Coal mine earthquake time series feature prediction method based on multi-source data spatio-temporal graph convolution network

Abstract

The coal mine earthquake time series feature prediction method based on multi-source data spatio-temporal graph convolution network belongs to the field of mine dynamic disaster prevention. First, by fusing microseismic data and charge data, the features of the two data sources are extracted and integrated using feature fusion technology to more comprehensively capture the dynamic characteristics of the underground environment. Then, a graph structure is constructed, taking sensors as nodes and using the Euclidean distance between sensors to define the edges between nodes to reflect the spatial correlation of geological activities. The spatio-temporal graph convolution network dynamically captures the correlation in the time and space dimensions. Next, WOA is used to optimize the hyperparameters in the model to the optimal parameter combination, improving the performance and accuracy of the model. Finally, the spatio-temporal attention mechanism is introduced, and by calculating the time and space attention matrices, the data input is dynamically adjusted to better capture the spatio-temporal features and improve the accuracy of the model prediction.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of mine dynamic disaster prevention, and particularly relates to a mine earthquake time series feature prediction, and in particular relates to a coal mine mine earthquake time series feature prediction method based on a multi-source data space-time graph convolution network. BACKGROUND

[0002] Mine earthquake is a coal and rock dynamic phenomenon occurring in mining, and is one of the mine natural disasters. How to reduce and reduce the accidents and disasters caused by mine earthquake is an important research topic at present, and one of the main ways to solve this problem is to monitor mine earthquake in real time and continuously, and to predict the time series features of mine earthquake monitoring data. The rapid development and wide application of machine learning method provide a new idea for making up for the shortage of numerical simulation method. As a data-driven method, machine learning method can effectively capture the nonlinear relationship between parameters in high-dimensional data, and has been widely used in geotechnical engineering.

[0003] In geological monitoring, different monitoring data sources usually provide multi-layer information about underground structure and geological features. Microseismic, ground sound, stress, charge, drill cuttings and other monitoring data reflect the multi-aspect characteristics of the underground environment, forming a multi-dimensional data system. In order to obtain more comprehensive and accurate underground environment change information, and thus improve the prediction accuracy, it is necessary to make full use of these multi-source data for research. Mine earthquake prediction is equivalent to a space-time data prediction problem. Data sensors record at fixed time points and fixed distribution sites. The observation results of adjacent nodes and time stamps are not independent, but are dynamically related to each other, so effective extraction of the space-time correlation of data is the key to solving these problems. Although deep learning method shows potential in mine earthquake prediction, existing models mostly rely on single data source, ignoring the potential association and complementarity between different data sources. In addition, the model lacks interpretability and fails to fully consider the complex physical properties of underground engineering. The current theoretical model has the limitation of relying on a single index, and it is difficult to integrate laboratory and field monitoring states to realize accurate prediction of mine earthquake occurrence time. How to effectively extract the internal mode of nonlinear and complex space-time data and make accurate prediction is still a major challenge in current research. SUMMARY

[0004] To solve the problem of existing mine earthquake time series feature prediction, the present application designs and realizes a regression prediction method based on whale optimization algorithm (WOA) optimization convolution neural network (CNN) and long short-term memory network (LSTM) combined model, and verifies the reliability of the method through example application.

[0005] The technical scheme adopted by the present application is as follows:

[0006] The coal mine earthquake time series feature prediction method based on multi-source data spatio-temporal graph convolution network has the following steps:

[0007] Step 1: WM-STGCN network model construction: The model is divided into three parts: feature fusion, spatio-temporal graph convolution, and whale optimization algorithm hyperparameter optimization WOA.

[0008] In the feature fusion module, feature encoders are used to extract data features from the original microseismic data and charge data. Single-element feature extraction is performed to obtain global time series features for each data type. Relationship learning is performed on the data features to obtain local interaction features between microseismic and charge. The global time series features of each data type and the local interaction features between microseismic and charge are fused to obtain fused features, realizing data fusion of microseismic and charge.

[0009] The spatio-temporal graph convolution module calculates the time and space attention matrices to dynamically capture the correlation in the two dimensions. The data adjusted by the attention mechanism is input into the spatio-temporal convolution layer to obtain dependency relationships from different dimensions.

[0010] The WOA hyperparameter optimization module optimizes the hyperparameters required in the model to find the optimal parameter combination.

[0011] The overall structure and usage of the WM-STGCN model are as follows:

[0012] First, the time series data is cut into three time dimension sequence segments along the time axis according to the length of hours, days, and weeks, which are used as the input of the hour component, day component, and week component, respectively.

[0013] Second, the features of microseismic data and charge data are extracted from sensor data by feature encoders. Global time series features within a single field and local interaction features between multiple fields are learned by intra-field data relationship learning modules and inter-field data relationship learning modules, respectively. Finally, the fusion of intra-field features and multi-field interaction features is realized by a multi-feature fusion module.

[0014] Then, the model is composed of multiple spatio-temporal blocks, each of which has a spatio-temporal attention module and a spatio-temporal convolution module. The residual learning framework is used in each component.

[0015] The hyperparameter optimization part uses the whale optimization algorithm to optimize the learning rate, spatio-temporal convolution kernel number, model block number, and convolution order hyperparameters for data fusion and spatio-temporal graph convolution, and finds the best parameter combination.

[0016] Finally, the outputs of the hour, day, and week components are further combined according to the parameter matrix to obtain the final prediction result.

