Mine micro-seismic time sequence prediction method

Abstract

The invention relates to the technical field of mine safety monitoring and prediction, and particularly discloses a mine micro-seismic time sequence prediction method, which comprises the following steps of: decomposing original micro-seismic data into a plurality of intrinsic mode sequences by using an improved completely adaptive noise set empirical mode decomposition method; sampling each intrinsic mode sequence by a consistent random sampling method, calculating a sampling sample entropy, and setting a sample entropy threshold to reconstruct different intrinsic mode sequences into a high-frequency sequence, an intermediate-frequency sequence and a low-frequency sequence; the three sequences are processed through variational mode decomposition, center frequencies under different decomposition numbers are calculated to determine the optimal decomposition number, and finally a multivariate feature time sequence of the microseismic data is obtained; an input and output data set is constructed by adopting a sliding decomposition mechanism, a multivariate feature time sequence in a period of time is used as input, micro-seismic data in a later period of time is used as output, and time sequence prediction is carried out in combination with a Transform model with autocorrelation decomposition.

Description

Technical Field

[0001] The present invention relates to the technical field of mine safety monitoring and prediction. More specifically, the present invention relates to a method for predicting mine microseismic time series. Background Art

[0002] During the deep coal mining process, microseismic events generated by the redistribution of rock mass stress are often used as important indicators for mine safety early warning. The microseismic sequence data monitored by microseismic is a typical non-linear and non-stationary time series data, which can usually reflect the dynamic change trend of variables over time. Before a rock burst occurs, the frequency and energy release of microseismic show unique evolution trends and characteristics. With the rapid development of machine learning and deep learning technologies, the prediction of microseismic time series based on deep learning is expected to further improve the feasibility, accuracy, and timeliness of impact early warning. Existing literature 1 (Lv Jinguo, Pan Li. Time series method for microseismic early warning of rock bursts [J]. Journal of China Coal Society, 2010, (12): 2002-2005.) achieved effective early warning of the future microseismic energy release trend by using the autoregressive integrated moving average (ARIMA) model and the threshold autoregressive model in time series analysis; Existing literature 2 (Lan Tianwei, Guo Xutao, Zhang Zhijia, et al. Prediction of microseismic events in rock burst mines based on MEA-BP neural network [J]. Scientific Reports, 2023, 13(1).) used a BP neural network model optimized by the mind evolution algorithm (MEA) to predict microseismic events of rock bursts in mines, effectively improving the prediction accuracy.

[0003] Due to the non - stationary characteristics of micro - seismic energy data, it is difficult to guarantee the accuracy when directly applying time - series models to predict the original data. The modal decomposition technology effectively solves the complex problem of the intertwined frequency components, trend components, and noise components in non - stationary and multi - scale signals. Especially when combined with neural networks, remarkable results have been achieved. Extending the single - stage modal decomposition to the two - stage modal decomposition can more effectively extract the hidden information in the original data sequence and further improve the prediction accuracy of the model. Existing literature 3 (Xiang Ling, Liu Jianing, Su Hao, et al. Research on multi - step wind speed prediction based on CEEMDAN two - stage decomposition and LSTM [J]. Acta Energiae Solaris Sinica, 2022, 43(8): 334 - 339.) proposed a multi - step wind speed prediction method based on the combination of CEEMDAN two - stage decomposition and LSTM network; existing literature 4 (Chen Jinpeng, Hu Zhijian, Chen Weinan, et al. Multi - load prediction of integrated energy system based on two - stage modal decomposition combined with DBiLSTM - MLR [J]. Automation of Electric Power Systems, 2021, 45(13): 85 - 94.) proposed a multi - load prediction method based on two - stage modal decomposition, deep bidirectional long - short - term memory (DBiLSTM) neural network, and multiple linear regression (MLR) model. Based on the above, in order to achieve the prediction of micro - seismic energy intensity in practical applications and ensure a high prediction accuracy, it is urgent to develop a micro - seismic time - series prediction method that combines double - layer modal decomposition technology and a self - correlation mechanism - optimized Transformer (or Autoformer) model to improve the ability to capture the evolution trend of mine micro - seismicity and prediction accuracy. To solve the above problems, a technical solution is provided as follows. Summary of the Invention

