Method and system for dynamically predicting surrounding rock stress of deep large-mining-height working face

Abstract

The invention discloses a deep large-mining-height working face surrounding rock stress dynamic prediction method. The method comprises the following steps: S1, data acquisition and selection; s2, decomposing a stress signal; s3, trend term stress prediction; s4, periodic term stress prediction; and S5, predicting result fusion. The prediction system comprises the following modules: a data acquisition module, a data preprocessing module, a stress signal decomposition module, a trend term stress prediction module, a periodic term stress prediction module and a prediction result fusion module. Through the advanced CEEMDAN-Kmeans algorithm and multi-algorithm coupling model, the high-precision dynamic prediction of the surrounding rock stress is realized, the signal processing efficiency and accuracy and the prediction precision are improved, the generalization ability of the model to a complex stress change mode is also enhanced, and the prediction precision of the surrounding rock stress is improved. And powerful technical support is provided for coal mine safety production and roadway surrounding rock stress management.

Description

Technical Field

[0001] The present invention relates to a dynamic prediction method, and in particular to a method and system for dynamically predicting the surrounding rock stress of a deep large mining height working face. Background Art

[0002] Nowadays, with the depletion of high-quality coal resources, traditional mining methods lead to resource waste. Although the gob-side entry driving and gob-side entry retaining technologies have improved the recovery rate, the problem of roadway deformation has become increasingly serious, and frequent mine pressure phenomena such as surrounding rock stress accumulation, roof subsidence, and floor heave occur. In the mining of extra-thick coal seams, the deformation of the surrounding rock is particularly prominent, and the complex stress characteristics are prone to disasters such as damage to the support structure, roof fall, and gas outburst. The study of surrounding rock stress is the key to solving the mine pressure problem. Especially around the coal mining face, the strong mining-induced stress causes strong disturbances to the roadway and the working face, resulting in deformation and damage. The original rock stress field, mining-induced stress field, and support stress field in the coal mine underground jointly constitute the underground comprehensive stress field, and its research is crucial for the safe and efficient mining of coal mines. With the increase of mining depth, safety accidents caused by the increase of surrounding rock stress occur frequently. Studying the problem of surrounding rock stress prediction is of great significance for formulating support strategies and ensuring coal mine safety.

[0003] The existing intelligent diagnosis technologies have obvious deficiencies in the prediction of surrounding rock stress. First of all, the research mainly focuses on short-term time series prediction, while long-term prediction is more important for coal mine safety production. However, traditional algorithms such as LSTM and BP have large errors in long-term prediction. Secondly, the existing algorithms fail to fully consider various factors such as geological conditions and mining activities, and it is difficult to accurately capture the composite characteristics of surrounding rock stress data, resulting in limited prediction accuracy. In addition, the existing models consume a large amount of computing resources, have a long training time, insufficient generalization ability, and a high demand for manual screening of training data, and it is difficult to adapt to different scenarios. Therefore, developing a simple, efficient, and adaptable prediction model is an urgent problem to be solved at present. Summary of the Invention

[0004] In order to solve the deficiencies of the above technologies, the present invention provides a method and system for dynamically predicting the surrounding rock stress of a deep large mining height working face.

[0005] In order to solve the above technical problems, the technical solution adopted by the present invention is: a method for dynamically predicting the surrounding rock stress of a deep large mining height working face, characterized by including the following prediction steps:

[0006] Step 1: Data collection and selection: Select long-term monitoring data that can reflect the dynamic changes of the surrounding rock stress for collection;

[0007] Step 2: Stress signal decomposition: The surrounding rock stress data is decomposed using the adaptive noise complete ensemble empirical mode CEEMDAN to obtain several intrinsic mode functions (IMFs). The IMFs after CEEMDAN decomposition are then divided into trend terms, periodic terms, and error terms using the K-means clustering algorithm.

[0008] Step 3: Trend item prediction: The trend item stress value is predicted using the convolutional neural network-support vector machine (CNN-SVW) model to capture the long-term nonlinear change trend of the surrounding rock stress.

[0009] Step 4: Periodic item prediction: A sparrow search algorithm (SSA)-optimized convolutional neural network-bidirectional long short-term memory network hybrid model, namely, an SSA-CNN-BiLSTM hybrid model, is constructed for periodic item stress values to predict the periodic changes in surrounding rock stress.

[0010] Step 5: Prediction value superposition: Superimpose the trend term and periodic prediction results, and consider the influence of the error term to obtain the final surrounding rock stress prediction value.

[0011] As a further implementation of this plan:

[0012] In step 2, the specific process of CEEMDAN decomposition is:

[0013] Step 2.1: Add Gaussian white noise with a homogeneity of 0 to the original signal x(t) to construct a preprocessing sequence x for a total of K experiments. i (t)(i=1,2,~,K):

[0014] x i (t=x(t)+εδ i (t)

[0015] Where ε is the added Gaussian white noise weight coefficient, δ i (t) is the Gaussian white noise during the i-th processing;

[0016] Step 2.2: Preprocess the sequence x i (t) Perform EMD decomposition to obtain i components And find its mean as the first IMF component I1(t) obtained by CEEMDAN decomposition:

[0017]

[0018] r1(t)=x(t)-I1(t)