[0017] Step 2: Data processing:

[0018] Set the maximum number of sensor sampling times as q times per day, the current time is 0, and the predicted window size is T p , three sequence fragments with lengths of T h , T d and T w are intercepted along the time axis, respectively as the input of the hour, day and week three periodic components, wherein T h , T d and T w are multiples of T p .

[0019] In step 2, the specific method is:

[0020] Set the maximum number of sensor sampling times as q times per day, the current time is 0, and the predicted window size is T p . Three sequence fragments with lengths of T h , T d and T w are intercepted along the time axis, respectively as the input of the hour, day and week three periodic components, wherein T h , T d and T w are multiples of T p .

[0021] 2.1) Hourly periodic sequence

[0022]

[0023] Wherein: T h is the sequence fragment of the hour period, t0 represents the current time, N and F are the dimension sizes, representing the number of samples and the number of features respectively;

[0024] As shown in formula (1), the hourly periodic sequence is a historical time sequence directly adjacent to the prediction period. For mine earthquake data, the occurrence and end of mine earthquake events are gradual, so the mine earthquake data just produced will have an impact on future data information;

[0025] 2.2) Daily periodic sequence

[0026]

[0027] Wherein: T d is the sequence fragment of the day period, q is the maximum number of sensor sampling times per day, T p is the predicted window size, t0, N and F are the same as above;

[0028] As shown in formula (2), the daily periodic sequence is composed of past data of several days in the same time period as the prediction period. Considering the daily production of the mine area, mine earthquake data will change periodically every day;

[0029] 2.3) Weekly cycle sequence

[0030]

[0031] where: T w is the sequence segment of the day cycle, q, T p , t0, N, F are the same as above;

[0032] As shown in formula (3), the weekly cycle sequence is composed of the segments of the last few weeks, and the coal mine working mode on Monday is similar to the historical coal mine working mode on Monday, but is quite different from the coal mine working mode on weekends.

[0033] Step 3 Feature fusion: multi-feature fusion of microseismic data and charge data, processing data of different modalities in the same framework.

[0034] 3.1) Feature extraction

[0035] The data feature extraction method based on deep learning is adopted, combining the bidirectional long short-term memory network and the average pooling method to capture the timing information and balance the output at different times. The BiLSTM network is applied to capture the timing relationship in the sensor data, and the microseismic and charge data are processed by independent networks. The BiLSTM network has forward and backward hidden states. For microseismic data, after step 2 data processing, X h , X d , X w are obtained, where micor t (i) represents the input data, and then X

[0036]

[0037] where micor t (i) represents the input of microseismic data, and are the forward and backward hidden states respectively, is the hidden state at the corresponding time; in order to balance the output at different times, the features X at each time t are averaged, as shown in formula (7):

[0038]

[0039] where T represents the total number of timing data, is the feature representation after average pooling;

[0040] The above processing is performed on the charge data to complete feature extraction;

[0041] 3.2) Learning Relationships from Single Data Sources

[0042] By constructing a fully connected graph attention network to extract internal global temporal features for microseismic and charge data respectively, the features of microseismic and charge data are divided into their respective fully connected graphs. Using the self-attention mechanism of the Transformer network, each data feature is divided into several nodes, ensuring that each node has an adjacency relationship with other nodes in the field. Multi-head neighbor aggregation is performed on each node and all adjacent nodes of their respective fully connected graphs to obtain the aggregated features, as shown in equations (8), (9), and (10).

[0043]

[0044] Where Q represents the edge weight, which is the self-attention coefficient value calculated by the self-attention mechanism in the Transformer network; X μ Indicates features before aggregation; The aggregated features are represented as i = 1, ..., h, where h represents the number of self-attention heads, μ ∈ {m, c}, where m and c represent microseismic events and charges, respectively; d k express or In the dimension of , softmax represents the activation function. All represent weight matrices, and || represents cascade operations; since the fully connected graph attention network is a multi-layer graph learning network, it needs to perform multi-layer graph learning to obtain the global temporal features of the final single-factor data;

[0045] 3.3) Global Data Relationship Learning

[0046] To capture the complex relationship between microseisms and charges, a sparse graph attention network is constructed to learn the data relationship between the data features of microseisms and charges. Using the sparse graph attention network, the data features of microseisms and charges are divided into several nodes. For each node, the adjacency relationship with nodes in another data type is calculated. X attention coefficient values ​​are retained for each node from largest to smallest, and the remaining attention coefficient values ​​are set to zero, forming a sparse graph within the field. Multi-head neighbor aggregation is performed on each node and all its adjacent nodes in each fully connected graph to obtain the aggregated features, as shown in equations (11), (12), and (13).

[0047]

[0048] Where, μ m ,μ c These represent micro-vibrations and electrical charges, respectively. The feature after aggregation of the nodes in the microseismic charge that exist adjacent relationship with the node; and The feature before aggregation, The weight matrix is represented, and the local interaction feature between data is obtained by multi-layer graph learning;

[0049] The global time sequence feature of single factor data and the local interaction feature between data are fused to obtain the fusion feature, realize the microseismic charge data fusion, and finally obtain the fusion feature X, as shown in formula (14):

[0050]

[0051] Step 4 hyperparameter optimization

[0052] The model constructed in step 1 needs to select an optimization algorithm, and the whale optimization algorithm is selected as the tool for hyperparameter optimization. Through the intelligent search mechanism in the parameter space, the optimal combination of hyperparameters is found.

