A rock burst prediction method based on SSA-CNN-MoLSTM-Attention

Abstract

The present invention relates to the field of mine safety technology, and in particular to a rock burst prediction method based on sparrow search algorithm (SSA), convolutional neural network (CNN), deformation long short-term memory network (MoLSTM) and attention mechanism (Attention). The present invention combines wavelet denoising technology and advanced deep learning models, and improves the prediction accuracy and robustness of the model through sparrow search algorithm optimization. The present invention performs well in rock burst prediction, can effectively reduce prediction errors, and provides reliable technical support for rock burst prediction.

Description

Technical Field

[0001] The present invention relates to the field of mine safety technology, and in particular to a rock burst prediction method based on a sparrow search algorithm (SSA), a convolutional neural network (CNN), a deformation long short-term memory network (MoLSTM) and an attention mechanism (Attention). Background Art

[0002] Rock burst is one of the common geological disasters in the mining process. It is sudden, destructive and highly dangerous, posing a serious threat to mining workers and equipment. Existing rock burst prediction methods mostly rely on traditional statistical and empirical analysis methods, which are difficult to fully utilize the complex time series characteristics in microseismic data, and the prediction accuracy is limited. In addition, the prediction of rock burst is affected by a variety of complex factors, and its changes have strong nonlinear and dynamic characteristics. In addition, due to the limitations of the underground environment, the collected data often contain a lot of noise, so previous prediction methods often find it difficult to accurately capture these characteristics, resulting in low accuracy of the prediction results.

[0003] In recent years, with the development of deep learning technology, rock burst prediction methods based on deep learning have gradually become a research hotspot. However, how to effectively combine multiple advanced technologies, such as convolutional neural networks, time series modeling, and intelligent optimization algorithms, to improve the accuracy and robustness of prediction models remains a problem to be solved. Summary of the invention

[0004] The present invention provides a rock burst prediction method based on SSA-CNN-MoLSTM-attention, which combines rock burst prediction methods of multiple advanced technologies and improves the accuracy and reliability of prediction by preprocessing, feature extraction, time series modeling and parameter optimization of microseismic data.

[0005] A rock burst prediction method based on SSA-CNN-MoLSTM-attention includes the following steps:

[0006] S1. Perform wavelet denoising on the original microseismic data and then divide the data into a training set and a test set;

[0007] S2. Use convolutional neural network (CNN) to extract local features, i.e. periodicity, of microseismic data;

[0008] S3. Use Mogrifier LSTM to perform time series modeling on the extracted features;

[0009] S4. Introduce attention mechanism to optimize the model;

[0010] S5. Use sparrow search algorithm (SSA) to optimize model parameters;

[0011] S6. Use the optimized model parameters to train the CNN-MoLSTM-attention model and perform rock burst prediction.

[0012] As a further limitation of this patent, the wavelet denoising processing in S1 specifically refers to: selecting appropriate wavelet basis functions and decomposition layers, performing wavelet decomposition on the original microseismic data, and obtaining detail coefficients and approximation coefficients of each decomposition layer; processing the detail coefficients of each decomposition layer by applying a soft threshold or hard threshold method to remove noise components; and performing wavelet reconstruction on the detail coefficients and approximation coefficients that have been subjected to threshold processing to obtain denoised microseismic data.

[0013] As a further limitation of this patent, the convolutional neural network (CNN) feature extraction in S2 specifically refers to: applying multiple convolution layers to the input denoised data, each convolution layer containing several convolution kernels for extracting local features of the data; adding a pooling layer after the convolution layer to reduce the feature dimension, reduce the amount of calculation, and enhance the robustness of the feature; applying a nonlinear activation function between the convolution layer and the pooling layer to introduce nonlinear features.

[0014] As a further limitation of this patent, the Mogrifier LSTM timing modeling in S3 specifically refers to: performing timing modeling on the input features through the LSTM layer to capture the timing information in the data; applying the Mogrifier mechanism between the input gate and the forget gate of the LSTM unit to dynamically adjust the interaction between the input features and the hidden state.

[0015] As a further limitation of this patent, the attention mechanism in S4 specifically refers to: calculating the attention weight of each time step to determine its weight for the final time series representation; using the attention weight to weight the outputs of CNN and Mogrifier LSTM to obtain a comprehensive time series feature representation.