[0019] Where I1(t) is the first IMF component obtained by CEEMDAN decomposition; is x i(t) The i-th IMF component obtained after EMD decomposition; r1(t) is the residual component obtained after the first decomposition;

[0020] 2.3) Add Gaussian white noise to the residual signal at the j-th stage after decomposition, and continue with EMD decomposition:

[0021]

[0022] r j (t) = r j-1 (t) - I j (t)

[0023] In the formula, I j (t) is the j-th IMF component obtained by CEEMDAN decomposition; E j-1 is the (j - 1)-th IMF component after EMD decomposition of the sequence; ε j-1 is the noise coefficient added to the (j - 1)-th stage residual component by CEEMDAN decomposition; r j (t) is the residual component at the j-th stage;

[0024] 2.4) Repeat the above steps until the number of extreme points of the residual component decreases to a preset value, then the CEEMDAN decomposition ends; at this time, the original signal is decomposed into several IMF components and a residual component.

[0025] As a further implementation of this solution:

[0026] The preset value of the number of extreme points of the residual component is set to be less than or equal to 2.

[0027] As a further implementation of this solution:

[0028] Through K-means clustering, the IMFs after CEEMDAN decomposition are divided into a trend term, a periodic term, and an error term, which respectively reflect the long-term evolution law, periodic fluctuations, and random noise of the surrounding rock stress;

[0029] The optimization objective of K-means clustering is to minimize the within-cluster sum of squared errors SSE:

[0030]

[0031] In the formula, C j is the j-th cluster; μ j is the center or centroid of the j-th cluster; x is the data point.

[0032] As a further implementation of this solution:

[0033] In step 3, the stress value data of the trend item is used as the input of the convolutional neural network CNN. Through multiple convolutional layers and pooling layers, the low-level and high-level features of the stress value data of the trend item are gradually extracted;

[0034] The convolutional layer uses multiple convolutional kernels to perform local perception on the input surrounding rock stress data, and captures low-level features through weighted summation and non-linear activation functions. As the network level increases, the convolutional layer gradually captures high-level features;

[0035] The pooling layer downsamples the output of the convolutional layer;

[0036] Through multiple convolution and pooling operations, the convolutional neural network CNN captures the complex non-linear patterns in the stress value data of the trend item;

[0037] After the feature extraction by CNN, the obtained feature vectors are passed to the support vector machine SVM model for regression prediction.

[0038] As a further implementation of this solution:

[0039] In step 3, the support vector machine SVM model uses the feature vectors extracted by CNN, learns the non-linear change law of the stress data of the trend item based on the kernel function, and outputs the final prediction result.

[0040] As a further implementation of this solution:

[0041] The SAA-CNN-BiLSTM hybrid model combines the sparrow search algorithm SSA, the optimized convolutional neural network CNN, and the bidirectional long short-term memory network BiLSTM;

[0042] This hybrid model receives data through the input layer, extracts features using the CNN layer, then captures the forward and backward dependencies of the time series by the bidirectional LSTM layer, and finally gives the prediction result through the output layer.

[0043] As a further implementation of this solution:

[0044] During the training process of the hybrid model, the Adam optimizer is used to adjust the weights, and the sparrow search algorithm SSA is used to optimize the hyperparameters, including the convolutional kernel size and the number of LSTM neurons, to improve the model performance.

[0045] As a further implementation of this solution:

[0046] The sparrow search algorithm SSA simulates the foraging behavior of sparrows to optimize the hyperparameters. The specific process is as follows:

[0047] First, initialize the sparrow population, set the population size and the maximum number of iterations. At this time, generate multiple random solutions, and each solution represents a different combination of hyperparameters;

[0048] Then, evaluate the fitness of each solution, that is, calculate the prediction accuracy of the corresponding CNN-BiLSTM model on the training set;

[0049] Next, update the position of the solution according to the rules of sparrow foraging behavior, combining local and global search strategies;

[0050] When the fitness reaches the target or the number of iterations exceeds the limit, stop the search and output the optimal hyperparameter combination.

[0051] A dynamic prediction system for surrounding rock stress in deep large mining height working face, comprising:

[0052] A data acquisition module for obtaining monitoring data reflecting the dynamic changes of stress in real time.

[0053] A stress signal decomposition module for decomposing the surrounding rock stress signal by using a model combining complete ensemble empirical mode decomposition with adaptive noise CEEMDAN and K-means clustering algorithm, and dividing it into a trend term, a periodic term and an error term;

[0054] A trend term displacement prediction module for combining a convolutional neural network CNN and a support vector machine SVM to construct a prediction model for surrounding rock stress of the trend term for long-term evolution law prediction;

[0055] A periodic term displacement prediction module for constructing a hybrid model of a convolutional neural network - bidirectional long short-term memory network optimized by SSA, that is, an SSA-CNN-BiLSTM hybrid model to predict periodic fluctuations,

[0056] A prediction result fusion module for superimposing the prediction results of the trend term and the periodic term to obtain the final predicted value of the surrounding rock stress.

[0057] The present invention discloses a method and system for dynamically predicting the surrounding rock stress in a deep large mining height working face. Through the advanced CEEMDAN-Kmeans algorithm and multi-algorithm coupling model, high-precision dynamic prediction of the surrounding rock stress is realized, which not only improves the efficiency and accuracy of signal processing and prediction accuracy, but also enhances the generalization ability of the model to complex stress change patterns, providing strong technical support for coal mine safety production and roadway surrounding rock stress management. Description of the Drawings

[0058] Figure 1 It is a flow chart of the technical route of the present invention.

