Long and short term intelligent prediction and early warning method for microseismic frequency and energy of rock burst mine

Abstract

The long-term and short-term intelligent prediction and early warning method for microseismic frequency and energy of rock burst mine comprises the following steps: 1. Importing microseismic event frequency and energy historical data; 2. Preprocessing the data of step 1; 3. Making a training set and a test set; 4. Building a microseismic prediction model; 5. Inputting the training set data into the prediction model for training; 6. Inputting the test set data into the prediction model for prediction, and recording the evaluation index result value; 7. Using the grid search method to determine the model hyperparameters; 8. Selecting the optimal hyperparameter combination of the evaluation index to the model of step 4; 9. Inputting the historical microseismic frequency and energy data into the model to realize short-term prediction; 10. Continuously adding the prediction results to the historical microseismic data to realize long-term prediction; and 11. Comparing the prediction value with the critical index to realize early warning. The present application can realize long-term and short-term intelligent prediction and early warning of microseismic frequency and energy through the above steps, has the advantages of intelligent prediction and early warning, solves the problems of poor generalization and poor dynamic prediction of traditional mathematical and physical models, and effectively improves the accuracy and timeliness of the intelligent prediction and early warning method for microseismic frequency and energy of rock burst mine.

Description

Technical Field

[0001] The technical field of this invention is the prediction of microseismic events in coal mine safety, specifically a method for intelligent prediction and early warning of the frequency and energy of microseismic events in mines prone to rockbursts, both in the short and long term. Background Technology

[0002] With the increase in coal mining intensity and depth, mining causes instability in the underground rock mass structure, leading to frequent high-energy microseismic events or mine tremors in mines prone to rockbursts. In severe cases, these events can trigger coal and rock dynamic disasters such as rockbursts, causing damage to mine equipment, casualties, and other significant property losses. Currently, the main regional monitoring method for rockburst mines is microseismic monitoring. Most mines use statistical indicators such as the frequency and energy of microseismic events as early warning indicators. At present, most microseismic anomalies are only discovered after a rockburst accident has occurred. Limited by current theories, accurate prediction and early warning are difficult to achieve, thus failing to guarantee the timeliness of microseismic event prediction. This often leads to delays in rockburst warnings and missing the optimal opportunity to implement mitigation measures. Microseismic monitoring data in rockburst mines carries a wealth of useful information; however, traditional prediction methods mostly combine mine geological conditions with comparisons between monitoring data and critical indicators for early warning. This results in poor timeliness, insufficient utilization of microseismic data, and an inability to uncover patterns in historical microseismic data, leading to a waste of microseismic data. Underground coal mining conditions are complex, influenced by various factors such as mining disturbance, roof fracture, fault slippage, and mining speed. Therefore, scientifically and accurately predicting and providing early warning of rockburst disasters is an urgent task. This invention, based on historical microseismic monitoring data and incorporating machine learning algorithms, constructs long-term and short-term prediction and early warning models for the frequency and energy of microseismic events in rockburst-prone mines. This effectively improves the accuracy and timeliness of intelligent prediction and early warning methods for microseismic frequency and energy in rockburst-prone mines. Summary of the Invention

[0003] This invention proposes a method for long-term and short-term intelligent prediction of microseismic frequency and energy based on deep learning to mine historical microseismic data.

[0004] This invention is achieved through the following technical solution:

[0005] A short- and long-term intelligent prediction method for the frequency and energy of micro-seismic events in mines prone to rockbursts, comprising the following steps:

[0006] 1) Based on the mining and microseismic monitoring of the underground working face in the coal mine, import historical microseismic data, including the time of occurrence and energy of the microseismic events;

[0007] 2) Preprocess historical microseismic data;

[0008] 3) Use the processed historical microseismic data to create a training dataset and a test dataset for microseismic events;

[0009] 4) Combine SG filters and TSA-GRU neural network units to construct an SG-TSA-GRU network microseismic event frequency and energy prediction model;

[0010] 5) Input the microseismic event training set data into the SG-TSA-GRU network to train the model;

[0011] 6) Input the microseismic event test set data into the trained model and record the results of the evaluation indicators;

[0012] 7) Use grid search to traverse the network hyperparameters and repeat steps 5) and 6): Use exhaustive search to list the values ​​of each hyperparameter and feed them into the model for iterative training. Save the model prediction results evaluation index, loss curve, and comparison chart of the actual and predicted values.

[0013] 8) Select the optimal hyperparameters of the network model based on the evaluation index: Select the hyperparameters corresponding to the minimum RMSE value among the evaluation index results as the optimal hyperparameters of the model;

[0014] 9) The trained model is used for short-term prediction of microseismic frequency and energy, mainly to predict the frequency and energy of microseismic events the following day based on historical data;

[0015] 10) Using the predicted data and historical data obtained in step 9) as input, repeat the process of step 9) to predict the frequency and energy of micro-seismic events on the second day, and so on to achieve long-term prediction of micro-seismic frequency and energy.

[0016] In step 2), the specific method is as follows:

[0017] 2.1) Preprocess the microseismic data to obtain microseismic frequency time series data;

[0018] 2.1.1) Aggregate the data by date to sum the number of micro-seismic events that occurred within a day to obtain time series data of micro-seismic frequency within a day;

[0019] 2.1.2) For missing days in consecutive dates, interpolation is used to fill in the missing data. Assume the mapping relationship between microseismic frequency and time is y = f(t), where y represents the time series value of the microseismic frequency and t represents the time index. The filling method is as follows:

[0020]

[0021]

[0022] Where y is the missing value of the microseismic energy level to be filled, y i It represents the n values ​​within a sliding window surrounding the missing value y, where t is the time index of the missing value to be filled. i and t kIt is the time index of the values ​​within the sliding window surrounding the missing value y.

