Microseismic "time-space-intensity" based rock burst prediction method

Abstract

The application discloses a rock burst full-time-space prediction method based on microseismic time-space-intensity, which comprises a microseismic time-space characteristic index module, a time prediction module and a space prediction module, and the specific method is as follows: S1, collecting original microseismic data; S2, constructing a time series data set M and obtaining microseismic time-space characteristic indexes after processing the data set; S3, in the time prediction module, generating a new time data set M' based on the processed time series data set M and the microseismic time-space characteristic indexes, constructing a microseismic data precursor mode sequence by using the data set, and constructing a rock burst time prediction model by using the microseismic data precursor mode sequence and LSTM and full connection network; and S4, in the space prediction module, respectively establishing long and short time window dangerous area space cloud maps, constructing a rock burst dangerous area cloud map through weight superposition, and completing the construction of the rock burst space prediction method. The application can realize full-time-space prediction of rock burst time-space cooperation.

Description

Technical Field

[0001] This invention relates to a method for predicting rockbursts in all time and space, specifically a method for predicting rockbursts in all time and space based on microseismic "time-space-intensity" data, belonging to the field of coal mine safety technology. Background Technology

[0002] Rockburst monitoring and early warning is a current research hotspot in the field of coal mine safety.

[0003] Monitoring methods for rockbursts are mainly divided into three categories: mining methods, geophysical methods, and multi-parameter integrated monitoring and early warning methods. Mining methods assess and predict rockburst hazards based on different mining theories, and mainly include drill cuttings analysis, comprehensive index methods, coal and rock mass deformation observation methods, coal and rock strata impact tendency classification methods, numerical simulation methods, stress monitoring methods, and working face support resistance monitoring methods. Geophysical methods continuously record dynamic phenomena occurring within coal and rock masses to monitor the rockburst hazard status in real time, and mainly include microseismic methods, electromagnetic radiation methods, and acoustic emission (ground sound) methods. Multi-parameter integrated monitoring and early warning methods establish a mine detection and evaluation system based on multiple parameters such as comprehensive index methods, vibration wave CT, microseismic analysis, acoustic-electric integration, and stress monitoring, and conduct regional prediction and local early warning based on this system. This type of method comprehensively considers multiple factors contributing to rockburst occurrence, improving the accuracy of monitoring and early warning. Pan Yishan et al. constructed a mine seismic detection and positioning system and achieved good results in field application; Dou Linming et al. constructed a spatiotemporal architecture-based monitoring and early warning technology for rockburst "stress field-vibration wave field" for rockburst hazard classification and prediction; Qi Qingxin et al. formulated the corresponding prevention and control ideas and technical framework for rockburst hazard accurate identification and intelligent recognition technology.

[0004] Rockburst hazard prediction based on microseismic monitoring data is a current research hotspot, with related work encompassing both temporal and spatial prediction. Temporal rockburst prediction refers to predicting the probability of rockburst occurrence from a temporal perspective. Chen Jie et al. constructed a physical-driven Bayesian probabilistic integrated model and a data-driven deep learning model MSNet, integrating them to complete a "dual-drive" intelligent early warning architecture for rockbursts. Cao Anye et al. explained the shortcomings and deficiencies of using only physical indicators for rockburst prediction, proposing a temporal prediction method that integrates physical indicators and data features to predict high-energy events with different rockburst hazard levels. Malkowski et al. used the changing trends of hazard autoregressive coefficients and parameter b values ​​to predict the probability of high-energy microseismic events. Wang et al. used various statistical methods to analyze the spatiotemporal correlation of microseismic event energy. Spatial rockburst prediction refers to predicting the probability of rockburst occurrence in different regions from a spatial perspective. Pan Yishan et al.'s research shows that spatial clustering, gradient significance, and curvature significance indicators can describe and track the spatiotemporal evolution characteristics of deformation localization, demonstrating the feasibility of using deformation localization to predict rockburst hazard areas. Liu Yaoqi et al. constructed a comprehensive prediction index for rockburst hazard based on the principle of dynamic and static load superposition, effectively predicting high-energy events and rockburst manifestation areas. Wen Zhijie et al. predicted the possible locations of rockbursts and proposed control measures. Rong Hai et al. proposed a multi-factor pattern recognition prediction method for rockburst hazard in steeply dipping extra-thick coal seams, classifying the rockburst hazard levels in different areas of the Wudong Coal Mine. Zhang Hongwei et al., based on a multi-factor pattern recognition method of geodynamic zoning, achieved refined unit-based prediction of rockburst hazard. However, current rockburst prediction work often separates temporal and spatial prediction, using different time windows for prediction, making it difficult to analyze the prediction results collaboratively from both temporal and spatial dimensions. In addition, most existing microseismic indices for rockburst prediction are constructed based on a single temporal or spatial dimension, failing to fully explore the spatiotemporal correlation information of microseisms, thus making it difficult to achieve full-time-space (time-space dimension) prediction of rockbursts. Summary of the Invention

[0005] This invention aims to at least solve one of the technical problems existing in the prior art. To this end, this invention proposes a spatiotemporal prediction method for rockburst based on the "time-space-intensity" of microseismic events, which can achieve spatiotemporal prediction of rockburst in a coordinated manner.

[0006] According to the present invention, a method for predicting rockburst in all time and space based on microseismic "time-space-intensity" includes a microseismic spatiotemporal characteristic index module, a time prediction module, and a spatial prediction module. The specific method is as follows:

[0007] S1 acquires raw microseismic data, including microseismic time, source coordinates, and energy magnitude;

[0008] In the microseismic spatiotemporal characteristic index module, S2 constructs a time series dataset M using microseismic time and source coordinates. Then, it uses the PCA method to reduce the dimensionality of the time series dataset. Finally, it uses the KDE method to process the dimensionality-reduced data to obtain the probability density value of each data point. These probability density values ​​are defined as microseismic spatiotemporal characteristic indices, which can reflect the degree of data aggregation within this time period.