[0059] Figure 2 It is a structural diagram of the BiLSTM model of the present invention.

[0060] Figure 3 It is a structural diagram of the SSA-CNN-BiLSTM hybrid model of the present invention.

[0061] Figure 4 Curves of each mode obtained by decomposing the original signal using the CEEMAN algorithm in the embodiment

[0062] Figure 5 K-means clustering result graph in the embodiment

[0063] Figure 6 Prediction and error graph of surrounding rock stress for the trend term in the embodiment

[0064] Figure 7 Comparison graph of predicted value and true value for the periodic term in the embodiment

[0065] Figure 8 Error histograms of the SSA-CNN-BiLSTM hybrid model before and after optimization in the embodiment

[0066] Figure 9 Regression graph of the SSA-CNN-BiLSTM hybrid model after optimization in the embodiment

[0067] Figure 10 Comparison graph of superposition prediction results in the embodiment Detailed implementation manners

[0068] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific implementation manners

[0069] The present invention discloses a method for dynamically predicting the surrounding rock stress of a deep large mining height working face, which is based on the CEEMDAN-Kmeans decomposition and the SSA-CNN-BiLSTM hybrid model to achieve accurate prediction of the surrounding rock stress, providing a scientific decision-making basis for the management of the surrounding rock stress of coal mine roadways

[0070] The technical solution adopted by the present invention has a technical route set as Figure 1 shown, including the following prediction steps

[0071] Step 1: Data collection and selection: In a deep large mining height working face, long-term monitoring data that can reflect the dynamic changes of the surrounding rock stress should be selected

[0072] Step 2: Stress signal decomposition: Combining the complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) and the K-means clustering algorithm, the stress value is decomposed into a trend term, a periodic term, and an error term. Specifically, after performing the complete ensemble empirical mode CEEMDAN decomposition on the surrounding rock stress data, several intrinsic mode functions IMF are obtained; then, through the K-means clustering algorithm, the IMF after CEEMDAN decomposition is divided into a trend term, a periodic term, and an error term

[0073] Step 3: Trend term prediction: The stress value of the trend term is predicted using a Convolutional Neural Network-Support Vector Machine (CNN-SVW) model. This model combines the feature extraction ability of CNN and the non-linear regression advantage of SVW, and can effectively capture the long-term non-linear change trend of surrounding rock stress.

[0074] Step 4: Periodic term prediction: A sparrow search algorithm (SSA) optimized Convolutional Neural Network-Bidirectional Long Short-Term Memory Network hybrid model, namely the SSA-CNN-BiLSTM hybrid model, is constructed for predicting the stress value of the periodic term. This model can handle complex non-linear relationships and accurately predict the periodic changes of surrounding rock stress.

[0075] Step 5: Prediction value superposition: The predicted results of the trend term and the periodic term are superimposed to obtain the final predicted value of surrounding rock stress. This process realizes the dynamic prediction of surrounding rock stress.

[0076] In Step 1, data collection and selection are the primary links. It is necessary to clarify the monitoring locations, usually select the key areas of deep large mining height working faces, and install equipment such as high-precision bolt (cable) dynamometers. When collecting data, it should be ensured that the time span is sufficient to capture the long-term and short-term changes of surrounding rock stress. From the massive data, a representative and complete data set is selected, and these data should comprehensively reflect the change trend of stress.

[0077] In Step 2, Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) is an improvement of the EEMA and EMD algorithms. By adding white noise in a specific frequency band during the decomposition process, it effectively reduces mode mixing and component reconstruction errors, and decomposes the original signal into multiple Intrinsic Mode Functions (IMFs). This improvement enables CEEMDAN to have both less mode mixing and less component reconstruction error. It can better handle the complex characteristics of surrounding rock stress signals and provide clearer input data for subsequent prediction models. The specific decomposition process is as follows:

[0078] 2.1) Add Gaussian white noise with a mean of 0 to the original signal x(t) to construct a preprocessing sequence x i (t) (i = 1, 2, ~, K).

[0079] x i (t) = x(t) + εδ i (t)

[0080] In the formula, ε is the weight coefficient of the added Gaussian white noise, and δ i (t) is the Gaussian white noise at the i-th processing.

[0081] 2.2) Perform EMD decomposition on the above preprocessing sequence x i (t) to decompose it into i components And calculate its mean value as the first IMF component I1(t) obtained by CEEMDAN decomposition.

[0082]

[0083] r1(t) = x(t) - I1(t)

[0084] Wherein, I1(t) is the first IMF component obtained by CEEMDAN decomposition; is the i-th IMF component obtained after EMD decomposition of x i (t); r1(t) is the residual component obtained after the first decomposition.

[0085] 2.3) After adding Gaussian white noise to the residual signal in the j-th stage obtained by decomposition, continue with EMD decomposition.

[0086]

[0087] r j (t) = r j-1 (t) - I j (t)

[0088] Wherein, I j (t) is the j-th IMF component obtained by CEEMDAN decomposition; E j-1 is the (j - 1)-th IMF component after EMD decomposition of the sequence; ε j-1 is the noise coefficient added to the residual component in the (j - 1)-th stage by CEEMDAN decomposition; r j (t) is the residual component in the j-th stage.

