evaluate its performance on unseen data. Precision, recall, and other metrics on the test set are calculated and recorded. If the model performance does not meet requirements, it can return to the training phase for further iterations or adjustments. This process yields a pre-defined general seismic phase picking model that meets the requirements. Furthermore, a general model is trained, validated, and tested using a coal mine seismic phase dataset. Specifically, the dataset can be first divided: using randomness or a specific strategy (such as stratified sampling), the coal mine seismic phase dataset is divided into a training set, a validation set, and a test set. The training set is then used to train the model to fine-tune its parameters. The validation set is used to validate the general model, and the test set is used to test the validated model and evaluate its performance on unseen data. Precision, recall, and other metrics on the test set are calculated and recorded. If the performance of the general model does not meet the requirements, it can be returned to the training phase for further iterations or adjustments. This process yields a pre-defined seismic phase picking model that meets the requirements. This embodiment of the invention pre-trains the model using a natural seismic phase dataset to learn general seismic phase characteristics, and then fine-tunes it using a coal mine seismic phase dataset to adapt to specific noise environments, achieving a high-precision seismic phase picking model even with limited coal mine microseismic data.

[0040] Furthermore, the trained preset phase picking model is stored, and when phase picking is needed, the preset phase picking model can be directly retrieved based on the storage location.

[0041] 203. Input the three-component microseismic waveform data of each microseismic monitoring station into the preset phase picking model. The encoder extracts local convolutional features from the three-component microseismic waveform data. The BiLSTM module in the bottleneck layer extracts temporal correlation features from the local convolutional features. The GC Block module in the bottleneck layer extracts global context features from the temporal correlation features. The decoder identifies the phases based on the temporal correlation features and global context features to obtain the phase identification results for each microseismic monitoring station.

[0042] In this embodiment of the invention, in order to extract more implicit information from the three-component microseismic waveform data, it is necessary to perform fusion processing on the three-component microseismic waveform data. Based on this, the method includes: for each single-component microseismic waveform data in the three-component microseismic waveform data, determining the waveform complexity and signal-to-noise ratio of each single-component microseismic waveform data; determining the weight coefficient of the corresponding single-component microseismic waveform data based on the waveform complexity and the signal-to-noise ratio; performing weighted fusion on the single-component microseismic waveform data based on the weight coefficient to obtain microseismic waveform fusion features; and inputting the microseismic waveform fusion features of each microseismic monitoring station into the preset phase picking model.

[0043] Specifically, the three-component microseismic waveform data includes east-west single-component microseismic waveform data, north-south single-component microseismic waveform data, and vertical single-component microseismic waveform data. Taking any single-component microseismic waveform data from the east-west, north-south, and vertical directions as a target single-component microseismic waveform data as an example, the waveform complexity is first determined: Whether there is a change in sign between adjacent sampling points in the target single-component microseismic waveform data is determined. If there is a change, it indicates that the waveform has crossed the zero axis. The total number of times the waveform crosses the zero axis is determined as the zero-crossing count, and the ratio of the zero-crossing count to the total number of sampling points in the target single-component microseismic waveform data is taken as the zero-crossing rate. Simultaneously, the amplitude variation of the target single-component microseismic waveform data is determined. Then, the zero-crossing rate and the amplitude variation are weighted and summed to obtain a complexity evaluation value. Finally, the waveform complexity is determined based on the complexity evaluation value; a higher complexity evaluation value indicates a more complex waveform. Furthermore, it is necessary to determine the signal-to-noise ratio (SNR) of the target single-component microseismic waveform data: First, determine the signal segment and noise segment of the target single-component microseismic waveform data, and then calculate the root mean square (RMS) of the waveform data within the signal segment and the waveform data within the noise segment respectively. and Then, the signal-to-noise ratio of the target single-component microseismic waveform data is calculated according to the following formula. :

[0044] Furthermore, after determining the waveform complexity and signal-to-noise ratio (SNR) of the target single-component microseismic waveform data, a comprehensive analysis of these factors is needed to determine the weighting coefficients for the target single-component microseismic waveform data. For example, the higher the waveform complexity and the higher the SNR, the larger the weighting coefficient of the target single-component microseismic waveform data; conversely, the lower the waveform complexity and the lower the SNR, the smaller the weighting coefficient. Alternatively, according to actual needs, corresponding weighting coefficients can be set for higher waveform complexity and higher SNR, and a weighted sum of waveform complexity and SNR can be performed based on these weighting coefficients. Finally, the weighting coefficients for the corresponding single-component microseismic waveform data are determined based on the weighted summation result, and then the weighting coefficients for each single-component microseismic waveform data are normalized. In this embodiment of the invention, the waveform complexity and SNR can also be normalized first before calculating the weighting coefficients. This embodiment of the invention determines the weighting coefficients for each single-component microseismic waveform data by comprehensively analyzing waveform complexity and SNR, taking all information into consideration, thereby improving the accuracy of the weighting coefficient calculation.