[0004] To overcome the above - mentioned defects of the prior art, the present invention provides a method for predicting mine micro - seismic time series, developing a micro - seismic time - series prediction method that combines double - layer modal decomposition technology and a self - correlation mechanism - optimized Autoformer model to improve the ability to capture the evolution trend of mine micro - seismicity and prediction accuracy, so as to solve the problems raised in the above - mentioned background technology.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] A method for predicting mine micro - seismic time series includes the following steps:

[0007] Step 1, data pre - processing and primary modal decomposition: After denoising, data cleaning, and standardization pre - processing of the collected original micro - seismic event sequence, the complete ensemble empirical mode decomposition with adaptive noise (ICEEMDAN) is used to decompose the original micro - seismic event sequence into several modal decomposition features;

[0008] Step 2, Sample Entropy Calculation and Sequence Reconstruction: For each modal decomposition feature, the sample entropy is calculated using the consistent random sampling method, and different modal decomposition features are reconstructed into high-frequency sequences, medium-frequency sequences, and low-frequency sequences by setting the sample entropy threshold;

[0009] Step 3, Second Modal Decomposition: Apply variational modal decomposition to the high-frequency sequence, medium-frequency sequence, and low-frequency sequence, calculate the central frequencies under different decomposition numbers to determine the optimal decomposition number, and obtain the characteristic time series of the microseismic data;

[0010] Step 4, Dataset Construction: Use the sliding window method to combine the original microseismic event sequence with each modal decomposition feature sequence to construct a multi-feature dataset, forming input and output sample pairs, where the input is the multi-dimensional feature data within a set time window, and the output is the microseismic energy data at the subsequent time;

[0011] Step 5, Time Series Prediction Based on the Autoformer Model: Input the multi-feature dataset into the Autoformer model for time series prediction;

[0012] Step 6, Model Training and Validation: Compare the single-step and multi-step prediction tasks, adjust the input sequence length, prediction step, and network hyperparameters, and use the mean square error (MSE), root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE), and coefficient of determination (R 2 ) metrics to evaluate the performance of the Autoformer model.

[0013] As a further solution of the present invention, in Step 1, the specific steps of data preprocessing and primary modal decomposition are:

[0014] Step 11, Obtain the microseismic event sequence E L ={(t1, e1),..., (t i , e i ),..., (t n , e n )} during the monitoring period by deploying fiber optic geophones at the microseismic detection points, where t i represents the time point, e i represents the microseismic energy, and n is the total number of microseismic detection points;

[0015] Step 12, Perform denoising, data cleaning, and standardization preprocessing on the data in the microseismic event sequence E L ;

[0016] Step 13, Obtain the energy within Δ t time as and the microseismic energy at the next moment is e t+1, there exists a mapping function f, and the following relationship is reflected based on the mapping function:

[0017] Step 14, using the improved complete adaptive noise ensemble empirical mode decomposition (ICEEMDAN) to decompose the original microseismic event sequence into several modal decomposition features, expressed as:

[0018] E L =

[0019] {(t1, v 11 , v 12 , …, v 1j , e1), …, (t i , v i1 , v i2 , …, v ij , e i ), …, (t n , v n1 , v n2 , …, v nk , e n )};

[0020] In the formula: v ij is the j-th modal decomposition feature of the i-th microseismic detection point, n is the total number of microseismic detection points, and k is the number of modal decomposition features.

[0021] In step 2, sample entropy calculation and sequence reconstruction: The sample entropy is calculated for each modal decomposition feature using the uniform random sampling method, and different modal decomposition features are reconstructed into high-frequency sequences, medium-frequency sequences, and low-frequency sequences by setting the sample entropy threshold; during the modal decomposition process of the original microseismic event sequence, the sample entropy is used to evaluate the importance and information content of each IMF sequence. Since the reconstructed sequence needs to calculate the sample entropy of each IMF sequence, the time complexity of sample entropy calculation is O(N 2 ·m), where N is the sequence length, and it is usually considered that N is much larger than m. To reduce the calculation time of the sample entropy, the uniform random sampling method is adopted, that is, K data points are randomly selected from the original time series (usually K << N), and then the sample entropy is calculated only based on these data points. Through strategies such as random sampling, the calculation time can be effectively reduced while still maintaining reasonable result accuracy.