[0089] 2.4) Repeat the above steps until the number of extreme points of the residual component decreases to a certain number (usually set to be less than or equal to 2) and no further decomposition can be performed, then the CEEMDAN decomposition ends. At this time, the original signal is decomposed into several IMF components and a residual component.

[0090] CEEMDAN decomposes the surrounding rock stress signal into several Intrinsic Mode Functions (IMFs), and each IMF represents the stress components at different time scales. Although these IMFs have clear physical meanings, they are numerous, and some IMFs may contain similar characteristics, which will increase the computational complexity if directly used in the prediction model. To further process the decomposed surrounding rock stress signal, extract more representative features, and provide clearer input data for the subsequent prediction model, the K-means algorithm is used to cluster the decomposed surrounding rock stress signal. Through K-means clustering, the IMFs decomposed by CEEMDAN are divided into a trend term, a periodic term, and an error term, which respectively reflect the long-term evolution law, periodic fluctuation, and random noise of the surrounding rock stress. This method can not only simplify the data input but also effectively remove noise, ensuring that the subsequent prediction model receives a more accurate and reliable feature dataset.

[0091] K-means clustering is a commonly used unsupervised learning algorithm, and its goal is to divide the dataset into K clusters, such that each data point belongs to the cluster center with the closest distance. By repeatedly adjusting the positions of the cluster centers, K-means continuously optimizes the compactness within the clusters, thereby obtaining clusters that are as compact and separated from each other as possible. The optimization objective is to minimize the Sum of Squared Errors (SSE) within the clusters:

[0092]

[0093] In the formula, C j is the j-th cluster; μ j is the center (centroid) of the j-th cluster; x is the data point.

[0094] In step 3, the principle of using a Convolutional Neural Network (CNN) for feature extraction in surrounding rock stress prediction is to automatically extract hierarchical features from the input data through a series of convolutional layers. Using the stress values of the trend term as the input to the CNN, through multiple convolutional layers and pooling layers, the low-level and high-level features of the stress values of the trend term are gradually extracted. The convolutional layer uses multiple convolutional kernels to slide on the input data to calculate local features. The pooling layer then downsamples the output of the convolutional layer to reduce the data dimension while retaining important features. Through multiple convolutional and pooling operations, the CNN can capture the complex non-linear patterns in the trend term stress data. After the feature extraction by the CNN, the obtained feature vectors are passed to the SVM for regression prediction.

[0095] Specifically: First, the convolutional layer uses multiple convolutional kernels to locally perceive the surrounding rock stress data, and captures low-level features such as short-term fluctuations and local trends through weighted summation and non-linear activation functions. As the network depth increases, the convolutional layer gradually captures more high-level features such as long-term trends and periodic changes. The pooling layer further reduces the data dimension, retains significant feature information, and prevents overfitting. Finally, the extracted features are integrated and predicted through the fully connected layer. During the model training process, the CNN automatically learns and adjusts the weights of the convolutional kernels, enabling the model to effectively identify and extract features related to the surrounding rock stress prediction target, achieving accurate prediction results. The specific principle formula is as follows:

[0096]

[0097] In the formula, y, h, and u are all sequences, k represents the number of convolutions, and N is the length of u.

[0098] Support Vector Machine (SVM) is a powerful machine learning algorithm widely used in classification and regression problems. In the prediction of surrounding rock stress, SVM fits the non-linear relationship of stress data by finding the optimal hyperplane to distinguish different categories of data points and achieve the processing of complex stress change patterns. However, SVM has limitations in long-term time series prediction, especially in comprehensively capturing the multi-factor effects on surrounding rock stress (such as geological conditions, mining activities, etc.), resulting in limited prediction accuracy. In addition, the computational complexity of SVM is relatively high, the training time is long, and it is sensitive to the selection of kernel functions and parameters. Improper parameter tuning may affect the model performance. Therefore, although SVM performs well in short-term prediction, in the long-term prediction of surrounding rock stress, it is still necessary to combine with other algorithms or improve the model to improve accuracy and generalization ability.

[0099] In the regression problem, the goal of SVM is to find a function that can minimize the error between the predicted value and the actual value. In the SVM model of the present invention, SVM uses the feature vectors extracted by CNN to learn the non-linear change law of the trend term stress data and outputs the final prediction result.

[0100] Thus, the CNN-SVM fusion model utilizes the feature extraction ability of CNN and the classification advantage of SVM. CNN extracts high-level features from the data (such as local patterns of time series), and then inputs these features into SVM for classification or regression. SVM processes non-linear relationships through kernel functions to enhance the model generalization ability. In the prediction of surrounding rock stress, the combination of CNN-SVM can effectively capture the complex patterns of stress data and improve the prediction accuracy.

[0101] In step 4, the SAA-CNN-BiLSTM hybrid model combines a convolutional neural network (CNN) optimized by the Sparrow Search Algorithm (SSA) with a bidirectional long short-term memory network (BiLSTM), as Figure 3 shown. The model receives data through the input layer, extracts features using the CNN layer, captures the forward and backward dependencies of the time series by the bidirectional LSTM layer, and finally gives the prediction result through the output layer. During the model training process, the Adam optimizer is used to adjust the weights, and the SSA algorithm is used to optimize the hyperparameters, such as the convolutional kernel size, the number of LSTM neurons, etc., to improve the model performance.