[0009] In the time prediction module, S3 generates a new time dataset M' based on the processed time series dataset M and the spatiotemporal characteristics of microseismic data. It then uses this dataset to construct a precursor pattern sequence of microseismic data. Based on the precursor pattern sequence of microseismic data, it uses LSTM to build a time prediction model for rockburst. It also uses a fully connected network to achieve accurate prediction of rockburst time. The model is then trained and tested to complete the construction of the rockburst time prediction method, which is used to predict the hazard level of rockburst in the future.

[0010] In the spatial prediction module, S4 uses the microseismic spatiotemporal characteristic index ρ to construct spatial prediction cloud maps of dangerous areas with short time windows T1 and long time windows T2, respectively. By using weight superposition, the spatial prediction cloud maps of dangerous areas with short and long time windows are fused to construct a cloud map of the rockburst dangerous area. Based on the dangerous areas in the cloud map, accurate prediction of rockburst space is achieved, thus completing the construction of the rockburst spatial prediction method.

[0011] Compared with existing technologies, this invention includes a time prediction method and a spatial prediction method for rockbursts. The time prediction method, based on a deep recurrent neural network, includes data preprocessing, prediction model construction, model training, and model testing. Data preprocessing transforms the "time-space-intensity" characteristics and microseismic spatiotemporal characteristic indicators into precursor pattern sequences with specific time windows. The prediction model construction uses a deep recurrent neural network for data training and a fully connected network for model classification. In model training and testing, it was found that using a time window I of 6 hours to predict future duration N of 6 hours yields the best results. For spatial prediction and prevention of rockbursts, this invention constructs long and short time window data, normalizes and confirms the weights of each time window data, uses Kriging interpolation to process the normalized data, and obtains a grid matrix of microseismic spatiotemporal characteristic indicators by weight superposition. The grid matrix is ​​used to draw a spatial cloud map, establish the danger zone, and complete the construction of the spatial prediction method for rockbursts. This invention combines time and spatial prediction methods for rockbursts, achieving coordinated time-space prediction of rockbursts with accurate prediction results. Attached Figure Description

[0012] Figure 1 This is a flowchart of a method for predicting rockburst in all time and space based on microseismic "time-space-intensity" according to an embodiment of the present invention;

[0013] Figure 2 This is a flowchart of the microseismic spatiotemporal characteristic index processing process of a microseismic "time-space-intensity" all-time-space prediction method for rockburst based on a microseismic "time-space-intensity" according to an embodiment of the present invention.

[0014] Figure 3 This is a structural diagram of a time prediction model for rockburst based on a microseismic "time-space-intensity" all-time-space prediction method according to an embodiment of the present invention.

[0015] Figure 4 This is a cloud map of the impact hazard area based on the microseismic "time-space-intensity" all-time-space prediction method for rockburst according to an embodiment of the present invention; wherein, (a) is a cloud map of the impact hazard area with a short time window, (b) is a cloud map of the impact hazard area with a long time window, and (c) is a cloud map of the impact hazard area fused from the long and short time windows.

[0016] Figure 5 The results of the rockburst prediction experiment for the next 6h / 12h / 18h based on the microseismic "time-space-intensity" all-time-space prediction method according to an embodiment of the present invention are as follows:

[0017] Figure 6 This is a specific distribution of the prediction error of the 6-hour rockburst hazard level and prediction level based on the microseismic "time-space-intensity" all-time-space prediction method according to an embodiment of the present invention.

[0018] Figure 7 This is a specific distribution of the prediction error of the 12-hour rockburst hazard level and prediction level based on the microseismic "time-space-intensity" all-time-space prediction method according to an embodiment of the present invention.

[0019] Figure 8 This is a specific distribution of the prediction error of the 18-hour rockburst hazard level and prediction level based on the microseismic "time-space-intensity" all-time-space prediction method according to an embodiment of the present invention.

[0020] Figure 9 The results of the time window 6h / 12h / 18h rockburst prediction experiment of the microseismic "time-space-intensity" all-time-space prediction method according to an embodiment of the present invention are as follows:

[0021] Figure 10 This is a specific distribution of the hazard level and prediction level error of the 12-hour time window of the rockburst prediction method based on the "time-space-intensity" microseismic data according to an embodiment of the present invention.

[0022] Figure 11This is a specific distribution of the hazard level and prediction level error of the rockburst based on the microseismic "time-space-intensity" all-time-space prediction method according to an embodiment of the present invention, within a time window of 18 hours.

[0023] Figure 12 This is a specific error distribution of the rockburst hazard level and prediction level based on the microseismic "time-space-intensity" all-time-space prediction method according to an embodiment of the present invention, with only the microseismic time-space characteristic indexes added;

[0024] Figure 13 This is an engineering application of the microseismic "time-space-intensity" all-time-space prediction method for rockburst based on an embodiment of the present invention, showing the specific distribution of rockburst hazard level and prediction level error;

[0025] Figure 14 This is a statistical prediction diagram of high-energy events based on the spatiotemporal characteristic indicators of microseismic rockbursts according to an embodiment of the present invention.

[0026] Figure 15 This is a full-time and space-time prediction result diagram of the rockburst full-time and space-time prediction method based on microseismic "time-space-intensity" according to an embodiment of the present invention, wherein (a) is the full-time and space-time prediction result from 0:00 to 6:00 on March 8, (b) is the full-time and space-time prediction result from 0:00 to 6:00 on March 25, and (c) is the full-time and space-time prediction result from 0:00 to 6:00 on March 27.

[0027] Figure 16 This is a diagram showing the full-time and space-time prediction results of a high-energy event based on a microseismic "time-space-intensity" full-time and space-time prediction method for rockburst according to an embodiment of the present invention.