[0023] 2.2) Preprocessing the microseismic data yields microseismic energy level time series data;

[0024] 2.2.1) Aggregate the data by date to obtain the time series data of the total energy of micro-seismic events within a day by superimposing the energy of each micro-seismic event within a day;

[0025]

[0026] Where n is the number of micro-earthquakes occurring in a day, e i The energy released in each micro-seismic event.

[0027] 2.2.2) For missing days in consecutive dates, interpolation is used to fill in the missing data. Assume the mapping relationship between microseismic energy level and time is y = f(t), where y represents the time series value of the microseismic energy level and t represents the time index. The filling method is as follows:

[0028]

[0029]

[0030] Where y is the missing value of the microseismic energy level to be filled, y i It represents the n values ​​within a sliding window surrounding the missing value y, where t is the time index of the missing value to be filled. i and t k It is the time index of the values ​​within the sliding window surrounding the missing value y.

[0031] 2.2.3) To enhance the autocorrelation of microseismic data, the daily total energy time series is logarithmically divided to base 10 to form a microseismic energy level time series;

[0032] grade = log 10 energy (6)

[0033] Where energy is the microseismic energy time series and grade is the microseismic energy level time series.

[0034] In step 3), the specific method is as follows:

[0035] 3.1) Normalize the microseismic frequency and microseismic energy level sequences in the time series using the maximum-minimum normalization method, as shown in the formula:

[0036]

[0037] Where x max x is the maximum value of the sample data. min Let x be the minimum value of the sample data, and x be the value to be normalized.

[0038] 3.2) Construct the correspondence between the precursor sequences X and future values ​​Y of microseismic events for the time series of microseismic frequencies and energy levels, respectively, to obtain a supervised dataset. The construction rules are as follows:

[0039] X n ={x n x n+1 x n+2 , ..., x time_step+n-1} (8)

[0040] Y n =x time_step+n (9)

[0041] Where X n Y represents the nth model input data. n x represents the label corresponding to the nth model input data. n This represents the value of the time series data of the microseismic event on day n, where time_step is a fixed constant;

[0042] 3.3) Use the first 70% of the constructed input and output data as the microseismic training dataset and the last 30% as the microseismic test dataset.

[0043] The specific method in step 4) is as follows:

[0044] SG-TSA-GRU network model structure: SG filtering is used for data smoothing. The input is the X values ​​from the microseismic frequency dataset and microseismic energy level dataset obtained in step 3). n The output is X n The smoothed data; the TSA-GRU neural network unit is used to extract data features. The input is the data smoothed by SG filtering, and the output is the prediction result; the final prediction result needs to be compared with Y in step 3). n For comparison, the Adam optimization algorithm is used to backpropagate and update the TSA-GRU weight matrix;

[0045] The Savitzky-Golay filter is a filtering method based on local polynomial least squares fitting in the time domain. The fitting polynomial formula is:

[0046]

[0047]

[0048] Where α k ε represents the constant term of the fitted polynomial. N This represents the residual of the least squares fit;

[0049] The GRU neural network unit contains two gate structures: the update gate z. t and reset door r t The update gate controls the extent to which historical microseismic data from the previous moment is incorporated into the current state. A larger update gate value indicates that less historical microseismic data from the previous moment is retained, and more historical microseismic data from the current moment is retained. The update gate z... t and reset door r t The formula is:

[0050] z t =σ(W z ·[h t-1 x t (12)

[0051] r t =σ(W r ·[h t-1 x t (13)

[0052] Where x t It is the input at the current moment, h t-1 It is the hidden state from the previous moment, W z To update the gate parameter matrix to be trained, W r To reset the training parameter matrix of the gate, σ is the sigmoid activation function;

[0053] When calculating the hidden state at the current moment, it first calculates a candidate state. When calculating candidate states, the value of the reset gate is considered. The formula is:

[0054]

[0055] Where W is the parameter matrix to be trained, r t It resets the gate output, x t It is the input at the current moment, h t-1 It is the output from the previous moment;

[0056] Finally, the output formula of the network model at the current time step is obtained as follows:

[0057]

[0058] Where z t To update the gate output, h t-1 The output from the previous time step, z t To update the gate output, Candidate state;

[0059] Time Series Attention (TSA) is a deep learning model that excels at extracting temporal features. The weighted summation formula for Time Series Attention is:

[0060]

[0061] Where b j For time-attention outputs, j represents the number of outputs, which is consistent with the time step time_step, and a j，i Indicates attention weight allocation, v i This represents the result of multiplying the input by the weight matrix to be trained.

[0062] The specific method in step 5) is as follows:

[0063] The input portion of the training set is fed into the SG-TSA-GRU network model, and the output portion of the training set is used as the model's output. The SG-TSA-GRU network model continuously learns the mapping relationship between the input and output through an optimization algorithm, updating the weight matrix. The objective function loss is the mean squared error (MSE), which is calculated as follows:

[0064]

[0065] Where y(x) represents the true value, a L (x) represents the predicted output value of the neuron;

[0066] After a certain number of iterations, the loss decreases slowly. If the output loss value is within the specified error σ for ten consecutive iterations, the model is considered to have completed training. The specific training process is as follows:

[0067] The model training uses the Adam optimization algorithm to update network weights and optimize the model learning rate; the network weight update formula is as follows:

[0068]

[0069] Where t represents the number of iterations, α represents the learning rate, and w t Represents network weights, For m t The correction, For v t The correction is as follows; the specific correction formula is as follows:

[0070]

[0071]

[0072] Where β1 and β2 are constants, controlling exponential decay, m tIt is the exponential moving average of the gradient, obtained by using the first moment of the gradient; v t It is the squared gradient, obtained through the second moment of the gradient; m t and v t The update is as follows:

[0073] m t =β1*m t-1 +(1-β1)*g t (twenty one)

[0074]

[0075] Where g t It is the first derivative.