[0022] As a further solution of the present invention, in step 3, the second modal decomposition: Apply variational mode decomposition to the high-frequency sequence, medium-frequency sequence, and low-frequency sequence, calculate the central frequencies under different decomposition numbers to determine the optimal decomposition number, and obtain the characteristic time series of the microseismic data;

[0023] The decomposition effect of Variational Mode Decomposition (VMD) highly depends on the choice of the number of modes. When the number of modes is too small, the VMD algorithm may act as an adaptive filter bank, resulting in the filtering out of key information in the original signal, thereby reducing the prediction accuracy; while when the number of modes is too large, the central frequencies of adjacent mode components are close, which may lead to the generation of redundancy or additional noise. The number of mode decompositions K is initially estimated according to the sample entropy of each sequence, and based on the central frequency distribution under different numbers of mode decompositions, a suitable number of decompositions is selected to obtain n modal feature (V-IMF) sequences. The V-IMF sequences are combined with the mode decomposition features to obtain n + 1 multivariate feature time series sample sets.

[0024] As a further solution of the present invention, in step 4, dataset construction: The original microseismic event sequence and each mode decomposition feature sequence are combined using the sliding window method to construct a multivariate feature dataset, forming input and output sample pairs, where the input is the multi-dimensional feature data within a set time window, and the output is the microseismic energy data at a subsequent time; the specific method is as follows: First, set the time window Δt (i.e., the time step) and define the sliding step S. Within each sliding window, the first Δt time series values and the corresponding mode feature sequence component data are used as the input features of the model. At the end of each time window, the data values at the next moment or multiple subsequent moments are used as the output. Through this process, the model will learn how to predict one or more time series values in the future based on the features of the first Δt time steps.

[0025] As a further solution of the present invention, in step 5, the Autoformer model decomposes the feature time series into a trend component and a seasonal component; the input features of the encoder and decoder are embedded into a high-dimensional space. Each encoder module consists of multiple encoding layers, which are responsible for extracting the time series features; each layer of the encoder includes an autocorrelation layer, a feed-forward neural network, a Layer Normalization layer, and a Dropout layer;

[0026] As a further solution of the present invention, in step 5, the autocorrelation layer is used to capture the long-term dependencies in the time series. The feed-forward neural network consists of two fully connected networks and combines a non-linear activation function to enhance the feature representation ability; the Layer Normalization layer and the Dropout layer help to stabilize the training process and prevent overfitting.

[0027] As a further solution of the present invention, in step 5, each decoder in the Autoformer model consists of multiple decoding layers, where each layer includes an autocorrelation layer and a cross-autocorrelation layer, the interaction between the decoder layer and the encoder layer, and the reconstruction of the trend component and the seasonal component. In a long time series prediction task, the decoder output is combined with the encoder features to reconstruct the trend component and the seasonal component, and finally generate the prediction result.

[0028] Technical effects and advantages of a mine microseismic time series prediction method of the present invention: The present invention uses an improved complete adaptive noise ensemble empirical mode decomposition method to decompose the original microseismic data into multiple intrinsic mode sequences, then samples each intrinsic mode sequence by the uniform random sampling method, calculates its sample entropy, and sets a sample entropy threshold to reconstruct different intrinsic mode sequences into high-frequency sequences, medium-frequency sequences, and low-frequency sequences; applies variational mode decomposition to process these three sequences respectively, and calculates the central frequencies under different decomposition numbers to determine the optimal decomposition number, and finally obtains the multivariate characteristic time series of the microseismic data; constructs an input-output data set by using a sliding decomposition mechanism, takes the multivariate characteristic time series within a period of time as the input, and the microseismic data within the subsequent period of time as the output, and combines a Transformer model with autocorrelation decomposition for time series prediction. Description of the Drawings

[0029] Figure 1 It is the overall architecture diagram of the Autoformer model provided by the present invention;

[0030] Figure 2 It is the C-IMF sequence diagram generated by complete adaptive noise ensemble empirical mode decomposition provided by the present invention;