[0053] In step 4, the specific method is:

[0054] The initial data is used as the input of the model in step 1, and the specific formula of the algorithm is as follows:

[0055] X i (t+1)=X i (t)+A·rand()·dist(X i (t),

[0056] X best (t))-B·dist(X i (t),X rand (t)) (15)

[0057] Where X i (t) represents the position of the i-th whale at time t, A and B are control parameters of the algorithm, rand() is a random number generation function, dist(·,·) represents a distance function between two points, X best (t) and X rand (t) are the global optimal position and random position at the current time, respectively.

[0058] Step 5 spatiotemporal graph convolution

[0059] After obtaining the optimal hyperparameter combination, the spatial and temporal attention mechanisms are introduced, and then the fusion feature value X obtained in step 4 is calculated by spatiotemporal graph convolution.

[0060] 5.1) Spatial attention mechanism

[0061] In the spatial dimension, the influence between different channels is dynamic. By calculating the spatial attention matrix, the attention mechanism is used to dynamically capture this correlation, as shown in equations (16) and (17):

[0062]

[0063] wherein, is the input of the rth spatio-temporal block, C (r-1) is the number of channels of the input data, b s , s ∈R NxN , is a learnable parameter, and σ is an activation function;

[0064] 5.2) Temporal Attention Mechanism

[0065] In the temporal dimension, there is also dynamic correlation between data in different time slices of the same channel. The attention mechanism is used to dynamically capture this correlation, as shown in equations (18) and (19):

[0066]

[0067] wherein, U1∈R N , is a learnable parameter, and σ is an activation function;

[0068] 5.3) Spatial Dependence Modeling

[0069] Microseismic data and charge data are essentially graph structures. The features of each node are considered as signals on the graph. Graph convolution based on graph theory is used to process each time slice. The spectral method converts the graph into an algebraic form to analyze the topological properties of the graph. Since the convolution operation of the graph signal is equal to the product of these signals transformed into the spectral domain through the graph Fourier transform, when the size of the graph is large, the Chebnet model is used, as shown in equation (20). After adding the spatial attention mechanism, as shown in equation (21):

[0070]

[0071] 5.4) Temporal Dependence Modeling

[0072] After the graph convolution operation captures the adjacent information of each node on the graph in the spatial dimension, a standard convolution layer in the temporal dimension is further stacked to update the node signal by merging the information in adjacent time slices. After capturing the neighborhood information in the spatial dimension, a standard convolution layer is added in the temporal dimension to update the node information by merging the field time slice information, as shown in equation (22):

[0073]

[0074] Step 6 model training and prediction

[0075] The model is trained by the training set, and the model parameters are adjusted to minimize the prediction error;

[0076] The input is the training set and the defined graph node relationship matrix, first, initialize the learning parameters of the model, including defining the graph convolution layer, attention mechanism, full connection layer; then, select batch data for training through iteration, including microseismic data, charge data; in the iteration process, multi-feature fusion is carried out on the data, WOA algorithm is used for hyperparameter optimization, and spatial-temporal attention mechanism is used to obtain spatial-temporal dependence, and finally the prediction data sequence is output through the full connection layer

[0077] The beneficial effects of the present application are: the present application fuses microseismic data and charge data, uses feature fusion technology to extract and integrate the features of the two data sources, so as to more comprehensively capture the dynamic characteristics of the underground environment. Then, a graph structure is constructed, the sensors are used as nodes, and the Euclidean distance between the sensors is used to define the edges between the nodes, so as to reflect the spatial correlation of geological activities. The spatial-temporal graph convolution network is used to dynamically capture the correlation in the time and space dimensions. Then, the whale optimization algorithm WOA is used to optimize the hyperparameters in the model to find the optimal parameter combination and improve the performance and accuracy of the model. Finally, the spatial-temporal attention mechanism is introduced, the time and space attention matrices are calculated, and the data input is dynamically adjusted to better capture the spatial-temporal features and improve the accuracy of the model prediction. Through experiments, it is verified that the model has high precision and good applicability in mine earthquake prediction. BRIEF DESCRIPTION OF DRAWINGS

[0078] Figure 1 : The experimental system diagram of the present application;

[0079] Figure 2 : Comparison of the predicted value and the true value of the present application against the line chart;

[0080] Figure 3 : Comparison of model training time overhead;

[0081] Figure 4 : Comparison of the traffic flow prediction results of the present application on PEMS04 against the line chart;

[0082] Figure 5 : 3 -1 105 working face plane position map;

[0083] Figure 6 : Mine earthquake event warning diagram of the present application;

[0084] Figure 7 : Step 6 training process diagram of the present application. Detailed Implementation

[0085] The technical solutions in the embodiments of the present invention will be described in detail below with reference to the accompanying drawings. The following embodiments are applicable to the present invention, but are not intended to limit the scope of the invention.

[0086] A method for predicting the time series features of coal mine seismic events based on multi-source data spatiotemporal graph convolutional networks, the steps of which are as follows:

[0087] Step 1: Construction of the WM-STGCN Network Model

[0088] The proposed WM-STGCN framework can be divided into three main parts: feature fusion, spatiotemporal graph convolution, and Whale Optimization Algorithm (WOA) hyperparameter optimization. In the feature fusion module, a feature encoder extracts data features from the original microseismic and charge data. Single-element feature extraction is performed to obtain global temporal features for each data type. Relationship learning is then performed on these features to obtain local interaction features between microseismic and charge data. Finally, the global temporal features of each data type are fused with these local interaction features to obtain the fused features, achieving data fusion between microseismic and charge data. The spatiotemporal graph convolution module dynamically captures the correlation between the two dimensions by calculating temporal and spatial attention matrices. Data adjusted by the attention mechanism is input into the spatiotemporal convolutional layer to obtain dependencies from different dimensions. The WOA hyperparameter optimization module optimizes the hyperparameters required by the model, finding the optimal parameter combination to improve the model's performance and accuracy.