[0076] The specific method in step 6) is as follows:

[0077] Select a suitable evaluation metric for regression prediction problems: root mean square error;

[0078] The root mean square error (RMSE) is calculated by summing the squares of the differences between the true and predicted values. It primarily measures the deviation between the true and predicted values; a smaller RMSE value indicates higher model accuracy. The calculation formula is as follows:

[0079]

[0080] Where m is the total number of samples, y i , These are the actual and predicted values ​​of the total daily microseismic energy, respectively.

[0081] The specific method in step 9) is as follows:

[0082] 9.1) Short-term prediction of microseismic frequency: Load the trained model, input a set of historical microseismic frequency time series data with a time step of time_step, and the model predicts the microseismic frequency for time_step+1 days;

[0083] 9.2) Short-term prediction of microseismic energy levels: Load the trained model, input a set of historical microseismic energy level time series data with a time step of time_step, and the model predicts the microseismic energy level for time_step+1 days.

[0084] The specific method in step 10) is as follows:

[0085] 10.1) Long-term prediction of microseismic frequencies:

[0086] 10.1.1) Load the trained model, input a set of historical microseismic frequency time series data with a time step of time_step days, and the model predicts the microseismic frequency in time_step+1 days;

[0087] 10.1.2) Add the predicted microseismic frequencies to the historical microseismic frequency time series, move the sliding window forward one unit, and use it as the input for the new microseismic time series to continue predicting the microseismic frequencies for time_step+2 days;

[0088] 10.1.3) Repeat step 10.1.2 to achieve long-term prediction of microseismic frequencies;

[0089] 10.2) Long-term prediction of microseismic energy levels:

[0090] 10.2.1) Load the trained model, input a set of historical microseismic energy level time series data with a time step of time_step days, and the model predicts the microseismic energy level for time_step+1 days;

[0091] 10.2.2) Add the predicted microseismic energy levels to the historical microseismic energy level time series, move the sliding window forward one unit, and use it as the input for the new microseismic time series to continue predicting the microseismic energy levels for time_step+2 days;

[0092] 10.2.3) Repeat step 10.2.2 to achieve long-term prediction of microseismic energy levels.

[0093] The method for early warning using the aforementioned prediction method compares the predicted value of this invention with the early warning index based on the micro-vibration frequency and energy warning index in the coal mine working face anti-impact design report. When the predicted value is greater than the critical index, an audible and visual alarm is triggered: when the predicted value is between 1.0 and 1.2 times the critical index, it is set as a Level 1 blue warning; when the predicted value is between 1.2 and 1.5 times the critical index, it is set as a Level 2 yellow warning; and when the predicted value is greater than 1.5 times the critical index, it is set as a Level 3 red warning.

[0094] The beneficial effects of this invention are:

[0095] 1. This invention quantitatively predicts the frequency and energy of microseismic events using a deep learning regression model. It addresses the common problem of lacking readily available field monitoring parameters and the difficulty in quantifying the mapping relationship between these parameters and the frequency and energy of disaster events.

[0096] 2. The prediction method takes into account both long-term and short-term predictions, and provides graded early warnings for microseismic and rockburst disasters based on different prediction results.

[0097] 3. The prediction method can train prediction models with different knowledge based on massive microseismic data to make quantitative predictions and early warnings, which solves the problems of poor generalization and poor dynamics of traditional mathematical physics models. Attached Figure Description

[0098] Figure 1 This is a flowchart of the present invention;

[0099] Figure 2 Here is a diagram of the SG-TSA-GRU network model structure;

[0100] Figure 3 A metric for evaluating the total frequency of microseismic events under different hyperparameters;

[0101] Figure 4 Comparison of the actual and predicted total frequency of micro-seismic events at a mine working face over the next day;

[0102] Figure 5 A comprehensive energy level evaluation index for microseismic events under different hyperparameters;

[0103] Figure 6 Comparison of the actual and predicted total energy levels of microseismic events at a mine working face over the next day. Detailed Implementation

[0104] This invention provides a method for intelligent prediction and early warning of the frequency and energy of micro-seismic events in mines prone to rockbursts, characterized by the following steps:

[0105] 1) Based on the mining and microseismic monitoring of the underground working face in the coal mine, import historical microseismic data, including the time of occurrence: year, month, day, hour, minute, second, and microseismic energy;

[0106] When a microseismic event occurs, the rock mass fractures, energy is released, and seismic waves are generated. After multiple sensors collect the seismic waves, the program calculates the time of occurrence of the earthquake (year, month, day, hour, minute, second) and the time series data of the microseismic energy.

[0107] 2) Preprocess historical microseismic data;

[0108] 2.1) Preprocess the microseismic data to obtain microseismic frequency time series data;

[0109] 2.1.1) Aggregate the data by date to sum the number of micro-seismic events that occurred within a day to obtain time series data of micro-seismic frequency within a day;

[0110] 2.1.2) For missing days in consecutive dates, interpolation is used to fill in the missing data. Assume the mapping relationship between microseismic frequency and time is y = f(t), where y represents the time series value of the microseismic frequency and t represents the time index. The filling method is as follows:

[0111]

[0112]

[0113] Where y is the missing value of the microseismic energy level to be filled, y i It represents the n values ​​within a sliding window surrounding the missing value y, where t is the time index of the missing value to be filled. i and t kIt is the time index of the values ​​within the sliding window surrounding the missing value y.

[0114] 2.2) Preprocessing the microseismic data yields microseismic energy level time series data;

[0115] 2.2.1) Aggregate the data by date to obtain the time series data of the total energy of micro-seismic events within a day by superimposing the energy of each micro-seismic event within a day;

[0116]

[0117] Where n is the number of micro-earthquakes occurring in a day, e i The energy released in each micro-seismic event.