[0102] Since the surrounding rock stress data has time dependence and dynamic evolution characteristics, and the bidirectional long short-term memory network (BiLSTM) can capture both the forward and backward dependencies of the time series, effectively dealing with the nonlinear, non-stationary and dynamic evolution characteristics of the surrounding rock stress data. As Figure 2 shown, BiLSTM is a recurrent neural network structure composed of two independent LSTMs. The LSTM weights the input data through its gating units (including the forget gate, input gate and output gate), determining which information should be remembered, forgotten or output, so as to be able to process long sequence data. In BiLSTM, one LSTM unit is responsible for capturing the forward information of the input sequence, and the other LSTM unit is responsible for capturing the backward information. Their outputs are concatenated or combined at each time step to provide more comprehensive sequence information.

[0103] The Sparrow Search Algorithm (SSA) is a swarm intelligence optimization algorithm that simulates the foraging behavior of sparrows. Its core idea is to seek the optimal solution by simulating the social behavior and individual behavior demonstrated by sparrows during the foraging process. Using the Sparrow Search Algorithm (SSA) to optimize the parameters of the deep learning model can significantly improve the training efficiency and prediction accuracy of the model. The specific process is as follows: First, initialize the sparrow population, set the population size and the maximum number of iterations, generate multiple random solutions, and each solution represents a different combination of hyperparameters; then, evaluate the fitness of each solution, that is, calculate the prediction accuracy of the corresponding CNN-BiLSTM model on the training set; then, update the position of the solution according to the rules of the sparrow foraging behavior, combining local and global search strategies; when the fitness reaches the target or the number of iterations exceeds the limit, stop the search and output the optimal combination of hyperparameters. In this way, the SSA algorithm optimizes the hyperparameters of the CNN-BiLSTM model by simulating the foraging of sparrows.

[0104] The present invention simulates the foraging behavior of sparrows to search for the optimal parameter combination and optimize the weights and hyperparameters in the CNN-BiLSTM model. SSA searches for the optimal hyperparameters in the search space, including the convolutional kernel size, the number of BiLSTM neurons, and the learning rate, etc., to ensure that the model better adapts to the complex characteristics of the surrounding rock stress data. The convolutional neural network (CNN) extracts features from the optimized signal, captures the local time dependencies, and extracts the key spatial features in the surrounding rock stress signal. Finally, the optimized features are input into the bidirectional long short-term memory network (BiLSTM) to improve the accuracy and stability of stress prediction.

[0105] In step 5, the prediction results of the trend term and the periodic term are superimposed according to the time series principle, that is, the predicted value of the trend term is added to the predicted value of the periodic term, and at the same time, the processing of the error term is considered to reduce the noise interference and obtain the final predicted value of the surrounding rock stress.

[0106] Furthermore, the present invention also discloses a dynamic prediction system for the surrounding rock stress of a deep large mining height working face based on the CEEMDAN-Kmeans decomposition and SSA-CNN-BiLSTM hybrid model, including:

[0107] A data acquisition module for obtaining in real time the monitoring data reflecting the dynamic changes of stress.

[0108] A stress signal decomposition module for decomposing the surrounding rock stress signal by using a model combining the complete ensemble empirical mode decomposition with adaptive noise CEEMDAN and the K-means clustering algorithm, and dividing it into a trend term, a periodic term, and an error term;

[0109] A trend term displacement prediction module for combining the convolutional neural network CNN and the support vector machine SVM to construct a prediction model for the surrounding rock stress of the trend term for long-term evolution law prediction;

[0110] A periodic term displacement prediction module for constructing a hybrid model of a convolutional neural network-bi-directional long short-term memory network optimized by SSA, that is, an SSA-CNN-BiLSTM hybrid model to predict the periodic fluctuations,

[0111] A prediction result fusion module for superimposing the prediction results of the trend term and the periodic term and considering the influence of the error term to obtain the final predicted value of the surrounding rock stress.

[0112] It can be seen from this that the present invention discloses a dynamic prediction method based on CEEMDAN-Kmeans. Combining the advantages of the CNN-SVM and SSA-CNN-BiLSTM hybrid models, starting from different stages of deep learning prediction, a dynamic prediction idea of surrounding rock stress based on "feature derivation - decomposition prediction - model optimization" is proposed to achieve accurate prediction of surrounding rock stress. The present invention improves the accuracy and reliability of surrounding rock stress prediction through an advanced algorithm coupling model, enhances the long-term dependence modeling and fitting ability of the model, can more accurately capture the dynamic change law of surrounding rock stress, and provides a reliable theoretical basis and technical support for the surrounding rock stress prediction of deep large mining height working faces.

[0113] The following further introduces the prediction method of the present invention in combination with specific application implementations:

[0114] Step 1, data collection.

[0115] In this embodiment, Meihuajing Coal Mine in Ningdong Energy and Chemical Industry Base is taken as the research object, and the data is from the actual monitoring of the 232205 coal mining face in Meihuajing Coal Mine. To accurately predict the mine pressure change trend of the working face roof and the surrounding rock stress and deformation law of the roadway, the researchers obtained the support working resistance in real time through the support resistance monitoring sensor in the 232205 coal mining face of Meihuajing Coal Mine. At the same time, roof separation meters, borehole stress gauges and bolt / anchor cable stress gauges were arranged in the return airway of the 232205 working face to obtain roof separation data, surrounding rock stress data and bolt / anchor cable tensile stress data in real time.