[0031] Figure 3 It is the change diagram of the sample entropy of each sequence under different parameters provided by the present invention;

[0032] Figure 4 It is the reconstruction result diagram based on the sample entropy sequence provided by the present invention;

[0033] Figure 5 It is the variational mode decomposition result diagram of the high-frequency sequence provided by the present invention;

[0034] Figure 6 It is the model prediction effect diagram of different input sequence lengths under single-step prediction provided by the present invention;

[0035] Figure 7 It is the model prediction effect diagram of different input sequence lengths under multi-step prediction provided by the present invention;

[0036] Figure 8 It is the model training process diagram of multi-step prediction single-step provided by the present invention;

[0037] Figure 9 It is the learning rate adjustment process diagram of multi-step prediction single-step provided by the present invention;

[0038] Figure 10 It is the model training process diagram of multi-step prediction multi-step provided by the present invention;

[0039] Figure 11 The single-step prediction long model fitting curve graph provided by the present invention;

[0040] Figure 12 The multi-step prediction model fitting curve graph provided by the present invention. Specific implementation manners

[0041] Next, the technical solutions in the present invention will be clearly and completely described in conjunction with the accompanying drawings in the present invention. Obviously, the described technical solutions are only a part of the present invention, rather than all of them. Based on the technical solutions in the present invention, all other technical solutions obtained by those of ordinary skill in the art without making creative efforts fall within the protection scope of the present invention.

[0042] Embodiment 1

[0043] A mine microseismic time series prediction method includes the following steps:

[0044] Step 1, data preprocessing and primary modal decomposition: After denoising, data cleaning, and standardization preprocessing of the collected original microseismic event sequence, the complete adaptive noise ensemble empirical mode decomposition (ICEEMDAN) is used to decompose the original microseismic event sequence into several modal decomposition features;

[0045] Step 2, sample entropy calculation and sequence reconstruction: The sample entropy is calculated for each modal decomposition feature using the consistent random sampling method, and different modal decomposition features are reconstructed into high-frequency sequences, medium-frequency sequences, and low-frequency sequences by setting the sample entropy threshold;

[0046] Step 3, second modal decomposition: The variational modal decomposition is applied to the high-frequency sequence, medium-frequency sequence, and low-frequency sequence, and the central frequencies under different decomposition numbers are calculated to determine the optimal decomposition number, obtaining the characteristic time series of the microseismic data;

[0047] Step 4, data set construction: The sliding window method is used to combine the original microseismic event sequence and each modal decomposition feature sequence to construct a multi-feature data set, forming input and output sample pairs, where the input is multi-dimensional feature data within a set time window, and the output is the microseismic energy data at a subsequent time;

[0048] Step 5, time series prediction based on the Autoformer model: The multi-feature data set is input into the Autoformer model for time series prediction;

[0049] Step 6, model training and verification: Comparing single-step and multi-step prediction tasks, by adjusting the input sequence length, prediction step length, and network hyperparameters, using the mean square error (MSE), root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE), and coefficient of determination (R2 ) Evaluate the performance of the Autoformer model using metrics.

[0050] Specifically, in step 1, the specific steps of data preprocessing and initial modal decomposition are as follows:

[0051] Step 11: Obtain the microseismic event sequence E during the monitoring period by deploying fiber optic geophones at microseismic detection points L ={(t1, e1),...,(t i , e i ),...,(t n , e n )}, where t i represents the time point and e i represents the microseismic energy, and n is the total number of microseismic detection points;

[0052] Step 12: Perform denoising, data cleaning, and standardization preprocessing on the data in the microseismic event sequence E L ;

[0053] Step 13: Obtain the energy within Δ t time as The microseismic energy at the next moment is e t+1 , and there exists a mapping function f, reflecting the following relationship based on the mapping function:

[0054] Step 14: Use the improved complete adaptive noise ensemble empirical mode decomposition (ICEEMDAN) to decompose the original microseismic event sequence into several modal decomposition features, expressed as:

[0055] E L =

[0056] {(t1, v 11 , v 12 , …, v 1j , e1), …, (t i , v i1 , v i2 , …, v ij , e i ), …, (t n , v n1 , v n2 , …, v nk , e n )};

[0057] In the formula: v ij is the jth modal decomposition feature of the ith microseismic detection point, n is the total number of microseismic detection points, and k is the number of modal decomposition features.