[0118] 2.2.2) For missing days in consecutive dates, interpolation is used to fill in the missing data. Assume the mapping relationship between microseismic energy level and time is y = f(t), where y represents the time series value of the microseismic energy level and t represents the time index. The filling method is as follows:

[0119]

[0120]

[0121] Where y is the missing value of the microseismic energy level to be filled, y i It represents the n values ​​within a sliding window surrounding the missing value y, where t is the time index of the missing value to be filled. i and t k It is the time index of the values ​​within the sliding window surrounding the missing value y.

[0122] 2.2.3) To enhance the autocorrelation of microseismic data, the daily total energy time series is logarithmically divided to base 10 to form a microseismic energy level time series;

[0123] grade = log 10 energy (6)

[0124] Where energy is the microseismic energy time series and grade is the microseismic energy level time series.

[0125] 3) Use the processed historical microseismic data to create a training dataset and a test dataset for microseismic events;

[0126] 3.1) Normalize the microseismic frequency and microseismic energy level sequences in the time series using the maximum-minimum normalization method, as shown in the formula:

[0127]

[0128] Where x max x is the maximum value of the sample data. minLet x be the minimum value of the sample data, and let x be the value to be normalized.

[0129] 3.2) Construct the correspondence between the precursor sequences X and future values ​​Y of microseismic events for the time series of microseismic frequencies and energy levels, respectively, to obtain a supervised dataset. The construction rules are as follows:

[0130] X n ={x n x n+1 x n+2 , ..., x time_step+n-1} (8)

[0131] Y n =x time_step+n (9)

[0132] Where X n Y represents the nth model input data. n x represents the label corresponding to the nth model input data. n This represents the value of the time series data of the microseismic event on day n, where time_step is a fixed constant.

[0133] 3.3) Use the first 70% of the constructed input and output data as the microseismic training dataset and the last 30% as the microseismic test dataset;

[0134] 4) Combine SG filters and TSA-GRU neural network units to construct an SG-TSA-GRU network microseismic event frequency and energy prediction model;

[0135] The structure diagram of the SG-TSA-GRU network model is as follows: Figure 2 As shown: SG filtering is used for data smoothing. The input is the X values ​​from the microseismic frequency dataset and the microseismic energy level dataset obtained in step 3. n The output is X n The smoothed data. TSA-GRU neural network units are used to extract data features; the input is the data smoothed by SG filtering, and the output is the prediction result. The final prediction result needs to be compared with Y in step 3. n For comparison, the Adam optimization algorithm is used to backpropagate and update the TSA-GRU weight matrix.

[0136] The Savitzky-Golay filter is a filtering method based on local polynomial least squares fitting in the time domain. The fitting polynomial formula is:

[0137]

[0138]

[0139] Where αk ε represents the constant term of the fitted polynomial. N This represents the residual of the least squares fit.

[0140] The GRU neural network unit contains two gate structures: the update gate z. t and reset door r t The update gate controls the extent to which historical microseismic data from the previous moment is incorporated into the current state. A larger update gate value indicates that less historical microseismic data from the previous moment is retained, while more historical microseismic data from the current moment is retained. Update gate z t and reset door r t The formula is:

[0141] z t =σ(W z ·[h t-1 x t (12)

[0142] r t =σ(W r ·[h t-1 x t (13)

[0143] Where x t It is the input at the current moment, h t-1 It is the hidden state from the previous moment, W z To update the gate parameter matrix to be trained, W r The parameter matrix to be trained is to reset the gate, and σ is the sigmoid activation function.

[0144] When calculating the hidden state at the current moment, it first calculates a candidate state. When calculating candidate states, the value of the reset gate is considered. Candidate states The formula is:

[0145]

[0146] Where W is the parameter matrix to be trained, r t It resets the gate output, x t It is the input at the current moment, h t-1 It is the output from the previous moment.

[0147] Finally, the output formula of the network model at the current time step is obtained as follows:

[0148]

[0149] Where z t To update the gate output, h t-1 The output from the previous time step, z tTo update the gate output, It is in the candidate state.

[0150] Time Series Attention (TSA) is a deep learning model that excels at extracting temporal features. The weighted summation formula for Time Series Attention is:

[0151]

[0152] Where b j For time-attention outputs, j represents the number of outputs, which is consistent with the time step time_step, and a j，i Indicates attention weight allocation, v i This represents the result of multiplying the input by the weight matrix to be trained.

[0153] 5) Input the microseismic event training set data into the SG-TSA-GRU network to train the model;

[0154] The input portion of the training set is fed into the SG-TSA-GRU network model, and the output portion of the training set is used as the model's output. The SG-TSA-GRU network model continuously learns the mapping relationship between the input and output through an optimization algorithm, updating the weight matrix. The objective function loss is the mean squared error (MSE), which is calculated as follows:

[0155]

[0156] Where y(x) represents the true value, a L (x) represents the predicted output value of the neuron.

[0157] After a certain number of iterations, the loss decreases slowly. If the output loss value is within the specified error σ for ten consecutive iterations, the model is considered to have completed training. The specific training process is as follows:

[0158] The model training uses the Adam optimization algorithm to update network weights and optimize the model learning rate. The network weight update formula is as follows:

[0159]

[0160] Where t represents the number of iterations, α represents the learning rate, and w t Represents network weights, For m t The correction, For v t The correction is as follows: The specific correction formula is as follows:

[0161]

[0162]

[0163] Where β1 and β2 are constants, controlling exponential decay, m t It is the exponential moving average of the gradient, obtained by using the first moment of the gradient. t It is the squared gradient, obtained through the second moment of the gradient. t and v t The update is as follows:

[0164] m t =β1*m t-1 +(1-β1)*g t (twenty one)

[0165]

[0166] Where g t It is the first derivative.

[0167] 6) Input the microseismic event test set data into the trained model and record the results of the evaluation indicators;

[0168] Select a suitable evaluation index for regression prediction problems: root mean square error.