[0028] Figure 17 yes Figure 1 Enlarged schematic diagram of medium-to-long-term window cloud plot;

[0029] Figure 18 yes Figure 1 Enlarged schematic diagram of the cloud map for short to medium time windows;

[0030] Figure 19 yes Figure 1 An enlarged schematic diagram of the cloud map of the central impact hazard area. Detailed Implementation

[0031] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0032] The following disclosure provides numerous different embodiments or examples for implementing various structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. These are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples. Such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed. Additionally, examples of various specific processes and materials are provided in this invention; however, those skilled in the art will recognize the applicability of other processes and / or the use of other materials.

[0033] The following is for reference. Figures 1 to 19 This invention describes a spatiotemporal prediction method for rockburst based on microseismic "time-space-intensity" according to embodiments of the present invention, such as... Figure 1 , Figure 17 , Figure 18 and Figure 19 As shown, the specific steps include:

[0034] S1 acquires raw microseismic data, including microseismic time, source coordinates, and energy magnitude;

[0035] In the microseismic spatiotemporal characteristic index module, S2 constructs a time series dataset M using microseismic time and source coordinates. Then, it uses PCA to reduce the dimensionality of the time series dataset, and finally uses KDE to process the dimensionality-reduced data, obtaining the probability density value for each data point. These probability density values ​​are defined as microseismic spatiotemporal characteristic indices, which reflect the degree of data aggregation within a specific time period and are of great significance in subsequent temporal and spatial prediction applications. Figure 2 As shown, the specific steps include:

[0036] S2.1 Constructs a time series dataset using microseismic time and source coordinates, as follows:

[0037] Microseismic data were statistically analyzed using a fixed time window, and a microseismic dataset m was constructed, where the i-th microseismic dataset is m. i , is represented as:

[0038]

[0039] In the formula, x i ,y i ,z i ,t i Let X, Y, Z be the source coordinates of the microseismic data contained in the i-th data entry, and let M be the microseismic data record for the j-th time window. j The number of microseismic data is n, represented as:

[0040] Mj =[m0,m1,...,m n-1 (2)

[0041] When the microseismic dataset m is divided into k time windows, the resulting time series dataset M is represented as:

[0042] M = [M0, M1, ..., M k-1 (3)

[0043] S2.2 Due to the k-th time series dataset M k-1 The included microseismic dataset m contains four dimensions: source coordinates X, Y, Z, and microseismic time t. When the k-th time series dataset M... k-1 When the amount of data is too large, each data m i The large spatial distance between the data sets increases the difficulty of data analysis and results in poor visualization. Therefore, this embodiment uses Principal Component Analysis (PCA) to analyze the microseismic dataset m from step S2.1. i Data dimensionality reduction is performed on the data in this study. Principal Probe Analysis (PCA), a commonly used dimensionality reduction and data analysis method, transforms original high-dimensional data into a new set of low-dimensional features through linear transformation. These new features are linear combinations of the original features. Its goal is to find a new set of features that maximizes the variance of the data in the new feature space, thus preserving the most important information in the data. PCA can transform original high-dimensional data into low-dimensional data while retaining the main information of the data; these new features are called principal components. The principal components are arranged in descending order of data variance; therefore, the first few principal components contain most of the data variance, allowing the original data to be represented using lower-dimensional data. This method was used to process the microseismic dataset m, which was reduced from 4D to 2D. i 'The combined microseismic dataset M within the j-th dimensionality-reduced time window' j , is represented as:

[0044] M' j =[m'0,m1',...,m' n-1 (4)

[0045] S2.3 Processing the microseismic dataset M' using kernel density estimation (KDE) j The probability density ρ corresponding to each data point in the microseismic dataset is obtained. i The probability density ρ i This can indicate the microseismic dataset M' j The correlation between various microseismic data points. Therefore, early detection of the correlation between microseismic events can help predict high-energy events. The KDE formula is expressed as:

[0046]

[0047] in, is the probability density estimate at the point; u is the number of data points in the sample; x i K is the i-th sample data point; K(·) is the kernel function, representing the local probability density of each data point; h is the bandwidth parameter, which controls the width of the kernel function and affects the smoothness of the estimation.

[0048] Kernel density estimation (KDE) is a nonparametric statistical method used to estimate the probability density function of an unknown random variable. It estimates the overall probability density function by treating each data point as a small probability density function (also called a "kernel function" or "window function") and then summing all kernel functions together. KDE can effectively reveal the distribution characteristics of data, thus enabling the analysis of microseismic datasets M'. j The structure and characteristics of data in China.

[0049] S2.4 will use the microseismic dataset M' j probability density ρ i The set, as the spatiotemporal characteristic index ρ of microseismic events, is expressed as:

[0050] ρ=[ρ0,ρ1,...,ρ i-1 (6)

[0051] When 0 ≤ ρ < 0.2, it indicates a no-hazard area; when 0.2 ≤ ρ < 0.45, it indicates a weakly hazardous area.

[0052] When 0.45≤ρ<0.7, it represents a medium-risk area; when 0.7≤ρ≤1, it represents a high-risk area.

[0053] S3. Because Deep Recurrent Neural Networks (LSTM) can effectively capture long-term dependencies when processing sequence data, solving the gradient vanishing problem of traditional Recurrent Neural Networks (RNNs), this embodiment generates a new time dataset M' based on the processed time series dataset M and microseismic spatiotemporal characteristic indicators in the time prediction module, and uses this dataset to construct a precursor mode sequence of microseismic data; based on the precursor mode sequence of microseismic data, an LSTM is used to construct a rockburst time prediction model, and a fully connected network is used to achieve accurate prediction of rockburst time, thus completing the construction of the rockburst time prediction method. The specific method is as follows:

[0054] S3.1 Data Preprocessing: Statistical analysis is performed on the processed time series dataset M to obtain the frequency of maximum energy, average energy, and microseismic times in the time series. Combined with the spatiotemporal characteristic index ρ of microseismic events, a new time series dataset M' is generated. This dataset is then used to construct a precursor model sequence, as detailed below:

[0055] To make the data suitable for model prediction and training, the raw microseismic data, including microseismic occurrence time, source coordinates, energy magnitude, and spatiotemporal characteristic indicators, are preprocessed:

[0056] Statistical analysis of the j-th time window data in a time series dataset M yields data record s. j , is represented as:

[0057] s j =[id,e max ,e avg ,f,ρ max (7)

[0058] In the formula: id is the time window number, e max e represents the maximum energy of the micro-vibration within the time window. avg Let f be the average energy of the micro-seismic events within the time window, and ρ be the frequency of the micro-seismic events within the time window. max This represents the maximum value among the spatiotemporal characteristic indicators of microseismic events within the time window. When the microseismic dataset m is divided into k time windows, a new time series dataset M' is obtained, expressed as:

[0059] M'=[s0,s1,...,s k-1 (8)

[0060] Furthermore, based on the time window s j Construct a set of labels R:

[0061] R = [r0, r1, r2, ..., r j-1 (9)

[0062] Where, r j The label represents the corresponding time window.

[0063] Based on the predicted energy level results, the rockburst hazard level is determined. The hazard levels include strong, medium, weak and none, and the specific correspondence is shown in Table 1.

[0064] Table 1 Classification of rockburst hazard levels corresponding to microseismic energy

[0065]

[0066] A precursor pattern sequence is constructed based on the time series dataset M', where the e-th precursor pattern sequence is w. e :

[0067]

[0068] Where f is the sampling step and p is the length of the precursor pattern sequence, a set W containing q precursor model sequences is constructed, and its expression is:

[0069] W=[w0,w1,w2,...,w q-1 (11)

[0070] S3.2 Constructing a rockburst time prediction model: LSTM is used to model and extract features from precursor pattern sequences to obtain regression loss and data features. The parameters in the LSTM model are continuously updated through recursive backpropagation to minimize the loss of the model on the training dataset. A fully connected network is used to achieve accurate prediction of rockburst time, thus obtaining the rockburst time prediction model.

[0071] Specifically, the aforementioned precursor pattern sequence W is fed into an LSTM for training, implicit features are extracted, and a fully connected layer and activation function are used to classify the rockburst time prediction model. The loss value calculated by the classification loss function is selected as the regression loss of the LSTM. A recursive backpropagation method is used to continuously update the parameters in the rockburst time prediction model, minimizing the model's loss on the training dataset, thus obtaining the final rockburst time prediction model, which can predict the magnitude of microseismic events occurring within a future period (e.g., within the next 6 hours). The structure of the rockburst time prediction model of this invention is as follows: Figure 3 As shown.

[0072] The time prediction model for rockbursts uses fully connected layers and activation functions for normalization to obtain the probability of occurrence of each hazard level of rockburst, which is then converted into strong, moderate, weak, and no rockburst hazard, ultimately completing the prediction of the rockburst level. For example, if the probability of the predicted result of the rockburst time prediction model is 40% for 0, 30% for 1, 20% for 2, and 10% for 3, then the final prediction result is 0, i.e., weak hazard.

[0073] The activation function formula is as follows:

[0074]

[0075] In the formula, x k The output value of the kth node can be converted to the range [0,1] using the softmax function.

[0076] LSTM can be used to uncover relationships between microseismic event sequences. It captures dependencies between microseismic events through a "gating mechanism," avoiding the gradient vanishing problem. LSTM primarily uses three control gates: the forget gate, the input gate, and the output gate to handle input and output. The forget gate filters and retains the memory state from the previous time step, the input gate determines the amount of input data at the current time step that can be remembered by the memory cells, and the output gate controls the amount of memory state that can be output to the current time step.

[0077] S3.3 To verify the accuracy of the present invention's all-temporal-spatial prediction method for rockburst based on microseismic "time-space-intensity" analysis, this embodiment trains and tests the rockburst spatiotemporal prediction method of the present invention, as detailed below:

[0078] S3.3.1 Collecting Sample Datasets: Collect microseismic data over a specific time period as a sample dataset, and divide this dataset into training and testing sets according to a predetermined ratio, as follows:

[0079] The sample dataset collected in this embodiment is from the 2215 working face, located in a mining area in Ordos, Inner Mongolia Autonomous Region, within the territory of Uxin Banner. A total of 9969 data points were collected from January to May 2022. During training, the dataset was divided into training and testing sets at an 8:2 ratio. Microseismic data from January to May were collected and statistically analyzed in 6-hour time windows with a sliding time window every 6 hours. Based on the collected time window data, it was found that there were a total of 27 high-energy events, and 10 high-energy events occurred frequently during March and April. 5 High-energy events above J indicate that the selected time period of data can meet the requirements for model training. Data processing is shown in Table 3 below.

[0080] Table 3 Data Preprocessing Parameter Settings

[0081]

[0082] In the training of the S3.3.2 rockburst time prediction model, the precursor pattern sequence W is used as input. To improve the convergence performance of the time prediction model, Adam is selected as the model optimizer. The training parameters of the rockburst time training model are set as follows: 200 training epochs, batch size of 128, decay rate of 0.1, and learning rate of 0.001. Classification cross-entropy is used as the loss function, accuracy and F1 score are used as evaluation functions, and Spftmax activation function is used for normalization to obtain the probability of high-energy event levels occurring, thus outputting the energy level of future events. The expression for the classification cross-entropy loss function is as follows:

[0083]

[0084] Where, p k Let y be the probability that a sample belongs to the k-th class. k is the sample label, and C is the total sample size.

[0085] The specific testing of the rockburst time prediction method in this embodiment is as follows:

[0086] (1) Determine the evaluation indicators

[0087] This embodiment uses a confusion matrix to record the prediction results of this model. There are a total of 4 possible prediction results:

[0088] If the prediction is true and the actual result is true, it is denoted as TP;

[0089] If the prediction is true and the actual result is false, it is denoted as FP;

[0090] If the prediction result is false and the actual result is true, it is denoted as FN;

[0091] If the prediction is false and the actual result is false, it is denoted as TN;

[0092] The confusion matrix is ​​shown in Table 4.

