Mine micro-seismic event time sequence prediction method based on KF-LSTM-BO network

Abstract

The invention relates to a mine micro-seismic event time sequence prediction method, in particular to a mine micro-seismic event time sequence prediction method based on a KF-LSTM-BO network, and the method comprises the following steps: firstly, performing logarithm processing on time-space intensity data to enhance autocorrelation; meanwhile, in order to reduce the influence of accidental high-strength micro-seismic events on the data and reduce the difference between the data, a maximum and minimum value normalization method is adopted for normalization. In order to enhance the prediction effect of the irregular data, a KF filter is adopted to carry out de-noising reprocessing on the processed data. In consideration of inconvenience of manual parameter adjustment and high degree of dependence on deep learning knowledge, Bayesian optimization is introduced to perform hyper-parameter automatic adjustment and optimization on the KF-LSTM. According to the invention, the monitoring data can be acquired in real time for prediction, the safety of mining operation is improved, the production safety is ensured, and casualties are reduced at the same time.

Description

Technical Field

[0001] The present invention relates to a method for predicting the time series of microseismic events in a mine, and in particular to a method for predicting the time series of microseismic events in a mine based on a KF-LSTM-BO network. Background Art

[0002] The instantaneous rupture of rock mass caused by mining operations will excite elastic waves. These elastic waves carry information about the rupture source and propagate around through the elastic medium of the rock mass, thus causing microseismic events. Among them, microseismic events include large-energy microseismic events, which often induce coal-rock dynamic disasters such as rock burst pressure and damage to well tunneling equipment, causing heavy losses such as casualties. Therefore, the prediction of large-energy microseismic events has become a rigid demand and scientific research hotspot in the coal industry. There are many methods for predicting large-energy microseismic events, among which microseismic monitoring methods are important methods for monitoring and early warning of large-energy microseismic events in coal mines.

[0003] Microseismic events in mines are a common phenomenon in the mining process, which may lead to the destruction of the mine geological structure and the occurrence of miner safety accidents. In recent years, the monitoring and prediction technology of microseismic events in mines has played a key role in mine safety management. Although some progress has been made, the existing prediction methods still have certain limitations due to the complexity and uncertainty of geophysical processes. Therefore, it is of great theoretical and practical significance to explore more efficient and accurate methods for predicting microseismic events in mines.

[0004] The autocorrelation of microseismic data is a key factor in predicting microseismic events. Common methods for enhancing data autocorrelation include the difference method and the logarithm method. Due to the existence of large-energy microseismic events, the intensity data of microseismic spatiotemporal data have large differences at certain time points. It may be difficult to improve the autocorrelation by simply using the difference method to reduce the difference between the data. Therefore, this paper uses the logarithm method to enhance the autocorrelation. Summary of the invention

[0005] The purpose of the present invention is to provide a method for predicting the time series of mine microseismic events based on a KF-LSTM-BO network. The present invention introduces a Kalman filter (KF) before the LSTM network, establishes a network to extract quality-related features, and uses the loss function of this network (KF-LSTM) as the objective function of Bayesian optimization to achieve automatic tuning of the hyperparameters of the KF-LSTM network.

[0006] The objective of the present invention is achieved through the following technical solutions:

[0007] A method for predicting time series of mine microseismic events based on a KF-LSTM-BO network, the method comprising the following steps:

[0008] S1: Be familiar with the specific operation process of mining operations and understand the safety hazards and causes of microseismic events in mines;

[0009] S2: Receive microseismic vibration waves through the microseismic monitoring system’s seismic pickup, collect data through the collector, and perform calculations and analysis through the recorder and analyzer to obtain microseismic time, space, and intensity monitoring data;

[0010] S3: The time interval between two microseismic events is calculated using the difference method as time series data for prediction;

[0011] S4: Logarithmic data is used to enhance data autocorrelation, and the maximum and minimum value normalization method is used to preprocess the data;

[0012] S5: A labeled dataset is constructed using a sliding window format, and the training set and test set are divided into 7:3.

[0013] S6: Use KF filter to filter and denoise the data, and use the filtered data as the input of the long short-term memory network;

[0014] S7: Take MSE as the loss function of the KF-LSTM network, minimize the loss function to train the model, and use the loss function as the objective function of the Bayesian optimization algorithm to automatically adjust the model hyperparameters from a macro perspective to achieve automatic tuning.