[0169] The root mean square error (RMSE) is calculated by summing the squares of the differences between the true and predicted values. It primarily measures the deviation between the true and predicted values; a smaller RMSE value indicates higher model accuracy. The calculation formula is as follows:

[0170]

[0171] Where m is the total number of samples, y i , These are the actual and predicted values ​​of the total daily microseismic energy, respectively.

[0172] 7) Use a grid search method to traverse the network hyperparameters and repeat steps 5 and 6;

[0173] Use an exhaustive method to list the values ​​of each hyperparameter, and feed them into the model for iterative training in sequence. Save the evaluation index values ​​of the model prediction results, the loss curve, and the comparison chart between the actual values ​​and the predicted values.

[0174] 8) Select the optimal hyperparameters of the network model based on the evaluation indicators;

[0175] The hyperparameter corresponding to the minimum RMSE value among the evaluation index results is selected as the optimal hyperparameter of the model.

[0176] 9) The trained model will be used for short-term prediction of microseismic frequency and energy, mainly to predict the frequency and energy of microseismic events the following day based on historical data.

[0177] 9.1) Short-term prediction of microseismic frequency

[0178] Load the trained model, input a set of historical microseismic frequency time series data with a time step of time_step, and the model will predict the microseismic frequency for time_step+1 days.

[0179] 9.2) Short-term prediction of microseismic energy levels

[0180] Load the trained model, input a set of historical microseismic energy level time series data with a time step of time_step, and the model will predict the microseismic energy level for time_step+1 days.

[0181] 10) Using the predicted data and historical data obtained in step 9 as input, repeat the process of step 9 to predict the frequency and energy of micro-earthquakes on the second day, and so on to achieve long-term prediction of the frequency and energy of micro-earthquakes.

[0182] 10.1) Long-term prediction of microseismic frequencies;

[0183] 10.1.1) Load the trained model, input a set of historical microseismic frequency time series data with a time step of time_step days, and the model predicts the microseismic frequency in time_step+1 days;

[0184] 10.1.2) Add the predicted microseismic frequencies to the historical microseismic frequency time series, move the sliding window forward one unit, and use it as the input for the new microseismic time series to continue predicting the microseismic frequencies for time_step+2 days;

[0185] 10.1.3) Repeat step 10.1.2 to achieve long-term prediction of microseismic frequencies.

[0186] 10.2) Long-term prediction of microseismic energy levels;

[0187] 10.2.1) Load the trained model, input a set of historical microseismic energy level time series data with a time step of time_step days, and the model predicts the microseismic energy level for time_step+1 days;

[0188] 10.2.2) Add the predicted microseismic energy levels to the historical microseismic energy level time series, move the sliding window forward one unit, and use it as the input for the new microseismic time series to continue predicting the microseismic energy levels for time_step+2 days;

[0189] 10.2.3) Repeat step 10.2.2 to achieve long-term prediction of microseismic energy levels.

[0190] 11) Based on the micro-vibration frequency and energy warning indicators in the coal mine working face anti-impact design report, compare the predicted value of this invention with the warning indicators. If the predicted value is greater than the critical indicator, an audible and visual alarm will be triggered: when the predicted value is between 1.0 and 1.2 times the critical indicator, it will be set as a Level 1 blue warning; when the predicted value is between 1.2 and 1.5 times the critical indicator, it will be set as a Level 2 yellow warning; when the predicted value is greater than 1.5 times the critical indicator, it will be set as a Level 3 red warning.

[0191] Example 1:

[0192] A coal mine has a mining depth of nearly 1000m. During the underground mining process, several high-energy microseismic events occurred, with noticeable tremors felt on the surface. Historical microseismic data of working face 105, monitored by the mine's microseismic monitoring system from October 19, 2020 to June 27, 2021, were collected.

[0193] Step S1: Based on the mining and microseismic monitoring data of the underground coal mine face, import historical microseismic data, including the time of occurrence: year, month, day, hour, minute, second, and microseismic energy;

[0194] When a microseismic event occurs, the rock mass fractures, energy is released, and seismic waves are generated. After multiple sensors collect the seismic waves, the program calculates the time of occurrence of the earthquake: year, month, day, hour, minute, second, and time series data of the microseismic energy.

[0195] Step S2: Preprocess the historical microseismic data;

[0196] 2.1) Preprocess the microseismic data to obtain microseismic frequency time series data;

[0197] 2.1.1) Aggregate the data by date to sum the number of micro-seismic events that occurred within a day to obtain time series data of micro-seismic frequency within a day;

[0198] 2.1.2) Missing days in consecutive dates are filled using equations (1) and (2);

[0199] There are missing data for the 105 working face of the coal mine on March 5, 2021 and March 26, 2021. In equations (1) and (2), n is selected as 4, and the missing values ​​33 and 30 are calculated respectively.

[0200] 2.2) Preprocessing the microseismic data yields microseismic energy level time series data;

[0201] 2.2.1) Aggregate the data by date and use formula (3) to sum the energy of each micro-earthquake that occurred within a day to obtain the time series data of the total energy of micro-earthquakes within a day;

[0202] 2.2.2) Missing days in consecutive dates are filled using equations (4) and (5);

[0203] The 105 working face of the coal mine had missing data on March 5, 2021 and March 26, 2021. In equations (4) and (5), n was selected as 4, and the missing values ​​were calculated as 2.339552704036468 and 3.5790547892170883, respectively.

[0204] 2.2.3) In order to enhance the autocorrelation of microseismic data, the daily total energy time series is taken as the logarithm to the base 10, and the microseismic energy level time series is formed using equation (6);

[0205] Step S3: Use the processed microseismic historical data to create a microseismic event training dataset and a microseismic event test dataset;

[0206] 3.1) Using Equation (7), the microseismic frequency and microseismic energy level columns in the time series are normalized using the maximum-minimum normalization method;

[0207] 3.2) Construct the correspondence between the precursor sequence X and the future value Y of microseismic events for the microseismic frequency and the microseismic energy level time series respectively, and obtain the supervised dataset. The construction rules are Equations (8) and (9).