[0093] Table 4 Confusion Matrix Form

[0094]

[0095]

[0096] Accuracy (Acc) represents the percentage of correct predictions made relative to the total sample, and is expressed by the following formula:

[0097]

[0098] Precision, in relation to the predicted sample result, represents the probability that among all samples predicted as positive, the actual sample is positive. The formula is as follows:

[0099]

[0100] Recall, relative to the original sample, represents the probability that a sample that is actually positive will be predicted as positive. The formula is as follows:

[0101]

[0102] The F1 score takes into account both precision and recall, and is calculated using the following formula:

[0103]

[0104] In this embodiment, accuracy and F1 value are selected as evaluation indicators for the rockburst time prediction model.

[0105] (2) Determine the optimal forecast duration N

[0106] To verify the performance of the LSTM-based time prediction model for rockbursts (hereinafter referred to as the "time prediction model") proposed in this embodiment, the time prediction model was used to predict the level of high-energy rockburst events in the future duration N (6h, 12h, 18h).

[0107] When using 12 sets of 6-hour data to predict the next 6 hours, 12 hours, and 18 hours, the ratio of the number of samples in the training set to the number of samples in the test set is 468:117. The data preprocessing parameters are set as shown in Table 5.

[0108] Table 5 Data Preprocessing Parameter Settings

[0109]

[0110] The accuracy and F1 score of the training results were analyzed using the confusion matrix of the prediction results, as shown in Tables 6-8.

[0111] Table 6 Confusion Matrix of Predicted 6h Results

[0112]

[0113] Table 7 Confusion Matrix of Predicted Results for 12 Hours

[0114]

[0115] Table 8 Confusion Matrix of Predicted Results at 18h

[0116]

[0117] As shown in Tables 6-8, the prediction accuracies were 0.855, 0.838, and 0.821, respectively, and the F1 values ​​were 0.866, 0.790, and 0.752, respectively. It is evident that this time-based prediction model performs well in predicting rockbursts when the size of the predicted future time window N is changed. Specific prediction accuracy and F1 values ​​are as follows: Figure 13 As shown.

[0118] analyze Figure 5 It can be seen that when the predicted future time window N is set to 6 hours, the time prediction model performs best. At this time, the prediction accuracy is 0.855 and the F1 score is 0.866, both higher than the other two sets of experimental parameter settings. To record the experimental prediction error, this embodiment plots a graph showing the specific distribution of prediction level errors. This graph reflects the difference between the actual hazard level and the predicted level (for example, if the actual hazard level is strong and the predicted level is moderate, the difference is 1 level). Figures 7-8 As shown.

[0119] (3) Determine the optimal time window I

[0120] To investigate the impact of time window I on the time prediction model, under the condition that the prediction of future duration N is best at 6 hours, the size of time window I was changed to 12 hours and 18 hours respectively, and prediction experiments were conducted. The ratio of the number of samples in the training set to the test set was 467:117 and 466:117, and the data preprocessing parameters are shown in Table 9.

[0121] Table 9 Data Preprocessing Parameter Settings

[0122]

[0123] The accuracy and F1 score of the training results were analyzed using the confusion matrix of the prediction results, as shown in Tables 10 and 11; the specific distribution of prediction level errors is shown in the following figure. Figure 9-10 As shown.

[0124] Table 10 Confusion Matrix of Results for 12-Hour Time Window

[0125]

[0126] Table 11 Confusion Matrix of Results for 18h Time Window

[0127]

[0128] According to the results in Tables 6, 10, and 11, the experimental prediction accuracies were 0.855, 0.727, and 0.624, respectively, and the F1 scores were 0.866, 0.780, and 0.560, respectively. The specific prediction accuracy and F1 scores are as follows: Figure 9 As shown.

[0129] To record the experimental prediction error, this embodiment plots a graph showing the specific distribution of prediction level errors, such as... Figure 10 and Figure 11 As shown.

[0130] By predicting the specific distribution of the level error Figure 6 , Figure 10 and Figure 11 It can be seen that in these three sets of experiments, the error was at level 1 when the prediction time window I was 12 hours, and level 2 errors occurred only twice when the prediction time window I was 18 hours. This shows good control over the error level range. In summary, setting the time window I to 6 hours yielded the best results in this experiment.

[0131] The comparative experiment of predicting future duration N and time window I shows that the time prediction model has the best prediction effect when the time window I is set to 6 hours and the predicted future duration N is set to 6 hours.

[0132] To evaluate the impact of microseismic spatiotemporal characteristic indices on the time prediction model, the experimental results without the indices are shown in Table 12. The accuracy was 0.769 and the F1 value was 0.816. Compared with the experimental results with the indices (Table 6), the accuracy and F1 value of the time prediction model decreased by 8.6% and 5%, respectively. Therefore, the use of microseismic spatiotemporal characteristic indices in this invention can improve prediction accuracy.

[0133] Table 12 Confusion Matrix of Prediction Results Without Microseismic Spatiotemporal Characteristic Indicators

[0134]

[0135] Furthermore, to evaluate the impact of maximum energy, average energy, and frequency indices on the time prediction model, the experimental results were obtained by adding only the microseismic spatiotemporal characteristic indices, as shown in Table 13. The accuracy was 0.530, and the F1 score was 0.420. Compared with the experimental results (Table 6) with the maximum energy, average energy, and frequency indices added, the accuracy and F1 score of the time prediction model decreased by 32.5% and 44%, respectively, and the model failed to predict high-energy events in both trials. The specific error distribution is also detailed below. Figure 12 Error distribution of the maximum energy, average energy, and frequency indices ( Figure 11 In contrast, when only microseismic spatiotemporal characteristic indicators were included, there were 7 instances of error level 2. Therefore, using maximum energy, average energy, and frequency indicators can improve prediction accuracy.