[0116] Step 2, decomposition of surrounding rock stress time series.

[0117] The surrounding rock stress data is regarded as time series data and decomposed. Time series decomposition usually includes the identification and separation of trend terms, periodic terms and error terms.

[0118] First, the CEEMAN algorithm is used to decompose the original signal (as shown in Figure 4 ) to more accurately extract different types of patterns in the signal. Subsequently, the IMF components obtained by decomposition are used as inputs, and the K-means algorithm is used for clustering analysis. The signal is divided into trend terms, periodic terms and error terms (as shown in Figure 5 ).

[0119] This decomposition and recombination method based on the CEEMDAN-Kmeans algorithm effectively extracts multi-level information in the surrounding rock stress data, enables different time scales and features to be processed independently, and can provide more accurate and reliable input data for the subsequent prediction model. Table 1 shows some data of the decomposition results. According to the modal curve characteristics of the decomposition and combined with the time series decomposition model for modal recombination, the time series decomposition formula is as follows:

[0120] S(t) = f(t) + η(t) + e(t)

[0121] Where: S(t) is the time series of stress; f(t) is the trend term representing the overall change trend of stress; η(t) is the periodic term representing the daily or seasonal change pattern; e(t) is the error term representing the noise information therein.

[0122] Table 1 Partial data table of decomposition results

[0123]

[0124]

[0125] Step 3: Prediction of the trend term.

[0126] The trend term reflects the long-term non-linear change of surrounding rock stress, and traditional linear models are difficult to meet the prediction requirements. The CNN-SVM model combines the feature extraction ability of CNN and the non-linear regression advantage of SVM, and can effectively handle the complex change law of surrounding rock stress of the trend term to achieve high-precision prediction. Among them, the first 70% of the data is selected as the training set, and the last 30% of the data is predicted;

[0127] Figure 6 Shows the prediction results and error analysis of the surrounding rock stress of the trend term. The prediction results show that the surrounding rock stress of Meihuajing Coal Mine shows an obvious upward trend during the long-term mining process, and this trend is closely related to the increase in mining depth, the increase in mining intensity and the complexity of geological conditions. Especially under the condition of deep mining, the stress concentration phenomenon is more prominent. The long-term evolution of surrounding rock stress has a significant time effect. The stress will gradually tend to be stable over time, but when continuous mining or geological conditions change, the stress may still continue to fluctuate. The CNN-SVM model shows high accuracy in prediction, with its error controlled within 0.03, the correlation coefficient as high as 0.998762, and the root mean square error of 0.039458 KN. This model provides important decision-making support for the safe production of Meihuajing Coal Mine. By predicting the long-term change trend of surrounding rock stress in advance, support measures can be taken in time to effectively avoid disasters such as roof caving and roof fall.

[0128] Step 4: Prediction of the periodic term.

[0129] The periodic term reflects the periodic fluctuation of surrounding rock stress. Affected by various external factors, it has the characteristics of non-linearity and dynamics.

[0130] Traditional single models are difficult to accurately capture such complex non-linear changes. Therefore, a CNN-BiLSTM model optimized by the Sparrow Search Algorithm (SSA) is selected. The model structure includes an input layer, two layers of CNN convolutional neural networks, a forward LSTM layer, a backward LSTM layer, and an output layer. The mean squared error is used as the loss function, and the Adam optimizer is used to adjust the model weights to minimize the loss function. To further improve the model performance, the Sparrow Search Optimization Algorithm (SCSSA) is introduced to optimize the hyperparameters of the CNN-BiLSTM model. The specific optimization process is as follows: First, initialize the population, set the population size to 10, the maximum number of iterations to 30, and generate multiple random solutions. Each solution represents a different combination of hyperparameters, including the regularization parameter, learning rate, the number of neurons in the BiLSTM model, and the convolution kernel size. Then, evaluate the fitness of each solution by calculating the prediction accuracy or loss function value of the corresponding CNN-BiLSTM model on the training dataset. Next, update the position of each solution according to the rules of the sparrow foraging behavior, combining local search and global search strategies to improve the search efficiency. Finally, when the fitness reaches the predetermined target or the number of iterations exceeds the threshold, stop the search. Input the optimized hyperparameters into the deep learning model for training, complete data prediction, and achieve high-precision prediction.

[0131] By comparing the prediction results with the true values (as Figure 7 shown), it can be observed that there are significant periodic fluctuations in the surrounding rock stress of Meihuajing Coal Mine. Such fluctuations are usually closely related to factors such as mining activities, geological conditions, and seasonal changes. The short-term fluctuations in the surrounding rock stress reflected by the periodic term may be caused by the periodic characteristics of mining activities (such as the advancement of the working face, the adjustment of support, etc.) or changes in the external environment (such as changes in groundwater pressure, seasonal temperature changes, etc.). Specifically, under the influence of mining in the 232205 working face, the distribution of the surrounding rock stress has changed significantly, and stress concentration has even occurred in some areas. The prediction results of the SSA-CNN-BiLSTM hybrid model are basically consistent with the actual values, with a correlation coefficient reaching 0.99161 and a root mean square error of only 0.01357 KN, which further confirms the close relationship between the surrounding rock stress and mining activities, as well as the effectiveness of the model in predicting the dynamic adjustment process of the surrounding rock stress.