[0016] As a further limitation of this patent, the sparrow search algorithm (SSA) optimization in S5 specifically refers to: defining a fitness function based on the prediction error of the model on the validation set to evaluate the performance of each individual. Select the individual with the best fitness as the leader, and use its position as a reference to update the positions of other individuals. Perform local searches near the leader to refine the optimal solution. Perform extensive searches globally to avoid local optimal solutions. When an individual falls into a local optimum, apply migration and escape strategies to enhance global search capabilities.

[0017] As a further limitation of this patent, model training and prediction in S6 refers to using optimized model parameters to train the model based on historical microseismic data; using the trained model to predict future microseismic data to provide a prediction result of rock burst pressure.

[0018] Compared with the existing methods, the present invention has the following advantages:

[0019] 1. Traditional statistical methods are highly dependent on the experience and knowledge of experts, while methods based on simple machine learning models often cannot effectively handle complex nonlinear relationships and large-scale data. This method integrates convolutional neural networks (CNN) and deformation long short-term memory networks (Mogrifier LSTM), which can automatically learn complex spatial and temporal features in data and improve the ability to model nonlinear relationships. CNN is used to extract spatial features, while Mogrifier LSTM optimizes predictions through long-term dependencies, allowing the model to more accurately capture the relationship between microseismic energy and rock burst pressure.

[0020] 2. Conventional model parameters usually need to be adjusted manually, which may lead to suboptimal or unstable model behavior in complex systems. This method uses the sparrow search algorithm (SSA) to automatically optimize the hyperparameters of the deep learning model. SSA simulates the behavior of a sparrow colony and iteratively searches for the global optimal solution, thereby effectively improving the prediction accuracy and robustness of the model. This automated optimization not only saves time in adjusting model parameters, but also ensures the best performance of the model under various conditions.

[0021] 3. Due to environmental factors and construction factors, monitoring data often contain a lot of noise, which often leads to instability in prediction results. This method introduces wavelet noise reduction technology, which can effectively separate useful signals from complex monitoring data and remove noise. This technology ensures high-quality data processing in the model during analysis and prediction, greatly improving the reliability and accuracy of the prediction.

[0022] 4. Conventional models may lack the ability to focus on key features when processing large data sets, resulting in inefficient learning or model overfitting. This method introduces an attention mechanism that enables the model to dynamically adjust the degree of attention to different geological features. This mechanism enables the model to more accurately identify and utilize the most informative data, thereby improving the accuracy and stability of predictions.

[0023] 5. Traditional methods cannot effectively process large-scale geological data with complex spatial and temporal dependencies. This method processes spatial data features through CNN, while Mogrifier LSTM optimizes the long-term dependencies of time series data. This combination enables the model to perform well in processing complex data structures, especially when it comes to comprehensive analysis and prediction of multi-dimensional information. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] The present invention is further described in detail below in conjunction with the accompanying drawings:

[0025] Figure 1 It is a schematic diagram of the process of the present invention;

[0026] Figure 2 It is the overall framework diagram of the SSA-CNN-LSTM-Attention model of the present invention;

[0027] Figure 3 is a training flow chart of the model of the present invention;

[0028] Figure 4 Schematic diagram of the structure of the convolutional neural network of the present invention;

[0029] Figure 5 A memory unit structure diagram of the long short-term memory neural network of the present invention;

[0030] Figure 6 Schematic diagram of the structure of the attention mechanism of the present invention. DETAILED DESCRIPTION

[0031] A rock burst prediction method based on SSA-CNN-MoLSTM-attention builds a powerful framework that can capture the intrinsic relationship of complex time series data through the fusion model. Through this combination, the model will perform better when dealing with long sequences, complex patterns, spatiotemporal data, and prediction tasks that require fine tuning of parameters.

[0032] like Figure 1 As shown, the method specifically comprises the following steps:

[0033] S1. Perform wavelet denoising on the original microseismic data, and then divide the data into a training set and a test set; the wavelet denoising is to perform wavelet denoising on the original microseismic data to remove noise and retain effective signals. Specifically, it means: selecting appropriate wavelet basis functions and decomposition layers, performing wavelet decomposition on the original microseismic data, and obtaining detail coefficients and approximation coefficients of each decomposition layer; processing the detail coefficients of each decomposition layer by applying a soft threshold or hard threshold method to remove noise components; performing wavelet reconstruction on the detail coefficients and approximation coefficients after threshold processing to obtain denoised microseismic data.