[0015] The method for predicting the time series of microseismic events in mines based on the KF-LSTM-BO network, wherein the familiarity with the specific operation process of mining operations includes being familiar with the monitoring of microseismic events and the processing of data.

[0016] The method for predicting the time series of microseismic events in mines based on the KF-LSTM-BO network uses a differential method to calculate the time interval between two adjacent microseismic events for prediction; uses logarithmic data to enhance data autocorrelation, and uses a maximum and minimum normalization method to preprocess the data; builds a model, uses a Kalman filter (KF) to denoise the strong spatiotemporal features, uses the denoising results as LSTM input data, and predicts the time, space, and intensity of microseismic events through the LSTM network.

[0017] The method for predicting the time series of microseismic events in mines based on the KF-LSTM-BO network described herein minimizes the loss function to train the model, and automatically tunes the model hyperparameters through Bayesian optimization. The optimization model parameter training part is:

[0018] (1) Preprocess the data;

[0019] (2) Divide the training set and test set;

[0020] (3) Input the feature data into KF to obtain smooth data;

[0021] (4) Establish a KF-LSTM network to extract deep features of data;

[0022] (5) Build a KF-LSTM-BO model and use BO to automatically adjust the KF-LSTM hyperparameters to minimize the loss function and achieve automatic tuning;

[0023] (6) Using the optimal hyperparameters and minimizing the loss function, a microseismic event time series prediction model is obtained.

[0024] The method for predicting the time series of microseismic events in mines based on the KF-LSTM-BO network, wherein the test data is input into the trained microseismic event time series prediction model to predict the time, space and intensity of the next microseismic event, and ensure the safety of mining operations, is as follows:

[0025] (1) Perform the same data preprocessing on the test data as on the training data;

[0026] (2) The processed data is input into the trained model to obtain the predicted data, and the MSE is calculated to evaluate the model prediction effect. Finally, KF inverse processing and inverse normalization are used to obtain the actual data, so as to draw conclusions and output the model.

[0027] The advantages and effects of the present invention are:

[0028] 1. Based on the LSTM network, the present invention proposes a method for predicting the time series of microseismic events in mines by using a deep learning network (KF-LSTM-BO) that integrates the KF denoising method and BO automatic tuning. First, the spatiotemporal strong data is logarithmically processed to enhance the autocorrelation. At the same time, in order to reduce the impact of accidental high-intensity microseismic events on the data and reduce the difference between the data, the maximum and minimum normalization method is used for normalization. In order to enhance the prediction effect of irregular data, the KF filter is used to denoise and reprocess the processed data. Taking into account the inconvenience of manual parameter adjustment and the high degree of dependence on deep learning knowledge, Bayesian optimization is introduced to automatically tune the hyperparameters of KF-LSTM.

[0029] 2. The KF-LSTM network using the KF filter can better smooth the data and accurately predict, ensuring the accuracy of the data. With the pruning strategy, the network based on KF-LSTM-BO can more efficiently find the optimal parameters and accurately predict the results, with better prediction performance.

[0030] 3. The present invention satisfies the need for real-time collection of monitoring data for prediction, improves the safety of mining operations, ensures production safety, and reduces the occurrence of casualties. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 It is a structural diagram of the model of the present invention;

[0032] Figure 2 It is the prediction result diagram of the working face 103 of the present invention;

[0033] Figure 3 is the error threshold diagram of the working surface 103 of the present invention;

[0034] Figure 4 It is the prediction result diagram of the working face 105 of the present invention;

[0035] Figure 5 is the error threshold diagram of the working surface 105 of the present invention;

[0036] Figure 6 It is the overall flow chart of the present invention. DETAILED DESCRIPTION

[0037] The present invention is described in detail below with reference to the implementation examples shown in the accompanying drawings.

[0038] Glossary:

[0039] KF: Kalman Filter;

[0040] LSTM: Long-Short Term Memory;

[0041] BO: Bayesian Optimization;

[0042] A method for predicting time series of mine microseismic events based on a KF-LSTM-BO network, the method comprising the following steps:

[0043] S1: Be familiar with the specific operation process of mining operations and understand the safety hazards and causes of microseismic events in mines;

[0044] S2: Receive microseismic vibration waves through the microseismic monitoring system’s seismic pickup, collect data through the collector, and perform calculations and analysis through the recorder and analyzer to obtain microseismic time, space, and intensity monitoring data;

[0045] S3: The time interval between two microseismic events is calculated using the difference method as time series data;

[0046] S4: Logarithmic data is used to enhance data autocorrelation, and the maximum and minimum value normalization method is used to preprocess the data;

[0047] S5: A labeled dataset is constructed using a sliding window format, and the training set and test set are divided into 7:3.