[0045] Furthermore, the microseismic waveform feature vectors of each single-component microseismic waveform data are determined separately. Then, each microseismic waveform feature vector is processed into a shared feature vector. Next, based on the weighting coefficients of each single-component microseismic waveform data, the microseismic waveform feature vectors of the same latitude are weighted and fused to obtain the microseismic waveform fusion feature. Finally, the microseismic waveform fusion feature is input into a preset phase picking model for phase picking. This embodiment of the invention, by fusing each single-component microseismic waveform data, can extract more implicit information, making fuller use of information and thus improving the accuracy of subsequent phase picking.

[0046] Furthermore, when the microseismic waveform fusion features are input into the preset seismic phase picking model, the encoder can use multiple convolutional layers to perform local convolutional feature extraction on the input microseismic waveform fusion features. Then, the BiLSTM module can simultaneously utilize past and future contextual information to capture the long-range temporal dependencies before and after the arrival of the seismic wave. Through the joint modeling of forward LSTM (capturing past information) and backward LSTM (capturing future information), BiLSTM can more accurately locate the first arrival point of the seismic wave. The GC Block module uses graph convolution operations to exchange information between the three components of the temporal correlation features extracted by BiLSTM, enhancing the correlation between components and improving the model's ability to model multi-component joint features, thereby obtaining global contextual features. Finally, the decoder uses convolutional layers to map the temporal correlation features extracted by the bottleneck layer and the global contextual features into a seismic phase probability distribution, that is, the P-wave and S-wave arrival probability distribution curves for each microseismic monitoring station. By setting a probability threshold, the first arrival time of the P-wave and the first arrival time of the S-wave can be determined.

[0047] 204. Obtain a preset seismic source location model, wherein the preset seismic source location model includes a feature embedding layer for feature embedding, a graph feature extraction layer for graph feature extraction, a global readout layer for extracting global features, and a coordinate regression layer for seismic source location.

[0048] In this embodiment of the invention, to obtain a preset seismic source location model, it is necessary to train and construct the preset seismic source location model in advance. Based on this, the method includes: constructing a preset initial seismic source location model; obtaining a natural seismic source dataset and a coal mine seismic source dataset, wherein both the natural seismic source dataset and the coal mine seismic source dataset include sample event-station maps with annotation information, the annotation information being the seismic source location; dividing the natural seismic source dataset into a natural seismic source training set, a natural seismic source validation set, and a natural seismic source test set; training the preset initial seismic source location model using the natural seismic source training set; validating the trained preset initial seismic source location model using the natural seismic source validation set; testing the preset initial seismic source location model that meets the validation requirements using the natural seismic source test set; and finally, using the preset initial seismic source location model that meets the test conditions as a preset general seismic source location model; fine-tuning the model parameters of the preset general seismic source location model using the coal mine seismic source dataset; and using the preset general seismic source location model with fine-tuned parameters as the preset seismic source location model.

[0049] The natural seismic source dataset includes various types of microseismic data from nature, covering a wide range and containing a large amount of data. The coal mine seismic source dataset, on the other hand, has a relatively small sample size, containing only coal mine microseismic data. Specifically, during model training, a pre-defined initial seismic source location model is first constructed. Then, the natural and coal mine seismic source datasets are acquired, ensuring all necessary files are included. The data is converted to a format understandable by the pre-defined initial seismic source location model. The natural seismic source dataset is then used to train, validate, and test the model. Specifically, the dataset can be partitioned: using randomness or a specific strategy (such as stratified sampling), the natural seismic source dataset is divided into training, validation, and test sets. The model is then trained using the training set, validated using the validation set, and tested using the test set to evaluate its performance on unseen data. Precision, recall, and other metrics on the test set are calculated and recorded. If the model performance does not meet requirements, it can return to the training phase for further iterations or adjustments. This process yields a pre-defined general seismic source location model that meets the requirements. Further, the coal mine seismic source dataset is used to train, validate, and test the general model. Specifically, the dataset can be divided first: using random or specific strategies (such as stratified sampling), the coal mine seismic dataset can be divided into training, validation, and test sets. The training set is then used to train the model to fine-tune its parameters. The validation set is used to validate the general model, and the test set is used to test the validated general model and evaluate its performance on unseen data. Precision, recall, and other metrics on the test set are calculated and recorded. If the performance of the general model does not meet the requirements, it can be returned to the training phase for further iterations or adjustments. This process yields a pre-defined seismic source location model that meets the requirements. This embodiment of the invention pre-trains the model using a natural seismic source dataset to learn general seismic source characteristics, and then fine-tunes it using a coal mine seismic source dataset to adapt to specific noise environments, achieving a high-precision seismic source location model even with limited coal mine seismic source data.

[0050] Furthermore, the trained preset seismic source location model is stored, and when seismic source location is needed, the seismic source location model can be directly obtained based on the stored location.

[0051] 205. Obtain the three-dimensional position coordinates of each microseismic monitoring station, determine the arrival time of the seismic waves at each microseismic monitoring station based on the seismic phase identification results, determine the reference microseismic monitoring station, and determine the relative arrival time difference of the corresponding microseismic monitoring station based on the arrival time of the seismic waves at each microseismic monitoring station and the arrival time of the seismic waves at the reference microseismic monitoring station.