[0034] Wavelet denoising includes three steps: wavelet decomposition, threshold processing and signal reconstruction. The denoised data is divided into a training set and a test set and normalized. The data is normalized by using the maximum and minimum normalization method so that the training data is between 0 and 1.

[0035] The microseismic data include the occurrence time of the microseismic event, the distance the working face advanced when the microseismic event occurred, and the energy value of the microseismic event.

[0036] S2. Use a convolutional neural network (CNN) to extract the local features, i.e., periodicity, of microseismic data; the convolutional neural network (CNN) feature extraction performs multi-scale feature extraction and dimensionality reduction on the input data through multiple convolutional layers and pooling layers, specifically referring to: applying multiple convolutional layers to the input denoised data, each convolutional layer containing several convolution kernels for extracting local features of the data; adding a pooling layer after the convolutional layer to reduce the feature dimension, reduce the amount of calculation, and enhance the robustness of the feature; applying a nonlinear activation function between the convolutional layer and the pooling layer to introduce nonlinear features.

[0037] S3. Use Mogrifier LSTM to perform time series modeling on the extracted features to capture the long-term and short-term dependencies in the microseismic data; the Mogrifier LSTM time series modeling specifically refers to: performing time series modeling on the input features through the LSTM layer to capture the time series information in the data; applying the Mogrifier mechanism between the input gate and the forget gate of the LSTM unit to dynamically adjust the interaction between the input features and the hidden state.

[0038] S4. Introduce the attention mechanism to optimize the model; the attention mechanism in S4 performs weighted processing on the outputs of CNN and MoLSTM to enhance the weight of important time series information; specifically, by calculating the attention weight of each time step, its weight on the final time series representation is determined; the attention weight is used to weight the outputs of CNN and Mogrifier LSTM to obtain a comprehensive time series feature representation.

[0039] S5. Use the sparrow search algorithm (SSA) to optimize the model parameters; the sparrow search algorithm (SSA) optimization uses the sparrow search algorithm SSA to automatically find the best solution and optimize the model according to the optimal parameters. Specifically, it means: the fitness function is defined according to the prediction error of the model on the validation set, which is used to evaluate the performance of each individual. The individual with the best fitness is selected as the leader, and its position is used as a reference to update the positions of other individuals. Perform local searches near the leader to refine the optimal solution. Perform extensive searches on a global scale to avoid local optimal solutions. When an individual falls into a local optimum, migration and escape strategies are applied to enhance global search capabilities.

[0040] S6. Use the optimized model parameters to train the CNN-MoLSTM-attention model and perform rock burst prediction. The model training and prediction refers to using the optimized model parameters to train the model based on historical microseismic data; using the trained model to predict future microseismic data and provide rock burst prediction results.

[0041] The CNN-MoLSTM-Attention model includes: the first layer uses the CNN model to extract local features of the input data to extract the periodicity of microseismic events; the second layer uses the MoLSTM model to process the nonlinear characteristics of the microseismic data to extract the temporal characteristics of the microseismic events; the third layer uses the attention mechanism to weight the output of the upper layer using the attention weight to obtain a comprehensive temporal feature representation.

[0042] The MoLSTM applies a Mogrifier mechanism between the input gate and the forget gate of the LSTM unit to dynamically adjust the interaction between the input features and the hidden state. The Mogrifier mechanism can improve the model's ability to capture temporal information by performing a weighted summation of the input features and the hidden state.

[0043] The proposed model structure of the present invention is shown in the attached Figure 2 As shown, this method includes the following levels:

[0044] The microseismic monitoring data are obtained in the input layer, including the occurrence time of the microseismic event, the advancement distance of the working face when the microseismic event occurs, and the energy value of the microseismic event;

[0045] In the data preprocessing layer, a series of preprocessing will be performed on the microseismic data. Due to the limitations of the underground environment, the collected data often contains a lot of noise or data is missing; therefore, this layer will fill the missing data and reduce the noise of the original data. Therefore, the appropriate wavelet basis function and decomposition layer number will be selected to perform wavelet decomposition on the original microseismic data. The original signal will be decomposed into n-1 layers of detail coefficients and one layer of approximation coefficients. The soft threshold method or hard threshold method is applied to the decomposed detail coefficients for threshold processing to remove noise. The thresholded detail coefficients are combined with the unprocessed approximation coefficients, and the wavelet reconstruction method is used to restore the denoised microseismic signal.