[0132] Figure 8 The histogram of the prediction errors of the CNN-BiLSTM model before and after optimization by the Sparrow Search Algorithm (SSA) intuitively reflects the improvement effect of SSA optimization on the model prediction accuracy. The error distribution after optimization is more concentrated, and the error range is reduced, indicating that the prediction performance of the model has been significantly improved after optimization. Through Figure 9The optimized regression graph can intuitively show the prediction accuracy of the model. The data points are closely distributed near the fitting line, indicating that the prediction results of the model are highly consistent with the actual values and the performance of the model is good.

[0133] Step 5: Superimpose the predicted values.

[0134] Superimpose the predicted values of the trend term and the periodic term to obtain the final stress prediction value, and compare it with the denoised true value. As Figure 10 shown, the predicted values and the actual values are highly consistent in both the overall trend and local fluctuations. The trend term captures the long-term change law of the surrounding rock stress through the CNN-SVM model, while the periodic term accurately reflects the periodic fluctuations of the surrounding rock stress through the CNN-BiLSTM model optimized by SSA, featuring non-linearity and dynamics. The superposition of the two fully considers the non-linear, dynamic and multi-scale characteristics of the surrounding rock stress. In addition, the Sparrow Search Algorithm (SSA) optimizes the hyperparameters of the CNN-BiLSTM model through global and local searches, significantly improving the convergence speed and prediction performance of the model.

[0135] Furthermore, verify the accuracy and reliability of the prediction model by comparing it with the actual monitoring data.

[0136] In summary, the present invention designs a dynamic prediction method for surrounding rock stress based on a multi-algorithm coupling model in view of the characteristics of the dynamic change of surrounding rock stress in deep large mining height working faces. Through the advanced multi-algorithm coupling model, high-precision dynamic prediction of surrounding rock stress is achieved, thus effectively preventing surrounding rock stress disasters in coal mining and ensuring the safety of coal mine production and personnel and property. Compared with the prior art, the beneficial effects are as follows:

[0137] (1) The present invention uses the CEEMDAN-Kmeans algorithm to decompose the original surrounding rock stress data, successfully decomposing the stress value into a trend term, a periodic term and an error term, effectively overcoming the mode mixing problem in traditional empirical mode decomposition (EMD) and removing noise interference, providing a high-quality data basis for subsequent accurate stress prediction.

[0138] (2) The present invention combines the CNN-SVW model to predict the trend term and constructs an SSA-CNN-BiLSTM hybrid model to predict the periodic term. This multi-algorithm fusion method can more comprehensively capture the dynamic change characteristics of surrounding rock stress, significantly improving the accuracy and reliability of prediction.

[0139] (3) The present invention uses the Sparrow Search Algorithm (SSA) to optimize the hyperparameters of the CNN-BiLSTM model, searching for the best parameter combination by simulating the foraging behavior of sparrows, significantly improving the training efficiency and prediction accuracy of the fusion model, enabling the model to better adapt to the complex characteristics of surrounding rock stress data.

[0140] (4) The method proposed by the present invention is not only applicable to the prediction of surrounding rock stress in deep large mining height working faces, but also can be extended and applied to the prediction of mine pressure under other complex geological conditions, enhancing the generalization ability of the model.

[0141] The above embodiments are not limitations to the present invention, and the present invention is not limited to the above examples either. Changes, modifications, additions or substitutions made by those skilled in the art within the scope of the technical solution of the present invention also fall within the protection scope of the present invention.

Claims

1. A method for dynamically predicting the surrounding rock stress of a deep large mining height working face, characterized in that: It includes the following prediction steps: Step 1: Data collection and selection: Select long-term monitoring data that can reflect the dynamic changes of surrounding rock stress for collection; Step 2: Stress signal decomposition: Perform complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) on the surrounding rock stress data to obtain several intrinsic mode functions (IMFs); then, through the K-means clustering algorithm, divide the IMFs after CEEMDAN decomposition into trend terms, periodic terms, and error terms; Step 3: Trend term prediction: Use the convolutional neural network-support vector machine (CNN-SVW) model to predict the trend term stress values to capture the long-term non-linear change trend of the surrounding rock stress; Step 4: Periodic term prediction: Construct a hybrid model of a convolutional neural network-bi-directional long short-term memory network optimized by the sparrow search algorithm (SSA), namely the SSA-CNN-BiLSTM hybrid model, to predict the periodic term stress values to obtain the periodic changes of the surrounding rock stress; Step 5: Superposition of predicted values: Superpose the predicted results of the trend term and the periodic term, and consider the influence of the error term to obtain the final predicted value of the surrounding rock stress.

2. The dynamic prediction method for surrounding rock stress of deep large mining height working face according to claim 1, characterized in that: In Step 2, the specific process of CEEMDAN decomposition is as follows: Step 2.1: Add Gaussian white noise with zero mean to the original signal x(t) to construct the preprocessing sequence x i (t) (i = 1, 2, ~, K): x i y(t) = x(t) + εδ i y(t) where ε is the weighting coefficient of the added Gaussian white noise, and δ i (t) is the Gaussian white noise at the i-th processing; Step 2.