[0052] Specifically, any one of the microseismic monitoring stations can be set as the reference microseismic monitoring station. The three-dimensional spatial coordinates of each microseismic monitoring station are acquired in real time using the Global Positioning System or the BeiDou Navigation Satellite System. Based on the phase identification results (wave arrival time probability distribution curve) output by the preset phase picking model, the time point corresponding to the maximum probability is taken as the wave arrival time of the corresponding station. For each microseismic monitoring station, the difference between its wave arrival time and the reference station's wave arrival time is calculated to obtain the relative arrival time difference.

[0053] 206. In each microseismic monitoring station, determine the K nearest neighbor microseismic monitoring stations corresponding to any given microseismic monitoring station. Based on the three-dimensional position coordinates, determine the Euclidean distance and the absolute value of the time difference between the seismic wave arrival at each microseismic monitoring station and its corresponding K nearest neighbor microseismic monitoring stations.

[0054] Specifically, for each microseismic monitoring station, based on its three-dimensional position coordinates, the Euclidean distance between the station and all other stations is calculated. The stations are sorted from nearest to farthest and the K stations with the smallest distances are selected as the K nearest neighbor microseismic monitoring stations. Then, the Euclidean distance between the microseismic monitoring station and its corresponding K nearest neighbor microseismic monitoring stations is determined. Based on the arrival times of the seismic waves at each station determined in the aforementioned steps, the absolute value of the time difference between the arrival times of the seismic waves at each station and its K nearest neighbor stations is calculated.

[0055] 207. Construct an event-station diagram based on three-dimensional position coordinates, Euclidean distance, relative time difference, and the absolute value of the seismic wave time difference. In this diagram, microseismic monitoring stations are nodes, three-dimensional position coordinates and relative time difference are node features, and Euclidean distance and the absolute value of the seismic wave time difference are edge features connecting a microseismic monitoring station to its K nearest neighbor microseismic monitoring stations.

[0056] Specifically, each station is treated as a node, the three-dimensional position coordinates and relative time difference are used as the node features of the corresponding node, and the Euclidean distance and the absolute value of the seismic wave time difference are used as the edge features of the connection between the corresponding station and its K nearest neighbor microseismic monitoring stations. An event-station graph is constructed based on the nodes, node features, edges between nodes and edge features.

[0057] 208. Input the event-station map into the preset seismic source location model. Extract the hidden state of high-dimensional nodes from the event-station map through the feature embedding layer. Extract deep node features from the hidden state of high-dimensional nodes through the graph feature extraction layer. Extract the global representation vector from the deep node features through the global readout layer. Use the coordinate regression layer to locate the seismic source of the target coal mine through the global representation vector.

[0058] In this embodiment of the invention, after constructing the event-station map, it is necessary to analyze the event-station map using a preset source location model to determine the source location of the target coal mine. Based on this, step 208 specifically includes: performing linear mapping, normalization, and random discarding on the node features in the event-station map sequentially through the feature embedding layer to obtain the high-dimensional hidden state of each node; iteratively updating the high-dimensional hidden state of the node through multi-layer convolutional blocks in the graph feature extraction layer; for each node, fusing the high-dimensional hidden state of the node's neighboring nodes with the edge features corresponding to the edges between the node and its neighboring nodes during the message passing process of each layer convolutional block, and converting the fusion result through a multi-layer perceptron to obtain the deep node features of the node; aggregating the deep node features of each node using a global average pooling function through the global readout layer to obtain a global representation vector; and mapping and regressing the global representation vector through the linear output layer in the coordinate regression layer to obtain the source location result containing the mean of the predicted three-dimensional coordinates of the source.

[0059] Specifically, the constructed event-station graph is input into a pre-defined seismic source localization model. First, a feature embedding layer linearly maps the node features of each node in the graph to increase feature dimensionality. Then, normalization is performed to eliminate the influence of different feature scales. Random discarding is then performed to enhance the model's generalization ability, resulting in the high-dimensional hidden state of each node. Next, these high-dimensional hidden states are fed into multi-layered cascaded convolutional blocks in the graph feature extraction layer for iterative updates. During message passing in each cascaded convolutional block, for each node, the high-dimensional hidden states of all its neighboring nodes are fused with the corresponding edge features between the node and its neighbors. The fusion result is then non-linearly transformed using a multilayer perceptron to obtain the deep node features of the node after this layer's update. This process is iteratively executed layer by layer in multiple cascaded convolutional blocks, gradually integrating node features into a wider range of neighbor information. Finally, a global readout layer uses a global average pooling function to aggregate the deep node features of all nodes in the graph, summarizing the information scattered across each node into a single global representation vector. Finally, the global representation vector is mapped and regressed through the linear output layer in the coordinate regression layer, directly outputting the source location result containing the predicted mean of the three-dimensional coordinates of the source, thus completing the end-to-end location from the event-station map to the spatial location of the microseismic event.

[0060] This invention utilizes multi-layered convolutional blocks to continuously aggregate neighbor node information and fuse edge features during message passing, enabling each station node's features to gradually incorporate the spatial-temporal joint information of its surrounding stations. Global average pooling then aggregates the complementary information of all stations into a unified representation. Finally, linear regression is used to directly output the three-dimensional coordinates of the seismic source, achieving end-to-end accurate mapping from multi-station observation data to the seismic source location, effectively reducing the impact of noise or missing data from a single station on the localization results.