[0048] S6: Use KF filter to filter and denoise the data, and use the filtered data as the input of the long short-term memory network;

[0049] S7: Take MSE as the loss function of the KF-LSTM network, minimize the loss function to train the model, and use the loss function as the objective function of the Bayesian optimization algorithm to automatically adjust the model hyperparameters from a macro perspective to achieve automatic tuning.

[0050] S8: Input the test data into the KF-LSTM-BO model for prediction to obtain the time interval for the next microseismic event, thereby inferring the possible time of the next microseismic event, reducing the probability of mining accidents and improving mining safety.

[0051] The present invention adopts the method described in S3 and S4 to perform data preprocessing, and the specific steps are as follows:

[0052] S31: The differential method introduces the formula as follows:

[0053] x′(t)=x(t+1)-x(t),x(0)=0(1)

[0054] S41: The formula for calculating autocorrelation is as follows:

[0055]

[0056] Where h is the time interval, cov(X t ,X t+h ) represents the variable sequence X t ,X t+h The covariance of t σ t+h Represents the standard deviation, which is used as a standardization factor to prevent the influence of the magnitude of the variable change. The autocorrelation ACF value of each event interval point is between -1 and 1. The larger the value, the higher the autocorrelation; the smaller the value, the lower the autocorrelation.

[0057] S42: The formula for introducing the maximum and minimum value normalization method is as follows:

[0058]

[0059] in is the minimum value of the microseismic time series data, is the maximum value of the microseismic event time series data, x d is the value of microseismic event time data to be normalized.

[0060] The present invention uses the Bayesian optimization algorithm described in S7 to perform hyperparameter tuning, and the specific steps are as follows:

[0061] S71: The formula introduced by Bayesian optimization is as follows:

[0062]

[0063]

[0064] In the Bayesian optimization algorithm, Represents the best y value in the parameter adjustment history library after t-1 iterations. EI represents the expected improvement, which is the probability distribution model M generated based on the current t-1 observations. t-1 The expected improvement obtained.

[0065] As a further improvement, the present invention uses Bayesian optimization to perform hyperparameter tuning and uses a pruning strategy to optimize the tuning process and improve the tuning efficiency.

[0066] As the instruction manual Figure 1 A method for predicting time series of microseismic events in mines based on a KF-LSTM-BO network is shown. The method realizes smoothing of high-noise data and effective prediction of microseismic events based on an LSTM network and KF. The process of predicting microseismic events in the present invention is as follows:

[0067] Training phase:

[0068] (1) Data preprocessing and KF smoothing;

[0069] (2) Initialize network structure and parameters;

[0070] (3) Input the KF smoothing result into the LSTM network to obtain the predicted value;

[0071] (4) Use the Adam optimizer and the loss function MSE to perform multiple rounds of training to stabilize the loss function value;

[0072]

[0073] (5) The stabilized loss function value is used as the objective function, and the Bayesian optimization method is used to optimize the hyperparameters of the KF-LSTM network as a whole. Multiple rounds of iterations are performed to find the optimal hyperparameters and train the optimal model.

[0074] (6) Save the optimal network structure and parameters.

[0075] Testing phase:

[0076] The test phase uses the same preprocessing method as the training phase for preprocessing, and uses the same parameters for KF smoothing, and uses the trained optimal model to obtain the predicted value. By comparing the predicted value with the label, the error threshold is obtained, and the plot is as shown in the attached manual. Figure 5 shown.

[0077] Example 1: 105 working surface

[0078] Table 1105 Optimal hyperparameters for working face time series - 100 rounds

[0079]

[0080]

[0081] Table 1 shows the best hyperparameters of this model in the time dimension after 100 rounds of Bayesian optimization and 105 work surfaces. The MSE value of this model is 0.02142403355325218.

[0082] In conjunction with the instruction manual Figure 5 And the instruction manual Figure 4 It can be seen that the error accuracy of the KF-LSTM model is very high, and the accuracy is greater than 80% at an error threshold of 0.13. This is because KF smoothes the original data, removes the influence of noise on the prediction process, and retains the deep features of the original data, greatly improving the prediction performance.