[0089] The overall structure and usage of the WM-STGCN model can be summarized as follows:

[0090] First, the time series data is cut into three time-dimensional sequence segments along the time axis according to the length of hour, day, and week, which are used as inputs for the hour component, day component, and week component, respectively.

[0091] Secondly, features of microseismic data and charge data are extracted from sensor data using a feature encoder. Intra-field data relationship learning modules and inter-field data relationship learning modules are used to learn global temporal features within a single field and local interaction features between multiple fields, respectively. Finally, a multi-feature fusion module is used to fuse intra-field features and multi-field interaction features.

[0092] The model is then composed of multiple spatiotemporal blocks. Each spatiotemporal block contains a spatiotemporal attention module and a spatiotemporal convolution module. To optimize training efficiency, a residual learning framework is employed in each component.

[0093] The hyperparameter optimization part optimizes the model hyperparameters required by the data fusion and spatio-temporal graph convolution part, finds the best parameter combination, and further improves the model performance. Specifically, the whale optimization algorithm is used to optimize the learning rate, the number of spatio-temporal convolution kernels, the number of model blocks, the convolution order and other hyperparameters.

[0094] Finally, the outputs of the three components are further combined according to the parameter matrix to obtain the final prediction result.

[0095] Step 2 data processing

[0096] Set the maximum sampling number of the sensor to q times per day, the current time to 0, and the prediction window size to T p . Take three sequence fragments with lengths of T h , T d and T w along the time axis as inputs of the three periodic components of time, day and week, respectively, where T h , T d and T w are multiples of T p .

[0097] (1) Time periodic sequence

[0098]

[0099] As shown in equation (1), the time periodic sequence is a historical time sequence directly adjacent to the prediction period. For mine earthquake data, due to its strong autocorrelation, the occurrence and end of mine earthquake events are gradual. Therefore, the mine earthquake data just produced will inevitably affect the future data information.

[0100] (2) Day periodic sequence

[0101]

[0102] As shown in equation (2), the day periodic sequence is composed of data of the past few days in the same time period as the prediction period. Considering the daily production of the mine, mine earthquake data may change periodically every day, such as regular start of work, end of work, rest, etc.

[0103] (3) Week periodic sequence

[0104]

[0105] As shown in equation (3), the week periodic sequence is composed of recent weeks. Generally speaking, the coal mine work mode on Monday has certain similarity with the historical coal mine work mode on Monday, but may be quite different from the coal mine work mode on weekends.

[0106] Step 3 feature fusion

[0107] Microseismic and charge data provide complementary information in the field of geology, but due to the heterogeneity of their respective characteristics, how to effectively integrate them becomes a key problem. Through multi-feature fusion, different modal data can be processed in the same framework, improving the overall coherence and comprehensiveness of information.

[0108] (1) Feature extraction

[0109] In this paper, a deep learning-based data feature extraction method is adopted to extract meaningful data features of microseismic and charge from raw sensor data. This method combines the techniques of bidirectional long short-term memory network (BiLSTM) and average pooling to capture temporal information and balance the outputs at different times. In order to effectively capture the temporal relationship in sensor data, BiLSTM networks are applied respectively, where microseismic and charge data are processed independently. BiLSTM network has forward and backward hidden states. For microseismic data, after step 2 data processing, X h , X d , X w , micor t (i) are obtained as input data. Then, at each sensor location i, the representation is obtained as shown in equations (4), (5), and (6).

[0110]

[0111] where micor t (i) represents the input of microseismic data, and are the forward and backward hidden states, is the hidden state at the corresponding time. In order to balance the outputs at different times, average pooling needs to be performed on the features at each time t, as shown in equation (7).

[0112]

[0113] where T represents the total number of time series data, is the feature representation after average pooling. Similar processing is performed for charge data to complete feature extraction.

[0114] (2) Single data relationship learning

[0115] To delve deeper into the internal correlations of microseismic and charge data features, this paper employs internal data relationship learning, constructing fully connected graph attention networks to extract internal global temporal features for both microseismic and charge data. The microseismic and charge data features are divided into their respective fully connected graphs. The self-attention mechanism of the Transformer network is used to divide each data feature into several nodes, ensuring that each node has an adjacency relationship with other nodes in the field. Multi-head neighbor aggregation is performed on each node and all adjacent nodes in each fully connected graph to obtain the aggregated features, as shown in equations (8), (9), and (10).

[0116]

[0117]

[0118] Where Q represents the edge weight, which is the self-attention coefficient value calculated by the self-attention mechanism in the Transformer network; X μ Indicates features before aggregation; The aggregated features are represented as i = 1, ..., h, where h represents the number of self-attention heads, μ ∈ {m, c}, where m and c represent microseismic events and charges, respectively; d k express or In the dimension of , softmax represents the activation function. Both represent weight matrices, and || represents cascading operations. Since fully connected graph attention networks are multi-layer graph learning networks, multiple layers of graph learning are required to obtain the final global temporal features of the single-factor data.

[0119] (3) Global data relationship learning

[0120] To capture the complex relationship between microseisms and charges, a sparse graph attention network is constructed to learn the data relationships between the data features of microseisms and charges. Using the sparse graph attention network, the data features of microseisms and charges are divided into several nodes. For each node, its adjacency relationship with nodes in another data type is calculated. Specifically, each node retains X attention coefficient values ​​from largest to smallest, and sets the remaining attention coefficient values ​​to zero, forming a sparse graph within the field. Multi-head neighbor aggregation is performed on each node and all its adjacent nodes in each fully connected graph to obtain the aggregated features, as shown in equations (11), (12), and (13).