[0061] This invention provides another deep learning-based microseismic source localization method. Compared with the current grid search method for microseismic source localization, this invention simultaneously utilizes waveform information in the east-west, north-south, and vertical directions, capturing seismic wave polarization characteristics more completely than single-component picking, thus improving the accuracy of phase identification. By reading the waveform twice in both the forward and reverse directions using BiLSTM, the discrete convolutional features output by the encoder are concatenated into continuous and complete temporal features, solving the problem that pure convolutional networks can only see local windows and cannot understand the context of the waveform, significantly improving the accuracy of seismic wave arrival time picking. Based on the ordered temporal features, GC... The Block module calculates global attention weights and extracts global contextual features, highlighting effective phase components and suppressing noise. This cascaded design, which first models the temporal sequence and then globally weights the signals, allows the attention mechanism to function based on ordered features, making it more accurate than directly applying attention to the original waveform. By uniformly encoding the distribution of stations in physical space and wave propagation time constraints into a graph structure, dependence on wave velocity models is avoided, thereby improving the accuracy of source location. The graph feature extraction layer in the preset source location model can comprehensively analyze node and edge features, thus improving the accuracy of source location. The above-mentioned source location process of this invention can be completed in one step without global optimization, and avoids getting trapped in local optima, thus improving the efficiency and accuracy of source location.

[0062] Furthermore, as Figure 1 In specific implementation, embodiments of the present invention provide a microseismic source localization device based on deep learning, such as... Figure 3 As shown, the device includes: a waveform acquisition unit 31, a seismic phase model acquisition unit 32, a seismic phase identification unit 33, a positioning model acquisition unit 34, a construction unit 35, and a seismic source positioning unit 36.

[0063] The waveform acquisition unit 31 can be used to acquire three-component microseismic waveform data collected by multiple microseismic monitoring stations for the target coal mine in response to the microseismic source location signal for the target coal mine.

[0064] The seismic phase model acquisition unit 32 can be used to acquire a preset seismic phase picking model, wherein the preset seismic phase picking model includes an encoder for extracting local convolutional features, a bottleneck layer for extracting temporal features and global context information, and a decoder for extracting microseismic phases. The bottleneck layer includes a BiLSTM (Bidirectional Long Short-Term Memory) module and a GC Block (Global Context Block) module.

[0065] The phase identification unit 33 can be used to input the three-component microseismic waveform data of each microseismic monitoring station into the preset phase picking model, extract local convolutional features from the three-component microseismic waveform data through the encoder, extract temporal correlation features from the local convolutional features through the BiLSTM module in the bottleneck layer, extract global context features from the temporal correlation features through the GC Block module in the bottleneck layer, and identify the phases of the temporal correlation features and the global context features through the decoder to obtain the phase identification result corresponding to each microseismic monitoring station.

[0066] The positioning model acquisition unit 34 can be used to acquire a preset seismic source positioning model, wherein the preset seismic source positioning model includes a feature embedding layer for feature embedding, a graph feature extraction layer for graph feature extraction, a global readout layer for extracting global features, and a coordinate regression layer for seismic source positioning.

[0067] The construction unit 35 can be used to obtain the three-dimensional position coordinates of each of the microseismic monitoring stations, determine the arrival time of the seismic waves at each of the microseismic monitoring stations based on the seismic phase identification results, and construct an event-station diagram based on the three-dimensional position coordinates and the absolute value of the seismic wave arrival time difference.

[0068] The source location unit 36 ​​can be used to input the event-station map into the preset source location model, extract the hidden state of high-dimensional nodes from the event-station map through the feature embedding layer, extract deep node features from the hidden state of high-dimensional nodes through the graph feature extraction layer, extract global representation vectors from the deep node features through the global readout layer, and obtain the source location result of the target coal mine by performing source location on the global representation vectors through the coordinate regression layer.

[0069] In specific application scenarios, in order to train and build a preset phase picking model, such as Figure 4 As shown, the device further includes a model building unit 37.

[0070] The model building unit 37 can be used to construct a preset initial phase picking model; acquire a natural phase dataset and a coal mine phase dataset, wherein both the natural phase dataset and the coal mine phase dataset include sample three-component microseismic waveform data with annotation information, the annotation information being a phase probability distribution curve; divide the natural phase dataset into a natural phase training set, a natural phase validation set, and a natural phase test set; train the preset initial phase picking model using the natural phase training set; validate the trained preset initial phase picking model using the natural phase validation set; test the preset initial phase picking model that meets the validation requirements using the natural phase test set; finally, use the preset initial phase picking model that meets the test conditions as the preset general phase picking model; fine-tune the model parameters of the preset general phase picking model using the coal mine phase dataset; and use the fine-tuned preset general phase picking model as the preset phase picking model.