[0208] 3.3) Use the first 70% of the constructed input and output data as the microseismic training dataset and the last 30% as the microseismic test dataset;

[0209] Step S4: Combine the SG filter and TSA-GRU neural network unit to construct an SG-TSA-GRU network microseismic event frequency and energy prediction model;

[0210] SG filtering is used for data smoothing. The input is the X values ​​from the microseismic frequency dataset and the microseismic energy level dataset obtained in step 3. n The output is X n The smoothed data. TSA-GRU neural network units are used to extract data features; the input is the data smoothed by SG filtering, and the output is the prediction result. The final prediction result needs to be compared with Y in step 3. n For comparison, the Adam optimization algorithm is used to backpropagate and update the TSA-GRU weight matrix.

[0211] Step S5: Input the microseismic event training set data into the SG-TSA-GRU network to train the model;

[0212] The input part of the training set is input into the SG-TSA-GRU network model, and the output part of the training set is used as the output of the model. The SG-TSA-GRU network model continuously learns the mapping relationship between the input and the output through the optimization algorithm, and updates the weight matrix using the objective function (15).

[0213] Step S6: Input the microseismic event test set data into the trained model and record the result values ​​of the evaluation index;

[0214] Step S7: Use a grid search method to traverse the network hyperparameters and repeat steps 5 and 6;

[0215] The values ​​of each hyperparameter are listed using an exhaustive method, and then sequentially fed into the model for iterative training. The evaluation metrics of the model's prediction results, the loss curve, and a comparison graph of the actual and predicted values ​​are saved. The exhaustive range of hyperparameters is:

[0216] window_length = [3, 5, 7, 9];

[0217] k = [1, 2, 3, 4] (the value of k must be less than the value of window_length);

[0218] time_step=[4,8,16,32,64,128];

[0219] batch_size=[4,8,16,32,64,128];

[0220] GRU_units=[4,8,16,32,64,128];

[0221] To reduce the number of iterations, a controlled variable method is used. First, time_step = 16, batch_size = 8, and GRU_units = 32 are fixed, and then the optimal values ​​of window_length and k are searched. After determining the optimal values ​​of window_length and k, the optimal values ​​of time_step, batch_size, and GRU_units are searched again while keeping window_length and k constant.

[0222] Step S8: Select the optimal hyperparameters of the network model based on the evaluation index.

[0223] First, fix time_step = 16, batch_size = 8, and GRU_units = 32. Then, find the optimal values ​​for window_length and k, which are window_length = 3 and k = 1, respectively. After determining the optimal values ​​for window_length and k, fix their values ​​again and find the optimal values ​​for time_step, batch_size, and GRU_units. For the GRU_units parameter, the optimal hyperparameter is GRU_units = 32.

[0224] For microseismic frequency prediction, the heatmap of microseismic frequency evaluation indices under different hyperparameters is shown below. Figure 3The optimal hyperparameters obtained from observing the heat map are time_step = 64 and batch_size = 4. At this time, the root mean square error (RMSE) of the microseismic frequency after comparing the actual value and the predicted value is 3.938.

[0225] For microseismic energy level prediction, the heatmap of microseismic energy level evaluation indices under different hyperparameters is shown below. Figure 6 The optimal hyperparameters obtained from observing the heatmap were time_step = 16 and batch_size = 32. At this point, the root mean square error (RMSE) of the microseismic energy level, after comparing the actual and predicted values, was 0.302.

[0226] Step S9: Use the trained model for short-term prediction of microseismic frequency and energy, mainly to predict the frequency and energy of microseismic events the next day based on historical data.

[0227] 9.1) Short-term prediction of microseismic frequency

[0228] Load the trained model, select the optimal hyperparameters obtained in step S8, use the last 30% of the entire dataset as the test dataset, input the real microseismic frequency data from day 1 to day 64 in the test dataset into the model, and predict the microseismic frequency on day 65.

[0229] Furthermore, the actual microseismic frequency data from day 2 to day 65 in the test dataset were input into the model to predict the microseismic frequency on day 66, and so on. The actual and predicted microseismic frequencies were then compared, resulting in a short-term prediction comparison chart for microseismic frequencies. Figure 4 .

[0230] 9.2) Short-term prediction of microseismic energy levels

[0231] Load the trained model, select the optimal hyperparameters obtained in step S8, use the last 30% of the entire dataset as the test dataset, input the real microseismic energy level data from day 1 to day 16 in the test dataset into the model, and predict the microseismic energy level on day 17.

[0232] Furthermore, the actual microseismic energy level data from days 2 to 17 in the test dataset were input into the model to predict the microseismic energy level on day 18, and so on. The actual and predicted microseismic energy levels were then compared, resulting in a comparison chart of the short-term predicted and actual microseismic energy levels. Figure 6 .

[0233] Step S10: Using the predicted data and historical data obtained in step 9 as input, repeat the process of step 9 to predict the frequency and energy level of micro-earthquakes on the second day, and so on to achieve short-term prediction of the frequency and energy level of micro-earthquakes.

[0234] 10.1) Long-term prediction of microseismic frequencies;

[0235] 10.1.1) Load the trained model, select the optimal hyperparameters obtained in step S8, input a set of historical microseismic frequency time series data with a time step of time_step, and the model predicts the microseismic frequency of time_step+1 days.

[0236] 10.1.2) Add the predicted microseismic frequencies to the historical microseismic frequency time series, move the sliding window forward one unit to form a new microseismic time series, and continue to predict the microseismic frequencies for time_step+2 days.

[0237] 10.1.3) Repeat step 10.1.2 to achieve long-term prediction of microseismic frequencies.