[0136] Table 13 Confusion matrix of prediction results with only microseismic spatiotemporal characteristic indicators added.

[0137]

[0138] In summary, the time prediction method proposed in this invention can improve the accuracy of predicting future high-energy events. In the specific error distribution of the energy levels, the energy level error is controlled within one level.

[0139] To further evaluate the capabilities of the time prediction model, this embodiment underwent engineering application practice. Based on the data segmentation results described above, a time window I of 6 hours was used to predict future duration N of 6 hours, with a learning rate of 0.001. The time prediction model settings remained unchanged. Considering that there were 13 high-energy events in March at the 2215 working face, a total of 236 time periods from January and February were used to predict the energy level of data in March. Specifically, the initial time prediction model was trained using the first 236 time periods, and then the energy event level of the 237th time period was predicted. The time window was then shifted by one time period, i.e., 6 hours, and the time prediction model was updated again. The initial time prediction model was trained using the first 237 data periods, and then the energy event level of the 268th time period was predicted. This process was repeated for a total of 122 simulation training runs. The prediction results are shown in Table 14, and the specific distribution of the prediction level error in the engineering application is as follows: Figure 13 As shown.

[0140] Table 14 Confusion Matrix of Prediction Results for Engineering Applications

[0141]

[0142] As shown in Table 14, there were a total of 13 high-energy events at the 2215 working face. This time prediction model, updated with each data update, achieved an accuracy of 0.812 and an F1 score of 0.768, successfully predicting 10 of these high-energy events. Specifically, in predicting high-energy events, the model correctly predicted three other high-energy events with a strong rockburst risk as moderate rockburst risks, achieving an accuracy of 0.770. Furthermore, among the 12 high-energy events predicted as having a strong rockburst risk, only two were actually predicted as having a moderate rockburst risk, resulting in a false alarm rate of 0.133.

[0143] Depend on Figure 13 As can be seen, 99 energy events were successfully predicted, while the prediction errors for the remaining 23 energy events were all within the first level. Engineering applications demonstrate that the time prediction model proposed in this embodiment can be effectively applied in practical engineering applications.

[0144] To compare the prediction performance of microseismic spatiotemporal characteristic indices in time prediction models and traditional threshold-based rockburst prediction methods (Si G, Cai W, Wang S, et al. Prediction of Relatively High-Energy Seismic Events Using Spatial–Temporal Parametrisation of Mining-Induced Seismicity[J]. Rock Mechanics Rock Engineering, 2020, 53, 5111–5132), a threshold method was used to analyze the microseismic spatiotemporal characteristic indices. Specific results are as follows: Figure 14 As shown.

[0145] Figure 14 The center dot represents a period without major energy events, while the pentagram represents a period with major energy events, with each period defined as 6 hours. The 70% range between the maximum and minimum values ​​of the monthly microseismic spatiotemporal characteristic index is used as a dividing line. Peak values ​​exceeding 70% are used to predict the likelihood of major energy events occurring within the next 3 days (12 time periods). Figure 14 It can be seen that 5 out of 13 high-energy events in March could be predicted successfully (i.e., the data between the two lines of the same color in the figure), with an accuracy of only 0.385. Compared with the prediction results using the threshold method, the time prediction method proposed in this paper improves the effect by 38.5%, and can predict high-energy events in advance. In engineering, this allows for the early implementation of rockburst prevention measures to reduce losses.

[0146] The spatial performance of the S4 microseismic spatiotemporal characteristic index ρ reflects the correlation between data points in the microseismic dataset and their degree of clustering. When high-energy events are located in high-probability regions, it indicates that the high-energy events are closely related to the microseismic dataset M'. j The microseismic data in the middle show a strong correlation.

[0147] To fully utilize the impact of long-term and short-term microseismic data on spatial prediction, this embodiment constructs spatial prediction cloud maps of hazardous areas for short-term time windows T1 and long-term time windows T2 using the microseismic spatiotemporal characteristic index ρ in the spatial prediction module. The spatial prediction cloud maps of hazardous areas for both long and short time windows are then fused using weighted superposition to construct a shock hazard area cloud map. Based on the hazardous areas in the shock hazard area cloud map, the spatial prediction method for rockburst is completed. The specific method is as follows:

[0148] S4.1 Data for short time window T1 and long time window T2 are constructed using the microseismic spatiotemporal characteristic index ρ;

[0149] S4.2 The working surface coordinates are gridded, and the normalized short-time window T1 and long-time window T2 microseismic spatiotemporal characteristic index ρ is filled in. The weights of the T1 and T2 time window data are calculated and determined respectively. Kriging interpolation was used to fill each grid to obtain grid matrices E1 and E2 for the spatiotemporal characteristics of microseismic indicators in the short-time window T1 and long-time window T2, thus obtaining spatial prediction cloud maps of the danger zones for T1 and T2. A schematic diagram of the effect of the short-time window T1 is shown below. Figure 4 As shown in (a), the effect of the long-term window T2 is illustrated in the diagram. Figure 4 As shown in (b);

[0150] The method for determining the weights of the data in time windows T1 and T2 is as follows: Microseismic data from time windows T1 and T2 are selected. The spatiotemporal characteristic index ρ corresponding to each microseismic data point in T1 and T2 is calculated. Then, the spatiotemporal characteristic indices ρ of T1 and T2 are normalized, and the variance of each characteristic index is calculated. In this process, it is assumed that the larger the variance of the index, the more information it conveys, and the greater its corresponding weight. Simultaneously, the amount of data within each unit time window is statistically analyzed. In this process, it is assumed that the larger the amount of data within a unit time window, the more information the index contains, and the greater its corresponding weight. Considering both weights, the weight values ​​for the data in time windows T1 and T2 are determined respectively, using the following formula:

[0151]

[0152]

[0153] in, The variances of the data for time windows T1 and T2 are respectively; These represent the number of microseisms per unit time window for time windows T1 and T2, respectively.