[0071] In specific application scenarios, in order to train and construct a preset source location model, the model construction unit 37 can also be used to construct a preset initial source location model; obtain a natural source dataset and a coal mine source dataset, wherein both the natural source dataset and the coal mine source dataset include sample event-station maps with annotation information, the annotation information being the source location; divide the natural source dataset into a natural source training set, a natural source validation set, and a natural source test set; train the preset initial source location model using the natural source training set; validate the trained preset initial source location model using the natural source validation set; test the preset initial source location model that meets the validation requirements using the natural source test set; finally, use the preset initial source location model that meets the test conditions as the preset general source location model; fine-tune the model parameters of the preset general source location model using the coal mine source dataset; and use the preset general source location model with fine-tuned parameters as the preset source location model.

[0072] In specific application scenarios, in order to construct an event-station diagram, the construction unit 35 includes a determination module 351 and a construction module 352.

[0073] The determining module 351 can be used to determine the reference microseismic monitoring station and determine the relative arrival time difference of the corresponding microseismic monitoring station based on the arrival time of the seismic waves of each microseismic monitoring station and the arrival time of the seismic waves of the reference microseismic monitoring station.

[0074] The determining module 351 can also be used to determine K nearest neighbor microseismic monitoring stations corresponding to any microseismic monitoring station in each microseismic monitoring station, and to determine the Euclidean distance and absolute value of the time difference between each microseismic monitoring station and its corresponding K nearest neighbor microseismic monitoring stations based on the three-dimensional position coordinates.

[0075] The construction module 352 can be used to construct an event-station map based on the three-dimensional position coordinates, the Euclidean distance, the relative time difference, and the absolute value of the seismic wave time difference. The microseismic monitoring station is used as a node in the event-station map, the three-dimensional position coordinates and the relative time difference are used as node features of the corresponding node, and the Euclidean distance and the absolute value of the seismic wave time difference are used as edge features of the edges connecting the microseismic monitoring station to its K nearest neighbor microseismic monitoring stations.

[0076] In specific application scenarios, in order to extract the high-dimensional hidden state of nodes in the event-station graph through the feature embedding layer, the source localization unit 36 ​​can be used to sequentially perform linear mapping, normalization, and random discarding on the node features in the event-station graph through the feature embedding layer to obtain the high-dimensional hidden state of each node; iteratively update the high-dimensional hidden state of nodes through multi-layer cascaded convolutional blocks in the graph feature extraction layer; for each node, during the message passing process of each layer of cascaded convolutional blocks, fuse the high-dimensional hidden state of the node's neighboring nodes with the edge features corresponding to the edges between the node and its neighboring nodes, and convert the fusion result through a multilayer perceptron to obtain the deep node features of the node; aggregate the deep node features of each node through the global readout layer using a global average pooling function to obtain a global representation vector; and map and regress the global representation vector through the linear output layer in the coordinate regression layer to obtain the source localization result containing the mean of the predicted three-dimensional coordinates of the source.

[0077] In specific application scenarios, in order to preprocess the three-component micro-vibration waveform data, the device also includes a waveform processing unit 38.

[0078] The waveform processing unit 38 can be used to determine the covariance matrix that is common to the three-component microseismic waveform data collected by each microseismic monitoring station; perform eigenvalue decomposition on the covariance matrix to obtain multiple eigenvalues ​​and eigenvectors corresponding to each eigenvalue; and determine the principal polarization direction that is common to the three-component microseismic waveform data based on each eigenvector. Each component of the three-component microseismic waveform data is projected onto the main polarization direction to obtain the main component microseismic waveform data. The main component microseismic waveform data is then subjected to forward filtering, signal inversion, reverse filtering, and signal inversion again to obtain the preprocessed main component microseismic waveform data. Based on the main polarization direction, the preprocessed main component microseismic waveform data is decomposed into preprocessed three-component microseismic waveform data.

[0079] In specific application scenarios, in order to perform phase picking, the phase identification unit 33 can be used to input the preprocessed three-component microseismic waveform data corresponding to each of the microseismic monitoring stations into the preset phase picking model.

[0080] In a specific application scenario, in order to perform phase picking, the phase identification unit 33 can be used to determine the waveform complexity and signal-to-noise ratio of each single component microseismic waveform data in the three-component microseismic waveform data; determine the weight coefficient of the corresponding single component microseismic waveform data based on the waveform complexity and the signal-to-noise ratio; perform weighted fusion of the single component microseismic waveform data based on the weight coefficient to obtain microseismic waveform fusion features; and input the microseismic waveform fusion features of each microseismic monitoring station into the preset phase picking model.

[0081] It should be noted that other corresponding descriptions of the functional modules involved in the microseismic source localization device based on deep learning provided in this embodiment of the invention can be found in the following references. Figure 1 The corresponding description of the method shown will not be repeated here.