[0046] After the data is processed, the sparrow search algorithm (SSA) is used to find the optimal parameters of the model. The SSA algorithm iteratively finds the optimal parameter combination by initializing the population, evaluating fitness, updating the positions of the leader and other individuals, applying forager and producer strategies, and migration and escape strategies. The fitness function is defined based on the prediction error of the model on the validation set to evaluate the performance of each individual. The individual with the best fitness is selected as the leader, and its position is used as a reference to update the positions of other individuals. The update strategy of the leader position is adjusted based on the optimization direction of the fitness function. Local search is performed near the leader to refine the optimal solution. The forager strategy searches for the local optimal solution by searching in a small range around the leader. A wide search is performed globally to avoid the local optimal solution. The producer strategy searches for the global optimal solution by randomly searching in the global range. When an individual falls into a local optimum, the migration and escape strategies are applied to enhance the global search capability. The migration strategy improves the efficiency of the global search by changing the position of the individual to escape the local optimal trap.

[0047] After finding the optimal parameters, the neural network model will be established. In the present invention, the convolutional neural network adopts one-dimensional convolution, and multiple convolutional layers are applied to the input denoised data. Each convolutional layer contains several convolution kernels for extracting local features of the data. A pooling layer is added after each convolutional layer to reduce the feature dimension, reduce the amount of calculation, and enhance the robustness of the feature. A nonlinear activation function is applied between the convolutional layer and the pooling layer to introduce nonlinear features.

[0048] After extracting local features, a deformable long short-term memory neural network is established to extract time series features. The input features are modeled through the LSTM layer to capture the time series information in the data. The LSTM unit contains an input gate, a forget gate, and an output gate to control the input, forgetting, and output of information. The Mogrifier mechanism is applied between the input gate and the forget gate of the LSTM unit to dynamically adjust the interaction between the input features and the hidden state.

[0049] The Attention mechanism is added to the CNN and MoLSTM layers to enhance the focus on important information, so as to extract the periodicity and timing of microseismic events and focus on key information, thereby improving the prediction performance and accuracy.

[0050] The final fully connected layer is used to fuse the features of the upper layer to form a global feature representation, which is used to predict continuous values ​​and output the final prediction results.

[0051] Embodiment 1:

[0052] Step 1: Local weighted regression is used to fill the missing values ​​of microseismic data. Local weighted regression is a non-parametric regression method that fits by assigning different weights to each data point. It can estimate missing values ​​based on existing data points without relying on data distribution assumptions. Symlet8 wavelet is used to decompose the data into three layers after interpolation. The data of each layer of decomposition is de-noised using a soft threshold function and then reorganized.

[0053] Step 2: Divide the processed data into training set and test set, and then normalize all the data. The normalization formula is as follows:

[0054]

[0055] Where: x′ is the normalized data; x is the original data before normalization; max(x) and min(x) are the maximum and minimum values ​​of the original data set, respectively.

[0056] Step 3: Use the sparrow search algorithm to find the best parameters. The SSA algorithm divides the population into three categories: producers, followers, and guards. These three types of individuals use different strategies to find the optimal solution in the search space. Producers are responsible for exploring new food sources, that is, potential optimal solutions, followers observe producers and follow producers who find better solutions, and guards move randomly in the population to avoid potential dangers, such as local optimal solutions. The position update formulas of these three types of populations are as follows:

[0057] (1) Producer position update formula

[0058]

[0059] Where: is the j-th dimension position information of the i-th sparrow in the t-th iteration; iter max is the maximum number of iterations; α is a random number in [0, 1]; R2 is a random number in [0, 1], indicating the warning value; ST is a constant in [0.5, 1], indicating the safety value; Q is a random number that obeys the normal distribution; L is a 1×d matrix with all elements being 1.

[0060] (2) Follower position update formula

[0061]

[0062] Where: is the worst individual in the tth iteration; is the position of the current optimal producer; A is a 1×d matrix, and its elements are randomly assigned 1 or -1; A+ is A T (AA T )-1 .

[0063] (3) Sentinel position update formula

[0064]

[0065] Where: is the current global optimal position; β is a random number that obeys a normal distribution with a mean of 0 and a variance of 1; K is a random number in [-1, 1], where the positive and negative numbers represent the moving direction of the sparrow, and the size represents the step length control parameter; f i is the fitness value of the current individual; f g is the current maximum fitness value; f w is the current minimum fitness value.