[0083] Example 2: 103 working surface

[0084] Table 2103 Optimal hyperparameters for working face time series - 100 rounds

[0085] Params Values batch_size 69 dropout_rate_1 0.6076339853715965 epochs 158 layers 1 learning_rate 0.002014621923881111 units 191

[0086] Table 2 shows the best hyperparameters used by this model in the time dimension after 100 rounds of Bayesian optimization and 103 work surfaces. The MSE value of the model is 0.05773200020961622.

[0087] In conjunction with the instruction manual Figure 3 And the instruction manual Figure 2 It can be seen that the error accuracy of the KF-LSTM model is very high, and the accuracy is greater than 80% at an error threshold of 0.11. This is because KF smoothes the original data, removes the influence of noise on the prediction process, and retains the deep features of the original data, greatly improving the prediction performance.

[0088] Different implementations on multiple different working surfaces demonstrate the adaptability of the model. This is because the automatic tuning of Bayesian optimization will make targeted adjustments to the different characteristics of each working surface to find suitable model parameters for each working surface, thereby finding the optimal model.

[0089] The specific embodiments of the present invention have been disclosed above. The present invention is applicable to various fields mentioned in the specification and other suitable fields. Without departing from the general concept defined by the claims and their equivalent scope, the present invention is not limited to the description examples given in the specification.

Claims

1. A method for predicting time series of microseismic events in mines based on KF-LSTM-BO network, characterized in that: The method comprises the following steps: S1: Be familiar with the specific operation process of mining operations and understand the safety hazards and causes of microseismic events in mines; S2: Receive microseismic vibration waves through the microseismic monitoring system’s seismic pickup, collect data through the collector, and perform calculations and analysis through the recorder and analyzer to obtain microseismic time, space, and intensity monitoring data; S3: The time interval between two microseismic events is calculated using the difference method as time series data for prediction; S4: Logarithmic data is used to enhance data autocorrelation, and the maximum and minimum value normalization method is used to preprocess the data; S5: A labeled dataset is constructed using a sliding window format, and the training set and test set are divided into 7:

3. S6: Use KF filter to filter and denoise the data, and use the filtered data as the input of the long short-term memory network; S7: Take MSE as the loss function of the KF-LSTM network, minimize the loss function to train the model, and use the loss function as the objective function of the Bayesian optimization algorithm to automatically adjust the model hyperparameters from a macro perspective to achieve automatic tuning.

2. According to claim 1, a method for predicting time series of mine microseismic events based on KF-LSTM-BO network is characterized in that: The familiarity with the specific operational procedures of mining operations includes being familiar with the monitoring of microseismic events and the processing of data.

3. The method for predicting time series of mine microseismic events based on KF-LSTM-BO network according to claim 1 is characterized in that: The method uses the difference method to calculate the time interval between two adjacent microseismic events for prediction; uses logarithmic data to enhance data autocorrelation, and uses the maximum and minimum normalization method to preprocess the data; builds a model, uses the Kalman filter (KF) to denoise the strong spatiotemporal features, uses the denoising result as LSTM input data, and predicts the time, space and intensity of microseismic events through the LSTM network.

4. The method for predicting time series of mine microseismic events based on KF-LSTM-BO network according to claim 1 is characterized in that: The loss function is minimized to train the model, and the model hyperparameters are automatically tuned through Bayesian optimization. The optimization model parameter training part is: (1) Preprocess the data; (2) Divide the training set and test set; (3) Input the feature data into KF to obtain smooth data; (4) Establish a KF-LSTM network to extract deep features of data; (5) Build a KF-LSTM-BO model and use BO to automatically adjust the KF-LSTM hyperparameters to minimize the loss function and achieve automatic tuning; (6) Using the optimal hyperparameters and minimizing the loss function, a microseismic event time series prediction model is obtained.

5. The method for predicting time series of mine microseismic events based on KF-LSTM-BO network according to claim 1 is characterized in that: The test data is input into the trained microseismic event time series prediction model to predict the time, space and intensity of the next microseismic event, and to ensure the safety of mining operations. The test part is: (1) Perform the same data preprocessing on the test data as on the training data; (2) The processed data is input into the trained model to obtain the predicted data, and the MSE is calculated to evaluate the model prediction effect. Finally, KF inverse processing and inverse normalization are used to obtain the actual data, so as to draw conclusions and output the model.

Images