[0082] Based on the above, Figure 1The method shown, correspondingly, also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, performs the following steps: In response to a microseismic source location signal for a target coal mine, acquiring three-component microseismic waveform data collected by multiple microseismic monitoring stations for the target coal mine; acquiring a preset phase picking model, wherein the preset phase picking model includes an encoder for extracting local convolutional features, a bottleneck layer for extracting temporal features and global context information, and a decoder for extracting microseismic phases, the bottleneck layer including a BiLSTM (Bidirectional Long Short-Term Memory) module and a GC Block (Global Context Block) module; inputting the three-component microseismic waveform data from each of the microseismic monitoring stations into the preset phase picking model, extracting local convolutional features from the three-component microseismic waveform data through the encoder, extracting temporal correlation features from the local convolutional features through the BiLSTM module in the bottleneck layer, and extracting temporal correlation features from the GC Block in the bottleneck layer. The Block module performs global context feature extraction on the temporal correlation features, and uses the decoder to perform seismic phase identification on the temporal correlation features and the global context features to obtain the seismic phase identification result for each microseismic monitoring station; it obtains a preset source location model, wherein the preset source location model includes a feature embedding layer for feature embedding, a graph feature extraction layer for graph feature extraction, a global readout layer for extracting global features, and a coordinate regression layer for source location; it obtains the three-dimensional position coordinates of each microseismic monitoring station, and determines the seismic phase identification result for each station. Based on the arrival times of seismic waves from the aforementioned microseismic monitoring stations, an event-station map is constructed using the three-dimensional location coordinates and the absolute value of the seismic wave arrival time difference. This event-station map is then input into the preset seismic source location model. A feature embedding layer extracts the hidden states of high-dimensional nodes from the event-station map, a graph feature extraction layer extracts deep node features from the hidden states of the high-dimensional nodes, a global readout layer extracts global representation vectors from the deep node features, and a coordinate regression layer uses the global representation vectors to locate the seismic source of the target coal mine.

[0083] Based on the above, Figure 1 The method shown and as Figure 3 The embodiment of the device shown in the invention also provides a physical structure diagram of a computer device, such as... Figure 5As shown, the computer device includes: a processor 41, a memory 42, and a computer program stored in the memory 42 and executable on the processor. Both the memory 42 and the processor 41 are mounted on a bus 43. When the processor 41 executes the program, it performs the following steps: In response to a microseismic source location signal for a target coal mine, it acquires three-component microseismic waveform data collected by multiple microseismic monitoring stations for the target coal mine; it acquires a preset phase picking model, wherein the preset phase picking model includes an encoder for extracting local convolutional features, a bottleneck layer for extracting temporal features and global context information, and a decoder for extracting microseismic phases. The bottleneck layer includes a BiLSTM (Bidirectional Long Short-Term Memory) module and a GC Block (Global Context) module. The module (Global Context Block) is used to input the three-component microseismic waveform data of each microseismic monitoring station into the preset phase picking model. The encoder extracts local convolutional features from the three-component microseismic waveform data, the BiLSTM module in the bottleneck layer extracts temporal correlation features from the local convolutional features, the GC Block module in the bottleneck layer extracts global context features from the temporal correlation features, and the decoder performs phase identification on the temporal correlation features and the global context features to obtain the phase identification result for each microseismic monitoring station. A preset source location model is obtained, which includes a feature embedding layer for feature embedding, a graph feature extraction layer for graph feature extraction, a global readout layer for extracting global features, and a coordinate regression layer for source location. The three-dimensional position coordinates of each microseismic monitoring station are obtained, and the phase identification result is used to determine the position of each station. Based on the arrival times of seismic waves from the aforementioned microseismic monitoring stations, an event-station map is constructed using the three-dimensional location coordinates and the absolute value of the seismic wave arrival time difference. This event-station map is then input into the preset seismic source location model. A feature embedding layer extracts the hidden states of high-dimensional nodes from the event-station map, a graph feature extraction layer extracts deep node features from the hidden states of the high-dimensional nodes, a global readout layer extracts global representation vectors from the deep node features, and a coordinate regression layer uses the global representation vectors to locate the seismic source of the target coal mine.

[0084] Through the technical solution of this invention, by simultaneously utilizing waveform information in three directions—east-west, north-south, and vertical—it captures the polarization characteristics of seismic waves more completely than single-component picking, thus improving the accuracy of seismic phase identification. By using BiLSTM to read waveforms twice in both the forward and reverse directions, the discrete convolutional features output by the encoder are concatenated into continuous and complete temporal features, solving the problem that pure convolutional networks can only view local windows and cannot understand the context of the waveform, significantly improving the accuracy of seismic wave arrival time picking. Based on the ordered temporal features, GC... The Block module calculates global attention weights and extracts global contextual features, highlighting effective phase components and suppressing noise. This cascaded design, which first models the temporal sequence and then globally weights the signals, allows the attention mechanism to function based on ordered features, making it more accurate than directly applying attention to the original waveform. By uniformly encoding the distribution of stations in physical space and wave propagation time constraints into a graph structure, dependence on wave velocity models is avoided, thereby improving the accuracy of source location. The graph feature extraction layer in the preset source location model can comprehensively analyze node and edge features, thus improving the accuracy of source location. The above-mentioned source location process of this invention can be completed in one step without global optimization, and avoids getting trapped in local optima, thus improving the efficiency and accuracy of source location.

[0085] It is obvious to those skilled in the art that the modules or steps of the present invention described above can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. Optionally, they can be implemented using computer-executable program code, thereby storing them in a storage device for execution by a computing device. In some cases, the steps shown or described can be performed in a different order than those presented herein, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Thus, the present invention is not limited to any particular combination of hardware and software.