[0058] Specifically, in step 2, sample entropy calculation and sequence reconstruction: For each modal decomposition feature, the sample entropy is calculated using the uniform random sampling method. By setting the sample entropy threshold, different modal decomposition features are reconstructed into high-frequency sequences, medium-frequency sequences, and low-frequency sequences. During the modal decomposition of the original microseismic event sequence, the sample entropy is used to evaluate the importance and information content of each IMF sequence. Since the sample entropy of each IMF sequence needs to be calculated for the reconstructed sequence, the time complexity of sample entropy calculation is O(N 2 ·m), where N is the sequence length, and it is usually considered that N is much larger than m. To reduce the calculation time of the sample entropy, the uniform random sampling method is adopted, that is, K data points are randomly selected from the original time series (usually K << N), and then the sample entropy is calculated only based on these data points. Through strategies such as random sampling, the calculation time can be effectively reduced while still maintaining reasonable result accuracy.

[0059] Specifically, in step 3, the second modal decomposition: Apply variational mode decomposition to the high-frequency sequence, medium-frequency sequence, and low-frequency sequence, calculate the central frequencies under different decomposition numbers to determine the optimal decomposition number, and obtain the characteristic time series of the microseismic data;

[0060] The decomposition effect of variational mode decomposition (VMD) highly depends on the choice of the number of modes. When the number of modes is too small, the VMD algorithm may act as an adaptive filter bank, resulting in the filtering of key information in the original signal, thus reducing the prediction accuracy; while when the number of modes is too large, the central frequencies of adjacent modal components are close, which may lead to the generation of redundancy or additional noise. Based on the sample entropy of each sequence, the number of modal decompositions K is initially estimated, and according to the central frequency distribution under different numbers of modal decompositions, a suitable decomposition number is selected to obtain n modal feature (V-IMF) sequences. Combine the V-IMF sequences with the modal decomposition features to obtain n + 1 multivariate feature time series sample sets.

[0061] Specifically, in step 4, dataset construction: Use the sliding window method to combine the original microseismic event sequence with each modal decomposition feature sequence to construct a multivariate feature dataset, forming input and output sample pairs, where the input is the multi-dimensional feature data within the set time window, and the output is the microseismic energy data at the subsequent time. The specific method is as follows: First, set the time window Δt (i.e., the time step) and define the sliding step S. Within each sliding window, the first Δt time series values and the corresponding modal feature sequence component data are used as the input features of the model. At the end of each time window, the data values at the next moment or multiple subsequent moments are used as the output. Through this process, the model will learn how to predict one or more future time series values based on the features of the first Δt time steps.

[0062] Specifically, such as Figure 1The overall architecture diagram of the Autoformer model is shown in the figure. In step 5, the Autoformer model decomposes the feature time series into trend components and seasonal components. The input features of the encoder and decoder are embedded in a high-dimensional space. Each encoder module consists of multiple encoding layers and is responsible for extracting time series features. Each encoder layer includes an autocorrelation layer, a feedforward neural network, a layer normalization layer, and a dropout layer. The autocorrelation layer is used to capture long-term dependencies in the time series. The feedforward neural network consists of two fully connected layers and is combined with a nonlinear activation function to enhance feature representation capabilities. The layer normalization layer and the dropout layer help stabilize the training process and prevent overfitting.

[0063] Each decoder in the Autoformer model consists of multiple decoding layers, each of which includes autocorrelation layers and cross-autocorrelation layers, interactions between decoder layers and encoder layers, and reconstruction of trend and seasonal components. In long time series prediction tasks, the decoder output is combined with the encoder features to reconstruct trend and seasonal components and ultimately generate prediction results.

[0064] Example 2

[0065] In a certain mining area, the horizontal elevation range extends from -200 meters to -1200 meters, the strike length of the mining area is 13.5 kilometers, the inclined width is 2.9 kilometers, and the total area is approximately 37.8623 square kilometers.