[0121]

[0122] Where, μ m ,μ c These represent micro-vibrations and electrical charges, respectively. The node in the microseismic charge represents the characteristics of the node after the aggregation of the adjacent relationship existing in the node in the microseismic charge. and The node in the microseismic charge represents the characteristics of the node after the aggregation of the adjacent relationship existing in the node in the microseismic charge. The node in the microseismic charge represents the characteristics of the node after the aggregation of the adjacent relationship existing in the node in the microseismic charge.

[0123] The node in the microseismic charge represents the characteristics of the node after the aggregation of the adjacent relationship existing in the node in the microseismic charge.

[0124]

[0125] Step 4 hyperparameter optimization

[0126] The performance of a deep learning model is often affected by hyperparameters, and the selection of hyperparameters is crucial to the training effect of the model. Therefore, hyperparameter optimization has become one of the key paths to optimize deep learning models. This process aims to adjust key hyperparameters such as learning rate, number of convolution kernels, and model block number, so that the model can better adapt to data, improve training effect and generalization ability. A reasonable combination of hyperparameters can accelerate the convergence speed of the model, improve the generalization ability of the model, and reduce the risk of overfitting or underfitting. If a fixed combination of hyperparameters is used without optimization, the model may fall into a local optimal solution during training, and cannot fully exploit the potential of the data. This may lead to performance degradation, insufficient generalization ability, and even poor performance on the test set. The absence of hyperparameter optimization may prevent the model from fully utilizing its advantages, so the optimization of hyperparameters is particularly important in deep learning tasks.

[0127] To address the problems caused by hyperparameters, the model constructed in step 1 needs to select an optimization algorithm. The whale optimization algorithm is chosen as the tool for hyperparameter optimization. The whale optimization algorithm simulates the foraging behavior of a whale population and has the ability to search globally, enabling it to efficiently find the optimal solution in a multi-dimensional parameter space. This algorithm draws on the collective behavior of whale populations in nature, evolving in parameter space through intelligent search mechanisms to find the optimal combination of hyperparameters. The core idea of the whale optimization algorithm is to simulate the foraging process of a whale population, constantly adjusting parameters to approach the optimal solution. Its mathematical expression includes an intelligent search mechanism for parameter space, which achieves efficient hyperparameter optimization through randomness and tracking of the global optimal solution. The specific formula of this algorithm is as follows, taking the initial data as the input of the model in step 1:

[0128] X i (t+1)=X i (t)+A·rand()·dist(X i (t),

[0129] X best (t))-B·dist(X i (t),X rand (t)) (15)

[0130] where X i (t) denotes the position of the i-th whale at time t, A and B are control parameters of the algorithm, is a random number generation function, dist(·,·) denotes the distance function between two points, X best (t) and X rand (t) are the global optimal position and random position at the current time, respectively.

[0131] The reason for choosing the whale optimization algorithm is its global search advantage. Compared with other optimization algorithms, the whale optimization algorithm has better global search characteristics, which enables it to efficiently find the optimal solution in a large parameter space. The global search property of the whale optimization algorithm enables it to find a superior parameter combination in complex deep learning tasks, improving the performance of the model. This global search capability plays a key role in the diversity and uncertainty of the hyperparameter space, helping the model better adapt to different data distributions and task scenarios.

[0132] Step 5 spatio-temporal graph convolution

[0133] After obtaining the optimal hyperparameter combination, a spatial and temporal attention mechanism is introduced, and then combined with the step 4 hyperparameters, the fusion feature value X obtained in step 2 is calculated by spatio-temporal graph convolution.

[0134] (1) Spatial attention mechanism

[0135] In the spatial dimension, the influence between different channels is dynamic, so an attention mechanism is used to dynamically capture this correlation, which is achieved by calculating a spatial attention matrix, as shown in equations (16) and (17).

[0136]

[0137] where, is the input of the r-th spatio-temporal block, C (r-1) is the number of channels of the input data, b s , V s ∈R NxN , is a learnable parameter, and σ is an activation function.

[0138] (2) Temporal attention mechanism

[0139] In the time dimension, the data in the same channel at different time slices also has dynamic correlation. The scheme uses an attention mechanism to dynamically capture this correlation, as shown in equations (18) and (19).

[0140]

[0141] wherein, U1∈R N , is a learnable parameter, and σ is an activation function.

[0142] (3) Spatial dependency modeling

[0143] Microseismic data and charge data are essentially graph structures, and the features of each node can be regarded as signals on the graph. Therefore, in order to fully utilize such topological characteristics, a graph convolution based on graph theory is used to process each time slice, and the signal correlation in the spatial dimension of the mine seismic data is utilized. The spectral method converts the graph into an algebraic form to analyze the topological properties of the graph. Since the convolution operation of the graph signal is equal to the product of these signals transformed into the spectral domain by the graph Fourier transform, however, when the size of the graph is large, the consumption of directly performing eigenvalue decomposition on the Laplacian matrix is very high, therefore, the Chebnet model is used to solve this problem, as shown in equation (20). After adding the spatial attention mechanism, as shown in equation (21).

[0144]

[0145] (4) Temporal dependency modeling

[0146] After the graph convolution operation captures the adjacent information of each node on the graph in the spatial dimension, a standard convolution layer in the time dimension is further stacked to update the node signal by merging the information in adjacent time slices. After capturing the neighborhood information in the spatial dimension, a standard convolution layer is added in the time dimension to update the node information by merging the field time slice information, as shown in equation (22).