[0086] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for locating microseismic sources based on deep learning, characterized in that, include: In response to the microseismic source location signal for the target coal mine, three-component microseismic waveform data collected by multiple microseismic monitoring stations for the target coal mine are acquired respectively. A preset seismic phase picking model is obtained, wherein the preset seismic phase picking model includes an encoder for extracting local convolutional features, a bottleneck layer for extracting temporal features and global context information, and a decoder for extracting microseismic phases. The bottleneck layer includes a BiLSTM (Bidirectional Long Short-Term Memory) module and a GC Block (Global Context Block) module. The three-component microseismic waveform data of each microseismic monitoring station are input into the preset seismic phase picking model. The encoder extracts local convolutional features from the three-component microseismic waveform data. The BiLSTM module in the bottleneck layer extracts temporal correlation features from the local convolutional features. The GC Block module in the bottleneck layer extracts global context features from the temporal correlation features. The decoder identifies the seismic phases of the temporal correlation features and the global context features to obtain the seismic phase identification results for each microseismic monitoring station. A preset seismic source location model is obtained, wherein the preset seismic source location model includes a feature embedding layer for feature embedding, a graph feature extraction layer for graph feature extraction, a global readout layer for extracting global features, and a coordinate regression layer for seismic source location. The three-dimensional position coordinates of each of the microseismic monitoring stations are obtained, the arrival time of the seismic waves at each of the microseismic monitoring stations is determined based on the seismic phase identification results, and an event-station diagram is constructed based on the three-dimensional position coordinates and the absolute value of the seismic wave arrival time difference. The event-station map is input into the preset seismic source location model. The hidden state of high-dimensional nodes is extracted from the event-station map through the feature embedding layer. The deep node features of the hidden state of high-dimensional nodes are extracted through the graph feature extraction layer. The global representation vector of the deep node features is extracted through the global readout layer. The seismic source location result of the target coal mine is obtained by using the global representation vector through the coordinate regression layer.

2. The method according to claim 1, characterized in that, Before acquiring the preset seismic phase picking model, the method further includes: Construct a pre-defined initial seismic phase picking model; Obtain a natural seismic phase dataset and a coal mine seismic phase dataset, wherein both the natural seismic phase dataset and the coal mine seismic phase dataset include sample three-component microseismic waveform data with annotation information, the annotation information being the seismic phase probability distribution curve; The natural earthquake phase dataset is divided into a natural earthquake phase training set, a natural earthquake phase validation set, and a natural earthquake phase test set. The preset initial earthquake phase picking model is trained using the natural earthquake phase training set, the trained preset initial earthquake phase picking model is validated using the natural earthquake phase validation set, and the preset initial earthquake phase picking model that meets the validation requirements is tested using the natural earthquake phase test set. Finally, the preset initial earthquake phase picking model that meets the test conditions is taken as the preset general earthquake phase picking model. The model parameters of the preset general seismic phase picking model are fine-tuned using the coal mine seismic phase dataset, and the preset general seismic phase picking model with fine-tuned parameters is used as the preset seismic phase picking model.

3. The method according to claim 1, characterized in that, Before obtaining the preset seismic source location model, the method further includes: Construct a pre-defined initial earthquake source location model; Obtain natural earthquake source datasets and coal mine earthquake source datasets, wherein both the natural earthquake source datasets and the coal mine earthquake source datasets include sample event-station maps with labeled information, the labeled information being the location of the earthquake source; The natural earthquake source dataset is divided into a natural earthquake source training set, a natural earthquake source validation set, and a natural earthquake source test set. The preset initial earthquake source location model is trained using the natural earthquake source training set, validated using the natural earthquake source validation set, and tested using the natural earthquake source test set. Finally, the preset initial earthquake source location model that meets the test conditions is taken as the preset general earthquake source location model. The model parameters of the preset general seismic source location model are fine-tuned using the coal mine seismic source dataset, and the preset general seismic source location model with fine-tuned parameters is used as the preset seismic source location model.

4. The method according to claim 1, characterized in that, An event-station diagram is constructed based on the three-dimensional location coordinates and the absolute value of the seismic wave time difference, including: Determine the reference microseismic monitoring stations, and determine the relative arrival time difference of the corresponding microseismic monitoring stations based on the arrival time of the seismic waves at each of the microseismic monitoring stations and the arrival time of the seismic waves at the reference microseismic monitoring stations; In each microseismic monitoring station, K nearest neighbor microseismic monitoring stations corresponding to any given microseismic monitoring station are identified. Based on the three-dimensional position coordinates, the Euclidean distance and the absolute value of the seismic wave arrival time difference between each microseismic monitoring station and its corresponding K nearest neighbor microseismic monitoring stations are determined. An event-station map is constructed based on the three-dimensional location coordinates, the Euclidean distance, the relative time difference, and the absolute value of the seismic wave time difference. The microseismic monitoring station is used as a node in the event-station map, the three-dimensional location coordinates and the relative time difference are used as node features of the corresponding node, and the Euclidean distance and the absolute value of the seismic wave time difference are used as edge features of the edges connecting the microseismic monitoring station to its K nearest neighbor microseismic monitoring stations.