[0066] A fiber-optic microseismic monitoring system has been deployed on the working face of the mining area. The system uses passive fiber-optic seismic detectors, high-resolution data acquisition technology, and precision clock synchronization technology. The sampling frequency is 2000Hz, which means that 8000 data items are received per second. Each data item contains 16 channels. The system can realize real-time dynamic monitoring of microseismic events in the mining area. Microseismic sensors are distributed within the well field and regionally centralized to ensure high-precision distributed data collection of mine microseismic events.

[0067] The microseismic data collected by the microseismic monitoring system covers key information such as microseismic energy, source location, and magnitude. During data processing, the collected microseismic data was first preprocessed to eliminate abnormal microseismic events caused by human activities and delete incomplete data, thus ensuring the continuity and integrity of the data. Ultimately, a new microseismic dataset was obtained.

[0068] To eliminate the influence of the wide range of microseismic energy values and thus improve the effectiveness of model training, a numerical fitting formula was used to convert microseismic energy into microseismic energy intensity. Changes in microseismic energy intensity reflect the changing trend of microseismic energy. The specific calculation formula is as follows.

[0069] M L = lg(A max ) + R Δ ;

[0070] R Δ = 1.66×lg(x) + 3.5;

[0071] Where: M L is the microseismic energy intensity, A max is the maximum amplitude, R Δ is the energy intensity fitting formula, and x is the microseismic energy.

[0072] Taking the prediction of microseismic energy intensity as an example, first, the time series data of the collected microseismic energy intensity is cleaned to remove noise and fill in missing values. After data cleaning, a set of time series data of microseismic energy intensity is obtained. Then, the complete ensemble empirical mode decomposition with adaptive noise (ICEEMDAN) decomposition method is used to perform the first modal decomposition on the data. Several C-IMF sequences are generated by ICEEMDAN decomposition, such as Figure 2 the C-IMF sequence diagram generated by the complete ensemble empirical mode decomposition with adaptive noise as shown. After completing the ICEEMDAN decomposition, next, uniform random sampling is performed on each C-IMF sequence, and then the sample entropy of each sampled sequence is calculated, such as Figure 3 the change diagram of the sample entropy of each sequence under different parameters provided. According to the size of the sample entropy, the C-IMF sequences are divided into three sequences: high-frequency, medium-frequency, and low-frequency. The method of setting thresholds is used for the division. Taking the sample entropy parameters m = 1 and r = 0.1 as an example, the sample entropy of the first four C-IMF sequences is above 1.5, and they are accumulated and reconstructed into a high-frequency sequence; the sample entropy of the last four C-IMF sequences is below 0.2, and they are accumulated and reconstructed into a low-frequency sequence; the sample entropy of the fifth and sixth C-IMF sequences is between 0.3 and 1, and they can be reconstructed into a medium-frequency sequence, such as Figure 4 the sequence reconstruction result diagram based on sample entropy provided. Finally, variational mode decomposition (VMD) is performed on the reconstructed high-frequency, medium-frequency, and low-frequency sequences. The different decomposition modal numbers are selected for each sequence by calculating the distribution of the central frequencies under different numbers. Finally, the decomposition number of the high-frequency sequence is 12, the decomposition number of the medium-frequency sequence is 6, and the decomposition number of the low-frequency sequence is 6, resulting in a total of 24 V-IMF sequences, such as Figure 5 the variational mode decomposition result diagram of the high-frequency sequence provided.

[0073] Construct an improved Autoformer model and use the above-constructed dataset to perform multi-step prediction single-step and multi-step prediction multi-step experiments on the microseismic energy intensity respectively. Before that, the optimal input sequence length needs to be determined first. In the multi-step prediction single-step experiment, the coefficient of determination (R2 ) The root mean square error (RMSE) varies with different input sequence lengths as shown in Figure 6 the model prediction effect diagrams for different input sequence lengths under single-step prediction provided. When the input sequence length is 24, the determination coefficient of the model reaches the highest and the root mean square error reaches the lowest. In the multi-step prediction multi-step experiment, the determination coefficient and root mean square error of the model vary with different input sequence lengths as shown in Figure 7 the model prediction effect diagrams for different input sequence lengths under multi-step prediction provided. Among them, the target sequence length is half of the input sequence length. The model with the best performance is when the input sequence length is 4 and the prediction sequence length is 2, and the determination coefficient R 2 reaches 0.827 and the root mean square error MSE is 0.0921. The model with the second-best performance is when the input sequence length is 24 and the output sequence length is 12, and the determination coefficient R 2 reaches 0.8174 and the root mean square error MSE is 0.0943. In order to reflect the microseismic change trend in the future period, under comprehensive evaluation, the model with an input sequence length of 24 and an output sequence length of 12 is adopted in the multi-step experiment.