[0066] Finally, SSA will optimize seven parameters, namely, learning rate, random dropout rate, convolution kernel size of CNN, number of neurons of MoLSTM, model iteration rounds, number of samples per model iteration;

[0067] Step 4: Establish the CNN-MoLSTM-Attention model according to the optimal parameters.

[0068] First: Establish a CNN-MoLSTM network based on the optimal parameters. According to the optimization results of SSA, CNN will use one hidden layer and MoLSTM will use two hidden layers. The optimizer of the entire fusion model uses Adam, the activation function uses ReLU, the learning rate is 0.0008, the random inactivation rate is 0.1599, the convolution kernel size of the CNN layer is 3, the number of neurons in the first layer of MoLSTM is 65, the number of neurons in the second layer of MoLSTM is 31, the iteration rounds of the model are 97, and the number of training samples in each iteration is 54.

[0069] Second: Convolutional neural network includes one-dimensional convolution layer, maximum pooling layer and fully connected layer, as shown in the attached figure. Figure 4 As shown in the figure. The convolution layer uses a set of convolution kernels to slide on the input data to extract local features of the data. The pooling layer is used to reduce the spatial dimension of the features, reduce the amount of calculation, and retain important features. The fully connected layer maps these flattened feature vectors to a higher-dimensional space. The attention mechanism is added after the convolution layer and before the pooling layer to reduce the calculation while enhancing the feature expression ability and improving the generalization ability of the model, thereby achieving the purpose of extracting the periodicity of microseismic events. The structure of the attention mechanism is shown in the figure. Figure 6 shown.

[0070] Third: The long short-term memory neural network includes input gate, forget gate and output gate, as shown in the attached Figure 5 As shown. At time step t, the parameters in the LSTM model are as follows:

[0071] f t =σ(W f *[h t-1 , X t ]+b f )

[0072] i t =σ(W i *[h t-1 , X t ]+b i )

[0073]

[0074] C t =f t *C t-1 +i t *C t

[0075] o t =α(W o *[h t-1 , X t ]+b o )

[0076] h t =o t *tanh(C t )

[0077] Where: the activation function used is the Sigmod function (σ) and the hyperbolic tangent function (tanh); i t , o t 、f t , C t , are the input gate, output gate, forget gate, the content of the memory unit and the content of the new memory unit respectively; W f , W i , W c The weight matrices corresponding to the functions respectively; h t-1 is the input value at time t-1; X t is the input value at time t; b i 、b c、 b o Divided into the bias term of the corresponding gate; h t is the hidden state at the current time step.

[0078] The MoLSTM used in the present invention is an improved long short-term memory network (LSTM), which is designed to enhance the performance of the model when processing time series data. It modifies the input and hidden states alternately to more effectively capture the complex relationship between sequence data. This mechanism enables Mogrifier LSTM to have better information integration capabilities than traditional LSTM models, can increase model complexity without increasing computational burden, and has stronger long-term dependency capture capabilities and stronger generalization capabilities.

[0079] The core of Mogrifier LSTM lies in its unique alternating modification mechanism, which alternately changes the input vector and hidden state at each time step. This process can be described by the following formula:

[0080] Let x (t) is the input vector at time step t, h (t-1) is the hidden state of the previous time step. Mogrifier LSTM updates x through q rounds of alternating modification process (t) and h (t-1) , where q is a pre-set positive integer.

[0081] (1) For odd rounds r = 1, 3, ..., q (if q is an odd number), calculate:

[0082] h (t-1) =2·σ(Q r x (t) )⊙h (t-1)

[0083] Where Q r is the parameter matrix of the rth round, σ represents the sigmoid function, and ⊙ represents the Hadamard product.

[0084] (2) For an even number of rounds r = 2, 4, ..., q (if q is an even number), calculate:

[0085] x (t) =2·σ(R r h (t-1) )⊙x (t)

[0086] Where R r is the parameter matrix of the rth round.

[0087] After completing the above alternating modifications, the updated x (t) and h (t-1) This will be used in the standard LSTM update step to compute the hidden state h at the current time step (t) and the cell state c (t)These update steps follow the standard LSTM formula, except that they use alternating modified x (t) and h (t-1) This mechanism enables Mogrifier LSTM to more effectively integrate input and past hidden state information at each time step, thereby improving the model's ability to process time series data.