[0147]

[0148] Step 6: Model training and prediction process

[0149] The model is trained through the training set, and the model parameters are continuously adjusted to minimize the prediction error. The input is the training set and the defined graph node relationship matrix. In the algorithm, the learning parameters of the model are first initialized, including defining the graph convolution layer, the attention mechanism, etc. Then, batch data is selected for training through iteration, including microseismic data and charge data. During the iteration process, multi-feature fusion is performed on the data, the WOA algorithm is used for hyperparameter optimization, and the spatiotemporal dependency is obtained through the spatiotemporal attention mechanism. Finally, the prediction data sequence is output through the fully connected layer The training loss is calculated according to the objective function, and the model parameters are updated by back propagation. The training algorithm running time is related to the iteration period, training data size and model parameters, and the overall running complexity is O(n 2 ), and the process of model training is shown in Figure 7 .

[0150] Example 1

[0151] In order to verify the present application, experiments were carried out, and a mine seismic data set HYSK collected by a certain mine in Liaoning from April 2022 to July 2022 was selected, which contains three-component waveform microseismic data of 100 mine seismic events of 9 microseismic stations and charge data of 6 charge probes. The microseismic data acquisition frequency is 5KHz, and the charge data is 1KHz. In order to optimize the data quality, zero padding, long-short time window method to align P wave, downsampling and Z-Score standardization and other processing steps were adopted to ensure the integrity and consistency of the data. Finally, the data set was divided into training set and test set according to the ratio of 7:3, input into the training model, and the model framework is shown in Figure 1 .

[0152] After the data is input into the model, feature fusion is first performed, the microseismic and charge data features are extracted by deep learning method, and the internal and global relationships between the data are learned by using full-connection graph attention network and sparse graph attention network to realize feature fusion. Further, whale optimization algorithm is used for hyperparameter optimization to find the optimal parameter combination. The spatial and temporal attention mechanism and graph convolution operation are used to capture the spatio-temporal dependence of the data, and finally the prediction result is output, as shown in Figure 2 .

[0153] The mean squared error (MSE) is selected as the model objective function, and the back propagation algorithm is used for training, and the WM-TGCN is designed as the input sequence length to realize multi-step prediction. Adam Optimezer optimizer is used to reduce the error. Mean absolute error (MAE), mean absolute percentage error (MAPE) and root mean square error (RMSE) are used to measure the performance of different models.

[0154] The model training time overhead is evaluated to confirm the strength and weakness of the actual application of the model. As Figure 3As shown in the comparison of model time overhead, WM-STGCN may exhibit slightly higher time cost compared to other models. However, it is worth noting that although the time overhead of WM-STGCN increases, the growth of its training time does not reach an exponential level, but rather a moderate increase within an acceptable range. This feature makes WM-STGCN still maintain good feasibility and practicality in actual application.

[0155] Finally, a generalization experiment is conducted to verify the effectiveness of the model in other fields. The public dataset PEMS04 in the transportation field is selected, which records the multidimensional data measured by sensors on California highways from January 1, 2018 to February 28, 2018, including total flow, average speed, and average occupancy. The control model uses the STGCN model. In WM-STGCN, the total flow is predicted by combining the average speed data of this type, and the experimental results are as follows Figure 4 As shown, compared with STGCN, WM-STGCN achieves the best learning effect. MAE and RMSE achieve the optimal evaluation results in prediction. As the number of prediction steps increases, the error of the WM-STGCN model does not explode, which indicates that it has strong robustness and proves that the proposed model has strong generalization ability.

[0156] An example of the present application is given below in conjunction with the accompanying drawings:

[0157] (1) In order to verify the effectiveness of the WM-STGCN model in actual mine earthquake prediction, it is applied to the prediction and early warning of mine earthquakes in a certain coal mine 3-1105 working face. The 3-1105 working face is located in the middle of the 3-11 mining area, southwest of the mine industrial square, with a terrain slope from northeast to southwest, with the highest point at the middle of the 18-12 and 19-10 boreholes near the ground, with an altitude of 1400.8m, and the lowest point at the ground corresponding to the main retreat passage of the working face, with an altitude of 1348.5m, with a maximum difference of 52.3m in terrain location elevation. The southeast direction of the 3-1105 working face is the auxiliary transportation crossheading (not excavated) of the 3-1107 working face, the northwest direction is the 3-1103 goaf (along the length of 966m) separated by a 6m coal pillar, part of which is solid coal, and the other surrounding directions are solid coal. The plan position of the 3-1105 working face is shown in Figure 5 .

[0158] (2) The monitoring data of the working face in June 2022 is selected for time series feature prediction, and the maximum energy value of the microseism on the same day is divided into 1, 2, 3, and 4 grade intervals, which correspond to four impact levels respectively.

[0159] (3) Finally,Figure 6 The early warning chart shows the comparison between the actual and predicted hazard levels of the 3-1105 working face during June 2022. Through detailed analysis, it can be seen that the hazard level calculated based on actual data and the hazard level calculated based on predicted data show a high degree of consistency. Especially in the case of high predicted hazard level, the prediction accuracy of actual mine earthquake events is as high as 80%. At the same time, on June 8, although the actual hazard level was rated as medium, the data result of the prediction model was high hazard level, and a mine earthquake event did indeed occur subsequently. This example not only verifies the accuracy of the WM-STGCN model proposed in this paper in predicting mine earthquake events, but also demonstrates its effectiveness and reliability in dealing with potential risks.