5. The method according to claim 4, characterized in that, High-dimensional node hidden state extraction is performed on the event-station graph through a feature embedding layer, including: The feature embedding layer sequentially performs linear mapping, normalization, and random discarding on the node features in the event-station graph to obtain the high-dimensional hidden state of each node. The graph feature extraction layer performs deep node feature extraction on the hidden state of the high-dimensional node, including: The hidden state of the high-dimensional node is iteratively updated by multi-layer concatenated convolutional blocks in the graph feature extraction layer. For each node, during the message passing process of each layer concatenated convolutional block, the hidden state of the high-dimensional node of the neighboring node of the node is fused with the edge features corresponding to the edge between the node and the neighboring node. The fusion result is then transformed by a multilayer perceptron to obtain the deep node features of the node. The global readout layer extracts global representation vectors from the deep node features, including: The global readout layer uses a global average pooling function to aggregate the deep node features of each node to obtain a global representation vector. The source location result of the target coal mine is obtained by performing source location on the global representation vector through the coordinate regression layer, including: The global representation vector is mapped and regressed through the linear output layer in the coordinate regression layer to obtain the source location result containing the mean of the three-dimensional coordinate prediction of the source.

6. The method according to claim 1, characterized in that, After acquiring three-component microseismic waveform data collected from multiple microseismic monitoring stations targeting the target coal mine, the method further includes: For the three-component microseismic waveform data collected by each microseismic monitoring station, determine the covariance matrix that is common to the three-component microseismic waveform data. The covariance matrix is ​​decomposed into eigenvalues ​​to obtain multiple eigenvalues ​​and eigenvectors corresponding to each eigenvalue. The main polarization direction corresponding to the three-component microseismic waveform data is determined based on each eigenvector. Each component of the three-component microseismic waveform data is projected onto the main polarization direction to obtain the main component microseismic waveform data. The main component microseismic waveform data is then subjected to forward filtering, signal inversion, reverse filtering, and signal inversion again in sequence to obtain the preprocessed main component microseismic waveform data. Based on the main polarization direction, the preprocessed main component microseismic waveform data is decomposed into preprocessed three component microseismic waveform data. The three-component microseismic waveform data of each of the aforementioned microseismic monitoring stations are input into the preset seismic phase picking model, including: The preprocessed three-component microseismic waveform data corresponding to each of the microseismic monitoring stations are input into the preset seismic phase picking model.

7. The method according to claim 1, characterized in that, The three-component microseismic waveform data of each of the aforementioned microseismic monitoring stations are respectively input into the preset seismic phase picking model, including: For each single-component microseismic waveform data in the three-component microseismic waveform data, the waveform complexity and signal-to-noise ratio of each single-component microseismic waveform data are determined respectively; the weighting coefficient of the corresponding single-component microseismic waveform data is determined based on the waveform complexity and the signal-to-noise ratio; and the single-component microseismic waveform data are weighted and fused based on the weighting coefficient to obtain the microseismic waveform fusion feature. The microseismic waveform fusion features of each of the microseismic monitoring stations are input into the preset seismic phase picking model.

8. A microseismic source localization device based on deep learning, characterized in that, include: The waveform acquisition unit is used to acquire three-component microseismic waveform data collected by multiple microseismic monitoring stations for the target coal mine in response to the microseismic source location signal for the target coal mine. The phase model acquisition unit is used to acquire a preset phase picking model, wherein the preset phase picking model includes an encoder for extracting local convolutional features, a bottleneck layer for extracting temporal features and global context information, and a decoder for extracting microseismic phases. The bottleneck layer includes a BiLSTM (Bidirectional Long Short-Term Memory) module and a GC Block (Global Context Block) module. The phase identification unit is used to input the three-component microseismic waveform data of each microseismic monitoring station into the preset phase picking model, extract local convolutional features from the three-component microseismic waveform data through the encoder, extract temporal correlation features from the local convolutional features through the BiLSTM module in the bottleneck layer, extract global context features from the temporal correlation features through the GC Block module in the bottleneck layer, and identify the phase through the temporal correlation features and the global context features through the decoder to obtain the phase identification result corresponding to each microseismic monitoring station. The positioning model acquisition unit is used to acquire a preset seismic source positioning model, wherein the preset seismic source positioning model includes a feature embedding layer for feature embedding, a graph feature extraction layer for graph feature extraction, a global readout layer for extracting global features, and a coordinate regression layer for seismic source positioning. A construction unit is used to obtain the three-dimensional position coordinates of each of the microseismic monitoring stations, determine the arrival time of the seismic waves at each of the microseismic monitoring stations based on the seismic phase identification results, and construct an event-station diagram based on the three-dimensional position coordinates and the absolute value of the seismic wave arrival time difference. The seismic source location unit is used to input the event-station map into the preset seismic source location model, extract the hidden state of high-dimensional nodes from the event-station map through a feature embedding layer, extract deep node features from the hidden state of high-dimensional nodes through a graph feature extraction layer, extract global representation vectors from the deep node features through a global readout layer, and perform seismic source location on the global representation vectors through a coordinate regression layer to obtain the seismic source location result of the target coal mine.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

10. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images