[0074] In the multi-step prediction single-step experiment, the model with an input sequence length of 24 and a prediction length of 1 is retrained, and the training process is as shown in Figure 8 the multi-step prediction single-step model training process diagram provided. During the experiment, a dynamic learning rate adjustment and early stopping mechanism are set, and the learning rate change curve is as shown in Figure 9 the multi-step prediction single-step learning rate adjustment process diagram provided. In the early stopping mechanism, training stops when the reduction amplitude of the training set loss or validation set loss is less than a certain threshold for 5 consecutive times. It can be seen from the figure that after the 12th generation of iteration, the training set loss, validation set loss, and test set loss will tend to be stable. Finally, on the test set, the model fitting curve of the predicted value and the true value is as shown in Figure 11 the single-step prediction long model fitting curve diagram provided. In the multi-step prediction multi-step experiment, the input sequence length is 24 and the prediction sequence length is 12, and the training process is as shown in Figure 10 the multi-step prediction multi-step model training process diagram provided. After the 10th generation, each loss rate curve tends to be stable. On the test set, the multi-step prediction effect of randomly selecting two data points is as shown in Figure 12 the single-step prediction long model fitting curve diagram provided. Figure (a) shows the prediction effect of the 20th data point and Figure (b) shows the prediction effect of the 60th data point. The evaluation indexes of the multi-step prediction single-step and multi-step prediction multi-step models are shown in Table 1.

[0075]

[0076] Based on the experimental results, it can be seen that the improved Autoformer model performs excellently in the microseismic time series prediction task. Especially when dealing with the data after double-layer modal decomposition, it can effectively capture the time series law in the microseismic sequence and make accurate predictions. With its new decomposition structure and autocorrelation mechanism, the Autoformer model can better capture long-term dependencies, thus still maintaining relatively excellent performance in multi-step prediction. In the multi-step prediction experiment, the coefficient of determination R 2 reaches 0.8174, and in the single-step prediction experiment, its coefficient of determination R 2 even reaches 0.9256, showing good prediction ability and fitting effect.

[0077] In the embodiment of the present invention, the improved complete adaptive noise ensemble empirical mode decomposition method is used to decompose the original microseismic data into multiple intrinsic mode sequences, then the consistent random sampling method is used to sample each intrinsic mode sequence, calculate its sample entropy, and set the sample entropy threshold to reconstruct different intrinsic mode sequences into high-frequency sequences, medium-frequency sequences, and low-frequency sequences; variational mode decomposition is applied to process these three sequences respectively, and the central frequencies under different decomposition numbers are calculated to determine the optimal decomposition number, and finally the multi-feature time series of microseismic data is obtained; a sliding decomposition mechanism is used to construct the input-output data set, the multi-feature time series within a period of time is used as the input, and the microseismic data within the subsequent period of time is used as the output, and a time series prediction is carried out in combination with the Transformer model with autocorrelation decomposition.

[0078] The above is only the specific implementation manner of this application, but the protection scope of this application is not limited thereto. Any person skilled in the art within the technical scope disclosed in this application can easily think of changes or substitutions, which should all be covered by the protection scope of this application. Therefore, the protection scope of this application should be subject to the protection scope of the claims.