Claims

1. A method for predicting coal mine seismic time series features based on multi-source data spatiotemporal graph convolutional networks, characterized by: Step 1: WM-STGCN Network Model Construction: The model is generally divided into three parts: feature fusion, spatiotemporal graph convolution, and Whale Optimization Algorithm Hyperparameter Optimization (WOA). In the feature fusion module, data features are extracted from the original microseismic data and charge data through the feature encoder. Single-element feature extraction is performed to obtain the global temporal features of each data type. The relationship between the two data features is learned to obtain the local interaction features between microseismic data and charge data. The global temporal features of each data type are fused with the local interaction features between microseismic data and charge data to obtain the fused features, thereby realizing the data fusion of microseismic data and charge data. The spatiotemporal graph convolution module dynamically captures the correlation between the two dimensions by calculating the temporal and spatial attention matrices. The data adjusted by the attention mechanism is then input into the spatiotemporal convolutional layer to obtain the dependency relationship from different dimensions. The WOA hyperparameter optimization module optimizes the hyperparameters required in the model to find the optimal parameter combination. Step 2: Data Processing The maximum number of samples taken by the sensor is set to q times per day, the current time is 0, and the prediction window size is T. p A section of length T is taken along the time axis. h T d and T w Three sequence segments are used as inputs for the three periodic components of time, day, and week, respectively, where T h T d and T w All are T p Multiples of; Step 3 Feature Fusion: Multi-feature fusion is performed on microseismic data and charge data to process data of different modes within the same framework; Step 4: Hyperparameter Optimization For the model constructed in step 1, an optimization algorithm needs to be selected. The whale optimization algorithm is selected as the hyperparameter optimization tool. It evolves in the parameter space through an intelligent search mechanism to find the optimal combination of hyperparameters. Step 5: Spatiotemporal graph convolution After obtaining the optimal hyperparameter combination, spatial and temporal attention mechanisms are introduced, and then spatiotemporal graph convolution is performed on the fused feature value X obtained in step 2 in combination with the hyperparameters in step 4. Step 6: Model Training and Prediction The model is trained using a training set, and the model parameters are adjusted to minimize the prediction error. The input consists of the training set and a defined graph node relationship matrix. First, the learning parameters of the model are initialized, including defining the graph convolutional layer, attention mechanism, and fully connected layer. Then, batches of data, including microseismic data and charge data, are selected iteratively for training. During the iteration process, multi-feature fusion is performed on the data, hyperparameter optimization is carried out using the WOA algorithm, and spatiotemporal dependence is obtained through a spatiotemporal attention mechanism. Finally, the predicted data sequence is output through a fully connected layer.

2. The method for predicting coal mine seismic time series features based on multi-source data spatiotemporal graph convolutional networks according to claim 1, characterized in that: The overall structure and usage of the WM-STGCN model are as follows: First, the time series data is cut into three time-dimensional sequence segments along the time axis according to the length of hour, day, and week, which are used as inputs for the hour component, day component, and week component, respectively. Secondly, features of microseismic data and charge data are extracted from sensor data through feature encoders. The global temporal features within a single field and the local interaction features between multiple fields are learned through intra-field data relationship learning modules and inter-field data relationship learning modules, respectively. Finally, the fusion of intra-field features and multi-field interaction features is achieved through a multi-feature fusion module. The model is then composed of multiple spatiotemporal blocks, each containing a spatiotemporal attention module and a spatiotemporal convolution module, with a residual learning framework used in each component. In the hyperparameter optimization section, the whale optimization algorithm is used to optimize the hyperparameters such as learning rate, number of spatiotemporal convolution kernels, number of model blocks, and convolution order for the data fusion and spatiotemporal graph convolution parts, and find the best parameter combination. Finally, the outputs of the hour, day, and week components are further merged based on the parameter matrix to obtain the final prediction result.

3. The method for predicting coal mine seismic time series features based on multi-source data spatiotemporal graph convolutional networks according to claim 1, characterized in that: In step 2, the specific method is as follows: The maximum number of samples taken by the sensor is set to q times per day, the current time is 0, and the prediction window size is T. p A section with a length of T is taken along the time axis. h T d and T w Three sequence segments are used as inputs for the three periodic components of time, day, and week, respectively, where T h T d and T w All are T p Multiples of; 2.1) Time-periodic sequences Wherein: T h For a time-period sequence segment, t0 represents the current time, and N and F are the dimensions, representing the number of samples and the number of features, respectively; As shown in equation (1), the time period sequence is a historical time series that is directly adjacent to the prediction period. For seismic data, the occurrence and end of seismic events are gradual. Therefore, the newly generated seismic data will affect future data information. 2.2) Daily Cycle Sequence Wherein: T d The sequence segment has a daily cycle, where q is the maximum number of samples taken by the sensor per day, and T is the number of samples taken per day. p The prediction window size, t0, N, and F are the same as above; As shown in equation (2), the daily cycle sequence is composed of data from the past few days within the same time period as the prediction cycle. Considering the daily production in the mining area, the mine earthquake data will change periodically on a daily basis. 2.3) Weekly Periodic Sequence Wherein: T w For daily periodic sequence segments, q, T p , t0, N, F are the same as above; As shown in equation (3), the weekly cycle sequence is composed of segments from the most recent weeks. The working pattern of coal mines on Mondays is similar to the historical working pattern of coal mines on Mondays, but it is quite different from the working pattern of coal mines on weekends.