[0088] After the MoLSTM layer is established, a layer of attention mechanism is added at its output position, so that it can improve its ability to capture key information, enhance the expression and generalization capabilities of the model, and provide better interpretability, thereby improving performance in various sequence prediction tasks.

[0089] Fourth: Add a random inactivation layer at the output position of CNN-MoLSTM to prevent overfitting of the neural network and improve the generalization ability of the model.

[0090] Fifth: Use the fully connected layer to fuse the features of the upper layer to form a global feature representation for predicting continuous values.

[0091] Step 5: Train the model. The specific training process is as follows. Figure 2 As shown, the model obtained based on the training obtains prediction information based on the test set.

[0092] Step 6: Use mean absolute error (MAE), root mean square error (RMSE) and coefficient of determination (R 2 )Evaluate the error between the model prediction results and the actual results, and further adjust the model. The calculation formula is as follows:

[0093]

[0094]

[0095]

[0096] Where n is the number of data points, y i is the actual value of the ith data point, is the predicted value of the ith data point, It is the average of all actual values. The smaller the MAE and RMSE, the smaller the model prediction error. The closer R2 is to 1, the smaller the difference between the model's predicted value and the true value, and the better the model fit.

[0097] In summary, the present invention can effectively predict rock burst pressure through the SSA-CNN-MoLSTM-Attention method.

Claims

1. A rock burst prediction method based on SSA-CNN-MoLSTM-attention, characterized in that: The following steps are involved: S1. Perform wavelet denoising on the original microseismic data and then divide the data into a training set and a test set; S2. Use convolutional neural network to extract local features of microseismic data, namely periodicity; S3. Use Mogrifier LSTM to perform time series modeling on the extracted features; S4. Introduce attention mechanism to optimize the model; S5. Use sparrow search algorithm to optimize model parameters; S6. Use the optimized model parameters to train the CNN-MoLSTM-attention model and perform rock burst prediction; The attention mechanism in S4 mentioned above specifically refers to: calculating the attention weight of each time step to determine its weight for the final time series representation; using the attention weight to weight the outputs of CNN and Mogrifier LSTM to obtain a comprehensive time series feature representation.

2. The rock burst prediction method based on SSA-CNN-MoLSTM-attention according to claim 1, characterized in that: The wavelet denoising process in S1 specifically refers to: selecting appropriate wavelet basis functions and decomposition layers, performing wavelet decomposition on the original microseismic data, and obtaining detail coefficients and approximation coefficients of each decomposition layer; processing the detail coefficients of each decomposition layer by applying a soft threshold or hard threshold method to remove noise components; and performing wavelet reconstruction on the detail coefficients and approximation coefficients after threshold processing to obtain the denoised microseismic data.

3. The rock burst prediction method based on SSA-CNN-MoLSTM-attention according to claim 1, characterized in that: The convolutional neural network feature extraction in S2 specifically refers to: applying multiple convolutional layers to the input denoised data, each convolutional layer contains several convolution kernels, which are used to extract local features of the data; adding a pooling layer after the convolutional layer to reduce the feature dimension, reduce the amount of calculation, and enhance the robustness of the feature; applying a nonlinear activation function between the convolutional layer and the pooling layer to introduce nonlinear features.

4. The rock burst prediction method based on SSA-CNN-MoLSTM-attention according to claim 1, characterized in that: Mogrifier LSTM time series modeling in S3 specifically refers to: performing time series modeling on the input features through the LSTM layer to capture the time series information in the data; applying the Mogrifier mechanism between the input gate and the forget gate of the LSTM unit to dynamically adjust the interaction between the input features and the hidden state.

5. The rock burst prediction method based on SSA-CNN-MoLSTM-attention according to claim 1, characterized in that: The optimization of the sparrow search algorithm in S5 specifically refers to: defining a fitness function based on the prediction error of the model on the validation set to evaluate the performance of each individual; selecting the individual with the best fitness as the leader, and using its position as a reference to update the positions of other individuals; performing local searches near the leader to refine the optimal solution; performing extensive searches globally to avoid local optimal solutions; and applying migration and escape strategies when an individual falls into a local optimum to enhance global search capabilities.

6. The rock burst prediction method based on SSA-CNN-MoLSTM-attention according to claim 1, characterized in that: Model training and prediction in S6 refers to using optimized model parameters to train the model based on historical microseismic data; using the trained model to predict future microseismic data and provide the prediction results of rock burst pressure.

Images