[0238] 10.2) Long-term prediction of microseismic energy levels;

[0239] 10.2.1) Load the trained model, input a set of historical microseismic energy level time series data with a time step of time_step days, and the model predicts the microseismic energy level for time_step+1 days;

[0240] 10.2.2) Add the predicted microseismic energy levels to the historical microseismic energy level time series, move the sliding window forward one unit to form a new microseismic time series, and continue to predict the microseismic energy levels for time_step+2 days;

[0241] 10.2.3) Repeat step 10.2.2 to achieve long-term prediction of microseismic energy levels.

[0242] 11) Based on the micro-seismic frequency and energy warning indicators in the coal mine working face anti-seismic design report, the predicted value of this invention is compared with the warning indicators. In this embodiment, the micro-seismic energy level warning indicator is set to 4. When the predicted value is greater than the critical indicator, an audible and visual alarm is triggered: when the predicted value is between 1.0 and 1.2 times the critical indicator, it is set as a Level 1 blue warning; when the predicted value is between 1.2 and 1.5 times the critical indicator, it is set as a Level 2 yellow warning; and when the predicted value is greater than 1.5 times the critical indicator, it is set as a Level 3 red warning.

[0243] from Figure 6 The prediction results show that the interval between lg(Sum E.) and (4.0, 4.8) should be set as a Level 1 blue alert; the interval between lg(Sum E.) and (4.8, 6.0) should be set as a Level 2 yellow alert; and the interval between lg(Sum E.) and (4.0) and (6.0) should be set as a Level 3 red alert. Figure 6 It can be seen that this method can predict that an event exceeding 10 will occur at the mine on June 11, 2021. 6 A large-energy microseismic event was detected, and a red alert was issued.

Claims

1. A method for intelligent prediction of the frequency and energy of microseismic events in mines prone to rockbursts, characterized in that, The steps are as follows: Step 1) Based on the mining and microseismic monitoring data of the underground coal mine face, import historical microseismic data, including the time of occurrence and energy of the microseismic events; Step 2) Preprocess the historical microseismic data; Step 3) Use the processed microseismic historical data to create a microseismic event training dataset and a microseismic event test dataset; Step 4) Combine the SG filter and the TSA-GRU neural network unit to construct an SG-TSA-GRU network microseismic event frequency and energy prediction model; SG-TSA-GRU network model structure: SG filtering is used for data smoothing. The input is the microseismic frequency dataset and microseismic energy level dataset obtained in step 3). The output is to The smoothed data; the TSA-GRU neural network unit is used to extract data features. The input is the data smoothed by SG filtering, and the output is the prediction result; the final prediction result needs to be compared with the result in step 3). For comparison, the Adam optimization algorithm is used to backpropagate and update the TSA-GRU weight matrix; The Savitzky-Golay filter is a filtering method based on local polynomial least squares fitting in the time domain. The fitting polynomial formula is: ； ； in This represents the constant term of the fitted polynomial. This represents the residual of the least squares fit; The GRU neural network unit contains two gate structures: an update gate and a... and reset door The update gate controls the extent to which historical microseismic data from the previous moment is incorporated into the current state. A larger update gate value indicates that less historical microseismic data from the previous moment is retained, and more historical microseismic data from the current moment is retained. and reset door The formula is: ； ； in It is the input at the current moment. It is the hidden state from the previous moment. To update the gate parameter matrix to be trained, To reset the gate parameter matrix to be trained, for Activation function; When calculating the hidden state at the current moment, it first calculates a candidate state. When calculating candidate states, the value of the reset gate is considered. The formula is: ； in It is the parameter matrix to be trained. It resets the gate output. It is the input at the current moment. It is the output from the previous moment; Finally, the output formula of the network model at the current time step is obtained as follows: ； in To update the gate output, The output from the previous moment. To update the gate output, Candidate state; Temporal attention is a deep learning model that excels at extracting temporal features. The weighted summation formula for temporal attention is: ； in For time-based attention output, This indicates the number of outputs and the time step. Consistent, This indicates the allocation of attention weights. This represents the result of multiplying the input by the weight matrix to be trained; Step 5) Input the microseismic event training set data into the SG-TSA-GRU network to train the model; Step 6) Input the microseismic event test set data into the trained model and record the result values ​​of the evaluation index; Step 7) Use grid search to traverse the network hyperparameters and repeat steps 5) and 6): Use exhaustive search to list the values ​​of each hyperparameter and feed them into the model for iterative training. Save the model prediction results evaluation index, loss curve, and comparison chart of the actual and predicted values. Step 8) Select the optimal hyperparameters of the network model based on the evaluation index: Select the hyperparameters corresponding to the minimum RMSE value among the evaluation index results as the optimal hyperparameters of the model; Step 9) Use the trained model for short-term prediction of microseismic frequency and energy, mainly to predict the frequency and energy of microseismic events the next day based on historical data; Step 10) Using the predicted data and historical data obtained in Step 9) as input, repeat the process of Step 9) to predict the frequency and energy of micro-seismic events the next day, and so on to achieve long-term prediction of micro-seismic frequency and energy.

2. The intelligent prediction method for the frequency and energy of microseismic events in mines prone to rockbursts according to claim 1, characterized in that, In step 2), the specific method is as follows: Step 2.1) Preprocess the microseismic data to obtain microseismic frequency time series data; Step 2.1.1) Aggregate the data by date to sum the number of micro-seismic events that occurred within a day to obtain the time series data of micro-seismic frequency within a day; Step 2.1.2) Fill in the missing days in the consecutive dates using interpolation, assuming the mapping relationship between microseismic frequency and time is as follows: , Represents the time series values ​​of microseismic frequencies. The time index is filled using the following method: ； ； in The missing microseismic energy level values ​​to be filled. Missing values Surrounding sliding window One value, It is the time index of the missing value to be filled. and Missing values The time index of the values ​​within the surrounding sliding window; Step 2.2) Preprocess the microseismic data to obtain microseismic energy level time series data; Step 2.2.1) Aggregate the data by date to obtain the time series data of the total energy of micro-seismic events within a day by superimposing the energy of each micro-seismic event within a day; ； in This represents the number of micro-earthquakes that occur within a day. The energy released in each microseismic event; Step 2.2.2) Fill in the missing days in the continuous dates using interpolation, assuming the mapping relationship between microseismic energy levels and time is as follows: , Represents the time series values ​​of microseismic energy levels. The time index is filled using the following method: ； ； in The missing microseismic energy level values ​​to be filled. Missing values Surrounding sliding window One value, It is the time index of the missing value to be filled. and Missing values The time index of the values ​​within the surrounding sliding window; Step 2.2.3) In order to enhance the autocorrelation of microseismic data, the daily total energy time series is logarithmically divided to base 10 to form a microseismic energy level time series; ； in It is a time series of microseismic energy. It is a time series of microseismic energy levels.