4. The method for predicting coal mine seismic time series features based on multi-source data spatiotemporal graph convolutional networks according to claim 1, characterized in that: The specific method in step 3 is as follows: 3.1) Feature Extraction A deep learning-based data feature extraction method is employed, combining a bidirectional long short-term memory network and average pooling to capture temporal information and balance outputs at different times. A BiLSTM network is applied to capture the temporal relationships in the sensor data. Microseismic and charge data are processed by independent networks. The BiLSTM network has hidden states in both forward and backward directions. For microseismic data, after data processing in step 2, X is obtained. h X d X w , here use micor t (i) represents the input data, which is then obtained at each sensor position i, as shown in equations (4), (5), and (6): Among them, micor t (i) represents the input of microseismic data. and These are the forward and backward hidden states, respectively. This refers to the hidden state at the corresponding time step; to balance the output at different times, the features at each time step t are... Perform average pooling as shown in equation (7): Where T represents the total number of time points in the time series data. It is the feature representation after average pooling; The above processing is performed on the charge data to complete feature extraction; 3.2) Learning Relationships from Single Data Sources By constructing a fully connected graph attention network to extract internal global temporal features for microseismic and charge data respectively, the features of microseismic and charge data are divided into their respective fully connected graphs. Using the self-attention mechanism of the Transformer network, each data feature is divided into several nodes, ensuring that each node has an adjacency relationship with other nodes in the field. Multi-head neighbor aggregation is performed on each node and all adjacent nodes of their respective fully connected graphs to obtain the aggregated features, as shown in equations (8), (9), and (10). Where Q represents the edge weight, which is the self-attention coefficient value calculated by the self-attention mechanism in the Transformer network; X μ Indicates features before aggregation; The aggregated features are represented as i = 1, ..., h, where h represents the number of self-attention heads, μ ∈ {m, c}, where m and c represent microseismic events and charges, respectively; d k express or In the dimension of , softmax represents the activation function. All represent weight matrices, and || represents cascade operations; since the fully connected graph attention network is a multi-layer graph learning network, it needs to perform multi-layer graph learning to obtain the global temporal features of the final single-factor data; 3.3) Global Data Relationship Learning To capture the complex relationship between microseisms and charges, a sparse graph attention network is constructed to learn the data relationship between the data features of microseisms and charges. Using the sparse graph attention network, the data features of microseisms and charges are divided into several nodes. For each node, the adjacency relationship with nodes in another data type is calculated. X attention coefficient values ​​are retained for each node from largest to smallest, and the remaining attention coefficient values ​​are set to zero, forming a sparse graph within the field. Multi-head neighbor aggregation is performed on each node and all its adjacent nodes in each fully connected graph to obtain the aggregated features, as shown in equations (11), (12), and (13). Where, μ m ,μ c These represent micro-vibrations and electrical charges, respectively. This represents the characteristics of nodes in a microseismic event after aggregating the nodes with which they have adjacent charges. and Indicates the features before aggregation. The weight matrix is ​​used to perform multi-layer graph learning to obtain the final local interaction features between data. The global temporal features of single-factor data are fused with the local interaction features between data to obtain fused features, thereby realizing the fusion of microseismic charge data. The final fused feature X is shown in Equation (14):

5. The method for predicting coal mine seismic time series features based on multi-source data spatiotemporal graph convolutional networks according to claim 1, characterized in that: In step 4, the specific method is as follows: Using the initial data as input to the model in step 1, the specific formula of the algorithm is as follows: X i (t+1)=X i (t)+A·rand()·dist(X i (t), X best (t))-B·dist(X i (t),X rand (t)) (15) Among them, X i (t) represents the position of the i-th whale at time t, A and B are the control parameters of the algorithm, is a random number generation function, dist(·,·) represents the distance function between two points, X best (t) and X rand (t) represents the global optimal position and the random position at the current time, respectively.

6. The method for predicting coal mine seismic time series features based on multi-source data spatiotemporal graph convolutional networks according to claim 1, characterized in that: In step 5, the specific details are as follows: 5.1) Spatial attention mechanism In the spatial dimension, the influence between different channels is dynamic. By calculating the spatial attention matrix, this correlation is dynamically captured using an attention mechanism, as shown in equations (16) and (17): in, It is the input of the r-th spacetime block, C (r-1) It is the number of input data channels, b s V s ∈R NxN , σ is a learnable parameter, and σ is the activation function; 5.2) Time Attention Mechanism In the time dimension, data from different time slices of the same channel also exhibit dynamic correlations. This correlation can be dynamically captured using an attention mechanism, as shown in equations (18) and (19): in, U1∈R N , σ is a learnable parameter, and σ is the activation function; 5.3) Spatial Dependency Modeling Microseismic data and charge data are essentially graph structures. The features of each node are regarded as signals on the graph. Graph convolution based on graph spectral theory is used to process each time slice. By utilizing the signal correlation of seismic data in the spatial dimension, the spectral method transforms the graph into an algebraic form and analyzes the topological properties of the graph. Since the convolution operation of graph signals is equal to the product of these signals transformed to the spectral domain through graph Fourier transform, when the graph scale is large, the Chebnet model is used, as shown in Equation (20). After adding the spatial attention mechanism, it is shown in Equation (21). 5.4) Time-dependent modeling After capturing the neighbor information of each node in the graph in the spatial dimension, the graph convolution operation further stacks a standard convolutional layer in the temporal dimension to update the node signal by merging the information in adjacent time slices. After capturing the neighborhood information in the spatial dimension, a standard convolutional layer is added in the temporal dimension to update the node information by merging the neighborhood time slice information, as shown in Equation (22):

Images