[0079] Finally: The above is only the preferred solution of the present invention and is not used to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for predicting the microseismic time series in a mine, characterized in that, It includes the following steps: Step 1, data preprocessing and primary modal decomposition: After denoising, data cleaning, and standardization preprocessing of the collected original microseismic event sequence, the complete adaptive noise ensemble empirical mode decomposition is used to decompose the original microseismic event sequence into several modal decomposition features; Step 2, sample entropy calculation and sequence reconstruction: The sample entropy is calculated for each modal decomposition feature using the uniform random sampling method, and different modal decomposition features are reconstructed into high-frequency sequences, medium-frequency sequences, and low-frequency sequences by setting the sample entropy threshold; Step 3, secondary modal decomposition: The variational mode decomposition is applied to the high-frequency sequence, medium-frequency sequence, and low-frequency sequence, and the central frequencies under different decomposition numbers are calculated to determine the optimal decomposition number, obtaining the characteristic time series of microseismic data; Step 4, dataset construction: The sliding window method is used to combine the original microseismic event sequence with each modal decomposition feature sequence to construct a multi-feature dataset, forming input and output sample pairs, where the input is the multi-dimensional feature data within the set time window, and the output is the microseismic energy data at the subsequent time; Step 5, time series prediction based on the Autoformer model: The multi-feature dataset is input into the Autoformer model for time series prediction; Step 6, model training and verification: Comparing single-step and multi-step prediction tasks, by adjusting the input sequence length, prediction step, and network hyperparameters, the performance of the Autoformer model is evaluated using mean square error, root mean square error, mean absolute error, mean absolute percentage error, and coefficient of determination indicators.

2. The mine microseismic time series prediction method according to claim 1, wherein In Step 1, the specific steps of data preprocessing and primary modal decomposition are: Step 11, obtaining a microseismic event sequence E during the monitoring period by deploying fiber optic geophones at microseismic detection points L ={(t1,e1),...,(t i ,e i ),...,(t n ,e n ), where t i represents the time point, e i represents the microseismic energy, and n is the total number of microseismic detection points; Step 12, perform denoising, data cleaning, and standardization preprocessing on the data in the microseismic event sequence E L ; Step 13, obtain Δ t The energy within the time of The microseismic energy at the next moment is e t+1 , there exists a mapping function f, and the following relationship is reflected based on the mapping function: Step 14, the original microseismic event sequence is decomposed into several modal decomposition features using the improved complete adaptive noise ensemble empirical mode decomposition, expressed as: E L = {(t1, v 11 , v 12 , …, v 1j , e1), …, (t i , v i1 , v i2 , …, v ij , e i ), …, (t n , v n1 , v n2 , …, v nk , e n )}; where: v ij is the j-th modal decomposition feature of the i-th microseismic detection point, n is the total number of microseismic detection points, and k is the number of modal decomposition features.

3. The microseismic time series prediction method for mine according to claim 1, wherein In Step 3, the secondary modal decomposition is specifically: The modal decomposition number K is initially estimated based on the sample entropy of the high-frequency sequence, medium-frequency sequence, and low-frequency sequence, and according to the central frequency distribution under different modal decomposition numbers, the decomposition number is selected to obtain n modal feature sequences, and the modal feature sequences are combined with the modal decomposition features to obtain n + 1 multi-feature time series sample sets.

4. The mine microseismic time series prediction method according to claim 1, wherein In Step 4, the specific steps of using the sliding window method to combine the original microseismic event sequence with each modal decomposition feature sequence to construct a multi-feature dataset and form input and output sample pairs are: Set the time window Δt and define the sliding step S. Within each sliding window, the first Δt time series values and the corresponding modal feature sequence component data are used as the input features of the model, and at the end of each time window, the data value at the next moment is used as the output.

5. The mine microseismic time series prediction method according to claim 1, characterized in that, In Step 5, the Autoformer model decomposes the characteristic time series into a trend component and a seasonal component; the input features of the encoder and decoder are embedded into a high-dimensional space, and each encoder module consists of multiple encoding layers, responsible for extracting time series features; each layer of the encoder includes an autocorrelation layer, a feed-forward neural network, a Layer Normalization layer, and a Dropout layer.

6. The mine microseismic time series prediction method according to claim 5, characterized in that In step 4 and step 5, each decoder in the Autoformer model consists of multiple decoding layers, where each layer includes an auto-correlation layer and a cross-auto-correlation layer, the interaction between the decoder layer and the encoder layer, and the reconstruction of the trend component and the seasonal component.

Images