3. The intelligent prediction method for the frequency and energy of microseismic events in mines prone to rockbursts according to claim 1, characterized in that, In step 3), the specific method is as follows: Step 3.1) Normalize the microseismic frequency and microseismic energy level columns in the time series using the maximum-minimum normalization method, as shown in the formula: ； in The maximum value of the sample data. The minimum value of the sample data. These are values ​​to be normalized; Step 3.2) Construct precursor sequences for microseismic events from the time series of microseismic frequencies and energy levels, respectively. With future value Based on the correspondence, a supervised dataset is obtained, and the construction rules are as follows: ； ； in Indicates the first Input data for the model, Indicates the first The labels corresponding to the input data of the model Indicates the first Values ​​of the time series data of the Tianwei earthquake event. It is a fixed constant; Step 3.3) Use the first 70% of the constructed input and output data as the microseismic training dataset and the last 30% as the microseismic test dataset.

4. The intelligent prediction method for the frequency and energy of microseismic events in mines prone to rockbursts according to claim 1, characterized in that, The specific method in step 5) is as follows: The input portion of the training set is fed into the SG-TSA-GRU network model, and the output portion of the training set is used as the model's output. The SG-TSA-GRU network model continuously learns the mapping relationship between the input and output through an optimization algorithm, updating the weight matrix. The objective function loss is the mean squared error (MSE), which is calculated as follows: ； in Represents the actual value. This represents the predicted output value of the neuron. After a certain number of iterations, the loss decreases slowly. If the output loss value is within the specified error for ten consecutive iterations... The model training is considered complete once the internal parameters are checked. The specific training process is as follows: The model training uses the Adam optimization algorithm to update network weights and optimize the model learning rate; the network weight update formula is as follows: ； in Indicates the number of times. Indicates the learning rate. Represents network weights, for The correction, for The correction is as follows: ； ； in and It is a constant, controlling for exponential decay. It is the exponential moving average of the gradient, obtained by using the first moment of the gradient; It is the squared gradient, obtained by using the second moment of the gradient; and The update is as follows: ； ； in It is the first derivative.

5. The intelligent prediction method for the frequency and energy of microseismic events in mines prone to rockbursts according to claim 1, characterized in that, The specific method in step 6) is as follows: Select a suitable evaluation metric for regression prediction problems: root mean square error; The root mean square error (RMSE) is calculated by summing the squares of the differences between the true and predicted values. It primarily measures the deviation between the true and predicted values; a smaller RMSE value indicates higher model accuracy. The calculation formula is as follows: ； in The total number of samples, , These are the actual and predicted values ​​of the total daily microseismic energy, respectively.

6. The intelligent prediction method for the frequency and energy of microseismic events in mines prone to rockbursts according to claim 1, characterized in that, The specific method in step 9) is as follows: Step 9.1) Short-term prediction of microseismic frequency: Load the trained model, input a set of historical microseismic frequency time series data with a time step of time_step, and the model predicts the microseismic frequency for time_step+1 days; Step 9.2) Short-term prediction of microseismic energy level: Load the trained model, input a set of historical microseismic energy level time series data with a time step of time_step, and the model predicts the microseismic energy level for time_step+1 days.

7. The intelligent prediction method for the frequency and energy of micro-seismic events in mines prone to rockbursts according to claim 1, characterized in that, The specific method in step 10) is as follows: Step 10.1) Long-term prediction of microseismic frequencies: Step 10.1.1) Load the trained model, input a set of historical microseismic frequency time series data with a time step of time_step days, and the model predicts the microseismic frequency in time_step+1 days; Step 10.1.2) Add the predicted microseismic frequencies to the historical microseismic frequency time series, move the sliding window forward one unit as the input for the new microseismic time series, and continue to predict the microseismic frequencies for time_step+2 days. Step 10.1.3) Repeat step 10.1.2) to achieve long-term prediction of microseismic frequencies; Step 10.2) Long-term prediction of microseismic energy levels: Step 10.2.1) Load the trained model, input a set of historical microseismic energy level time series data with a time step of time_step days, and the model predicts the microseismic energy level for time_step+1 days; Step 10.2.2) Add the predicted microseismic energy level to the historical microseismic energy level time series, move the sliding window forward one unit as the input for the new microseismic time series, and continue to predict the microseismic energy level for time_step+2 days; Step 10.2.3) Repeat step 10.2.2) to achieve long-term prediction of microseismic energy levels.

8. A method for early warning using the prediction method according to any one of claims 1-7, characterized in that: Based on the micro-vibration frequency and energy warning indicators in the coal mine working face anti-impact design report, the predicted values ​​of this invention are compared with the warning indicators. When the predicted value is greater than the critical indicator, an audible and visual alarm is triggered: when the predicted value is between 1.0 and 1.2 times the critical indicator, it is set as a Level 1 blue warning; when the predicted value is between 1.2 and 1.5 times the critical indicator, it is set as a Level 2 yellow warning; and when the predicted value is greater than 1.5 times the critical indicator, it is set as a Level 3 red warning.

Images