2. Perform EMD decomposition on the preprocessed sequence x i (t) to obtain i components and calculate their mean value as the first IMF component I1(t) obtained by CEEMDAN decomposition: r1(t) = x(t) - I1(t) Where, I1(t) is the first IMF component obtained by CEEMDAN decomposition; is the i-th IMF component obtained by performing EMD decomposition on x i (t); r1(t) is the residual component obtained after the first decomposition; 2.3) Add Gaussian white noise to the residual signal obtained after decomposition in the j-th stage, and continue with EMD decomposition: r j r(t) = j-1 I(t) - j I(t) Where, I j (t) is the j-th IMF component obtained by CEEMDAN decomposition; E j-1 is the (j - 1)-th IMF component after EMD decomposition of the sequence; ε j-1 is the noise coefficient added to the residual component in the (j - 1)-th stage by CEEMDAN decomposition; r j (t) is the residual component in the j-th stage; 2.4) Repeat the above steps until the number of extreme points of the residual component decreases to a preset value, then the CEEMDAN decomposition ends; at this time, the original signal is decomposed into several IMF components and a residual component.

3. The method for dynamically predicting the surrounding rock stress of a deep large mining height working face according to claim 2, characterized in that: The preset value of the number of extreme points of the residual component is set to be less than or equal to 2.

4. The deep large mining height working face surrounding rock stress dynamic prediction method according to claim 3, characterized in that: In Step 2, through K-means clustering, the IMFs after CEEMDAN decomposition are divided into trend terms, periodic terms, and error terms, which respectively reflect the long-term evolution law, periodic fluctuations, and random noise of the surrounding rock stress; The optimization objective of K-means clustering is to minimize the within-cluster sum of squares error (SSE): where C j is the j-th cluster; μ j is the center or centroid of the j-th cluster; x is a data point.

5. The dynamic prediction method for surrounding rock stress of deep large mining height working face according to claim 1, characterized in that: In Step 3, use the trend term stress value data as the input of the convolutional neural network (CNN). Through multiple convolutional layers and pooling layers, gradually extract the low-level and high-level features of the trend term stress value data; The convolutional layer uses multiple convolutional kernels to perform local perception on the input surrounding rock stress data, captures low-level features through weighted summation and non-linear activation functions, and as the network depth increases, the convolutional layer gradually captures high-level features; The pooling layer performs downsampling on the output of the convolutional layer; Through multiple convolutional and pooling operations, the convolutional neural network (CNN) captures the complex non-linear patterns in the trend term stress value data; After feature extraction by CNN, the obtained feature vector is passed to the support vector machine (SVM) model for regression prediction.

6. The dynamic prediction method for surrounding rock stress of deep large mining height working face according to claim 5, characterized in that: In Step 3, the support vector machine (SVM) model uses the feature vector extracted by CNN to learn the non-linear change law of the trend term stress data based on the kernel function and outputs the final prediction result.

7. The dynamic prediction method for surrounding rock stress of deep large mining height working face according to claim 1, characterized in that: The SAA-CNN-BiLSTM hybrid model integrates the sparrow search algorithm (SSA), the optimized convolutional neural network (CNN), and the bi-directional long short-term memory network (BiLSTM); The hybrid model receives data through the input layer, extracts features using the CNN layer, captures the forward and backward dependencies of the time series by the bidirectional LSTM layer, and finally gives the prediction result through the output layer.

8. The method for dynamically predicting the surrounding rock stress of a deep large mining height working face according to claim 7, characterized in that: During the training process of the hybrid model, the Adam optimizer is used to adjust the weights, and the sparrow search algorithm SSA is used to optimize the hyperparameters, including the convolutional kernel size and the number of LSTM neurons, to improve the model performance.

9. The deep large mining height working face surrounding rock stress dynamic prediction method according to claim 8, characterized in that: The sparrow search algorithm SSA simulates the foraging behavior of sparrows to optimize the hyperparameters. The specific process is as follows: First, initialize the sparrow population, set the population size and the maximum number of iterations, and generate multiple random solutions. Each solution represents a different combination of hyperparameters; Then, evaluate the fitness of each solution, that is, calculate the prediction accuracy of the corresponding CNN-BiLSTM model on the training set; Next, update the position of the solution according to the rules of the sparrow foraging behavior, combining local and global search strategies; When the fitness reaches the target or the number of iterations exceeds the limit, stop the search and output the optimal combination of hyperparameters.

10. A system for the dynamic prediction method of surrounding rock stress in a deep large mining height working face as described in claim 1, characterized in that: The system includes: A data acquisition module for real-time acquisition of monitoring data reflecting the dynamic changes of stress. A stress signal decomposition module for decomposing the surrounding rock stress signal using a model that combines the complete ensemble empirical mode decomposition with adaptive noise CEEMDAN and the K-means clustering algorithm, and dividing it into a trend term, a periodic term, and an error term; A trend term displacement prediction module for constructing a prediction model of the surrounding rock stress of the trend term by combining the convolutional neural network CNN and the support vector machine SVM for long-term evolution law prediction; A periodic term displacement prediction module for constructing a hybrid model of a convolutional neural network - bidirectional long short-term memory network optimized by SSA, that is, an SSA-CNN-BiLSTM hybrid model for predicting periodic fluctuations, A prediction result fusion module for superimposing the prediction results of the trend term and the periodic term to obtain the final predicted value of the surrounding rock stress.

Images