[0154] S4.3 Because spatial prediction cloud maps of impact hazard areas based on short-term windows can observe a large spatial range, they cannot effectively reflect the actual hazard area. Conversely, spatial prediction cloud maps of impact hazard areas based on long-term windows are difficult to effectively reveal the microseismic characteristics in the short term, i.e., the characteristics contained in microseismic events. Microseismic characteristics include the degree of microseismic concentration, energy magnitude, frequency, etc. Therefore, after superimposing grid matrices E1 and E2 according to weighted superposition, a microseismic spatiotemporal characteristic index grid matrix E3 with fused short and long-term windows is obtained. Impact hazard area cloud maps are then drawn based on grid matrix E3, as shown below. Figure 4 As shown in (c), the danger zone is determined based on the cloud map, completing the construction of the spatial prediction method. This method can effectively reflect the actual danger zone and effectively reveal the microseismic characteristics in the short term. The expression for the grid matrix E3 is as follows:

[0155]

[0156] The specific testing of the spatial prediction method for rockbursts in this embodiment is as follows:

[0157] This embodiment of the spatial prediction method selects 13 high-energy data points from the 2215 working face in March. The prediction duration N and time window I are both set to 6 hours, consistent with the temporal prediction method, thus demonstrating the application of spatial prediction in engineering. Through statistical analysis of these 13 high-energy events, T1 and T2 are set to 6 hours and 3 days respectively. These two sets of grid data are then overlaid to obtain a spatial cloud map, as shown below. Figure 15 The image shows full-time and spatiotemporal prediction cloud maps of three high-energy events. In the image, the pentagram positions indicate the locations of the high-energy events.

[0158] like Figure 15 The image shown displays the spatiotemporal prediction results for high-energy events with strong time prediction capabilities. Figure 15 This directly reflects the temporal dimension of microseismic event prediction and the spatial division of predicted hazardous areas within the time window, enabling the construction of a full-time and spatiotemporal prediction method. Simultaneously, based on the full-time and spatiotemporal prediction results of 13 high-energy events, the statistical spatial prediction and early warning results are displayed as follows: Figure 16 As shown.

[0159] according to Figure 16 It can be seen that among the 13 high-energy events predicted, 6 were located in areas with strong rockburst hazards, 3 were located in areas with moderate rockburst hazards, and 4 were located in areas with weak rockburst hazards. This shows that the spatial prediction method of the present invention is effective and can meet the needs of engineering applications.

[0160] In summary, by combining time-based prediction methods with spatial cloud maps, a full-time and spatial prediction method for rockburst based on the fusion of the "time-space-intensity" characteristics of microseisms was developed. This method has reference value for early prediction of high-energy events and for taking preventive measures in advance.

[0161] This invention presents a spatiotemporal prediction method for rockburst based on the fusion of microseismic temporal-spatial-intensity features. This method comprises three modules: microseismic spatiotemporal feature indices, time prediction, and spatial prediction. In the microseismic spatiotemporal feature indices module, this invention establishes indices that simultaneously consider both temporal and spatial information from microseismic data, reflecting the clustering of spatiotemporal features. These indices form the basis for both time and spatial prediction in this invention. This invention constructs a rockburst time prediction model based on a deep recurrent neural network (LSTM), using a gate control unit within the LSTM to control the transmission of time information and utilizing a fully connected network's hazard classification module to achieve accurate time prediction of rockburst. In the spatial prediction module, this invention employs a spatial prediction method for rockburst that fuses long-term and short-term time windows. This method analyzes both the long-term global characteristics of data from long-term windows and the short-term local characteristics of data from short-term windows, thus achieving spatial prediction of rockburst. The time and spatial prediction methods of this invention can be used in conjunction to achieve accurate spatiotemporal prediction of rockburst.

[0162] Other components of the all-temporal and spatiotemporal prediction method for rockburst based on the fusion of microseismic "temporal-spatial-intense" features according to embodiments of the present invention, such as deep recurrent neural networks (LSTM), backpropagation methods, and operations, are known to those skilled in the art and will not be described in detail here.

[0163] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," "counterclockwise," "axial," "radial," and "circumferential" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0164] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0165] In this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," "linking," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection, an electrical connection, or a communication connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0166] In this invention, unless otherwise explicitly specified and limited, "above" or "below" the second feature can mean that the first feature is in direct contact with the second feature, or that the first feature is in indirect contact with the second feature through an intermediate medium. Furthermore, "above," "over," and "on top" of the second feature can mean that the first feature is directly above or diagonally above the second feature, or simply that the first feature is at a higher horizontal level than the second feature. "Below," "below," and "under" the second feature can mean that the first feature is directly below or diagonally below the second feature, or simply that the first feature is at a lower horizontal level than the second feature.

[0167] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0168] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. A method for predicting rockburst across time and space based on microseismic "time-space-intensity" dynamics, characterized in that... It includes a microseismic spatiotemporal characteristic index module, a time prediction module, and a spatial prediction module. The specific methods are as follows: S1 collects raw microseismic data, including microseismic time, source coordinates, and energy magnitude; In the microseismic spatiotemporal characteristic index module, S2 constructs a time series dataset using microseismic time and source coordinates. Then, the PCA method is used to reduce the dimensionality of the time series dataset, and finally the KDE method is used to process the dimensionality-reduced data to obtain the probability density value of each data point. These probability density values ​​are defined as microseismic spatiotemporal characteristic indicators, which can reflect the degree of data aggregation in this time period. In the time prediction module, S3 is based on the processed time series dataset. Generate a new temporal dataset using microseismic spatiotemporal characteristic indicators. The study used this dataset to construct a precursor pattern sequence of microseismic data. Based on the precursor pattern sequence of microseismic data, an LSTM model was constructed to predict the timing of rockbursts. A fully connected network was used to achieve accurate prediction of rockburst timing. The model was trained and tested to complete the construction of a rockburst timing prediction method, which is used to predict the hazard level of rockbursts in the future. S4 utilizes microseismic spatiotemporal characteristic indicators in the spatial prediction module. Construct short time windows respectively and long window The spatial prediction cloud map of dangerous areas is used to fuse the spatial prediction cloud maps of dangerous areas with long and short time windows by weight superposition to construct the shock hazard area cloud map. Based on the dangerous areas in the cloud map, the spatial prediction of rockburst is realized, and the construction of the spatial prediction method of rockburst is completed. Step S4 specifically includes the following steps: S4.1 Utilizing the spatiotemporal characteristics of microseismic indices Construct short time windows respectively and long window Data; S4.2 Mesh the working surface coordinates and fill them with the normalized short time window. and long window Microseismic Spatiotemporal Characteristic Indicators Calculate and determine respectively and Weight of time window data , Short time windows are obtained by filling each grid using Kriging interpolation. and long window Grid matrix of microseismic spatiotemporal characteristic indicators and Thus obtain and The spatial prediction cloud map of the danger zone, in which... , The representation is: （14） （15） In the formula, , They are respectively and Variance of data within a time window; , They are respectively and The number of micro-seismic events per unit time window; S4.3 Based on the weight superposition, the grid matrix is... and After superposition, a grid matrix of microseismic spatiotemporal characteristic indicators fused from long and short time windows is obtained. According to the grid matrix A cloud map of the rockburst hazard area is drawn, and the hazard area is determined based on the cloud map, achieving accurate spatial prediction of rockbursts and completing the construction of a spatial prediction method for rockbursts. This includes the use of a grid matrix. Represented as: （13）。 2. The method for predicting rockburst in all time and space based on microseismic "time-space-intensity" according to claim 1, characterized in that, The microseismic spatiotemporal characteristic indicators in step S2 Obtained through the following methods: S2.1 A time series dataset is constructed using microseismic time and source coordinates, as detailed below: Microseismic data were statistically analyzed using a fixed time window, and a microseismic dataset was constructed. , among which, the The microseismic dataset is , represented as: (1) In the formula, The first Microseismic dataset The microseismic data includes the source coordinates X, Y, and Z of the earthquake and the microseismic time. Let the first... Microseismic data records for each time window are as follows The number of microseismic data is n, represented as: (2) When the microseismic dataset m is divided into By dividing the time window, a time series dataset is obtained. , represented as: (3) S2.2 PCA is used to analyze the microseismic dataset from step S2.

1. The data in the dataset was dimensionality reduced from 4 dimensions to 2 dimensions to obtain a microseismic dataset. The combination of the first Microseismic dataset within a reduced time window , represented as: (4) S2.3 Using KDE to process microseismic datasets To obtain the microseismic dataset The probability density corresponding to each data point The formula for KDE is expressed as: （5） in, is the probability density estimate at the point; u is the number of data points in the sample. It is the i-th sample data point; It is a kernel function, representing the local probability density of each data point; It is a bandwidth parameter that controls the width of the kernel function and affects the smoothness of the estimation; S2.4 Microseismic Data Set probability density As a spatiotemporal characteristic index of microseismic events , represented as: （6）。 3. The method for predicting rockburst in all time and space based on microseismic "time-space-intensity" according to claim 2, characterized in that, Step S3 specifically includes the following steps: S3.1 Data Preprocessing: Preprocessing the time series dataset... Statistical analysis was performed to obtain the frequency of maximum energy, average energy, and microseismic events in the time series, combined with spatiotemporal characteristic indicators of microseismic events. Generate new time series datasets And use this dataset to construct a precursor pattern sequence; S3.2 Constructing a rockburst time prediction model: LSTM is used to model and extract features from precursor pattern sequences to obtain regression loss and data features. The parameters in the LSTM model are continuously updated through backpropagation to minimize the loss of the model on the training dataset. A fully connected network is used to achieve accurate prediction of rockburst time, thus obtaining the rockburst time prediction model. S3.3 Training and testing of the rockburst time prediction model.

4. The method for predicting rockburst in all time and space based on microseismic "time-space-intensity" according to claim 3, characterized in that, The specific method for step S3.1 is as follows: Statistical analysis of time series datasets The Middle Data from a time window can yield the following data records: , represented as: (7) In the formula: Number the time window. This represents the maximum energy of the micro-vibrations within the time window. The average energy of micro-vibrations within the time window. The frequency of micro-vibrations within the time window. Spatiotemporal characteristics of microseisms within a time window The maximum value in the microseismic dataset; Classified as By dividing the time window, a time series dataset is obtained. Represented as: (8) Furthermore, based on the time window Construct a set of labels : (9) in, Labels representing the corresponding time windows; Based on time series datasets Construct a precursor pattern sequence, where the first... The precursor pattern sequence is as follows: Its expression is: (10) In the formula, For the sampling step, Given the length of the precursor pattern sequence, a sequence containing... Group precursor model sequence It is represented as: (11)。 5. The method for predicting rockburst in all time and space based on microseismic "time-space-intensity" according to claim 3, characterized in that, The specific method for training the rockburst time prediction model in step S3.3 is as follows: S3.3.1 Collect sample dataset: Collect microseismic data for a certain time period as a sample dataset, and divide the dataset into training set and test set according to a set ratio; S3.3.2 During training, the precursor pattern sequence will be used. As input, Adam is used as the model optimizer, classification cross-entropy is used as the loss function, accuracy and F1 score are used as evaluation functions, and activation function is used. Normalization is performed to obtain the probability of high-energy events occurring, thus outputting the magnitude of future energy levels. The classification cross-entropy loss function is used in this process. and activation function The expression is as follows: (16) (12) In the formula, For the sample to belong to the first The probability of a class These are sample labels. It is the total sample size; For the first The output value of each node is obtained through The function can convert the output values ​​of a multi-class classification task to the range [0,1].

Images