Deep learning-based impact dangerous area trajectory prediction visualization method
Through deep learning-based methods, the problem of unintuitive visualization of microseismic data in impact ground pressure monitoring and early warning is solved, and the accurate visualization of the trajectory of microseismic clustering areas is achieved, providing a more reliable reference for impact ground pressure space prediction and prevention and control.
Patent Information
- Application Number
- CN202411002108.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-25
- Publication Date
- 2024-11-26
AI Technical Summary
The existing shock ground pressure monitoring and early warning methods are difficult to effectively visually predict the trajectory of impact hazardous areas, resulting in the stacking and visualization of microseismic data that cannot be intuitive and cannot meet the needs of coal mine safety production.
Using a deep learning-based method, through microseismic dataset segmentation, feature extraction, data preprocessing and time-series data structure, the Patch Embedding layer, MLP module and Bi-LSTM decoder generates future trajectory prediction sequences, and combines rainbow color cards for visualization. Display, intuitively display the moving trajectory of the microseismic gathering area.
Accurate visual display of the moving trajectory of the microseismic cluster area is achieved, data quality and reliability are improved, noise influence is significantly reduced, and the reliability and intuitiveness of impact ground pressure space prediction are enhanced.
Smart Images

Figure CN119026903A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of coal mine monitoring and early warning, and specifically relates to a method for visualizing trajectory prediction of impact danger zones based on deep learning. Background Art
[0002] The monitoring and early warning of rock burst has become a hot topic in the current coal mine safety field. Studies have shown that the clustering area of microseismic data can be used as a precursor feature of the stress state change in the rock burst danger area during coal mining. By using the location and energy of microseismic events, the danger of rock burst can be evaluated, which is conducive to the real-time early warning of rock burst risk. The stress in the microseismic cluster area is relatively concentrated, and the area with concentrated stress is prone to accumulate elastic energy and rock burst. In the mining process, the stress transfer trajectory can be obtained through the trajectory of the rock burst danger area, and then the stress distribution can be determined, which is of great significance for the monitoring and early warning of rock burst in coal mines. At the same time, according to the analysis results, effective safety management of coal mines can be carried out in a targeted manner. At present, the visualization method of trajectory prediction of rock burst danger area needs to be further improved. Therefore, how to effectively realize the visualization prediction of rock burst danger area trajectory is one of the problems that need to be solved in the monitoring and early warning of rock burst.
[0003] Data visualization is a means of presenting data in an easy-to-understand way by using graphics and computer vision and fully combining the characteristics of data. It can express abstract data in a way that is easy for people to understand, and present invisible data as visible graphics, symbols, etc. According to the data type, data visualization can be divided into time series data visualization, network data visualization, hierarchical data visualization, multidimensional data visualization, spatial data visualization, and spatiotemporal data visualization. Data visualization is a graphical presentation based on a deep understanding and analysis of data. By converting complex data into intuitive and visible graphics, colors, and animations, it can help users understand complex phenomena and reveal the inherent laws of data. At present, data visualization has been widely used in many fields such as social media and network analysis, transportation, commercial finance, and healthcare, but there is a lack of research on visualization methods for rock burst danger areas. Microseismic data visualization technology can help mining engineers easily obtain the relationships and laws in microseismic data, and then scientifically and rationally guide coal mining operations. It is one of the indispensable technical means in microseismic monitoring. At this stage, accurate and intuitive microseismic data visualization has become the key to ensuring safe production in coal mines. However, the traditional microseismic data visualization method can no longer meet the current needs of rock burst monitoring and early warning. It has problems such as microseismic data stacking and non-intuitive data visualization. Therefore, it is very necessary to carry out research on the visualization method of microseismic data. Summary of the invention
[0004] In response to the problems existing in the above-mentioned prior art, the present invention provides a deep learning-based method for predicting and visualizing the trajectory of impact-hazardous areas. The method can realize the visualization of the moving trajectory of microseismic clusters based on a continuous and noisy real microseismic data set. The method can provide a reliable technical reference for the spatial prediction and prevention of impact ground pressure.
[0005] In order to achieve the above object, the present invention provides a method for predicting and visualizing the trajectory of an impact danger zone based on deep learning, comprising the following steps:
[0006] Step 1: Segmentation of microseismic dataset;
[0007] The original microseismic data set is segmented according to time periods to obtain multiple segmented microseismic data sets;
[0008] Step 2: Feature extraction;
[0009] For each segmented microseismic data set, a two-dimensional weighted kernel density method is used to estimate the kernel density. The weighting of data points is introduced in the process of calculating the probability density to obtain the probability density estimate of the data point. The data point with the largest probability density estimate is then taken as the center of the microseismic cluster area in this time period.
[0010] Step 3: construct time series data of microseismic clustering areas;
[0011] The calculated data are then smoothed and denoised using data filtering and denoising methods to obtain smoothed and denoised data, and the time series data of the microseismic cluster area are constructed using the smoothed and denoised data.
[0012] Step 4: Data preprocessing;
[0013] The time series data of the microseismic cluster area is segmented, and a sliding window is used to generate an input sequence consisting of l samples and an output sequence consisting of h samples. The sliding window size is set to 1, and the number of generated input sequences X and output sequences Y is n = T-(l+h)+1; the Patch Embedding layer is used to process the input sequence X, and the PatchEmbedding layer is used to segment the input data into small patches, and each small patch is embedded and converted into a token form;
[0014] Step 5: Encode the input sequence X;
[0015] The encoder composed of MLP modules is used to process the input sequence X and output the sequence information X of the hidden state. f,t ;
[0016] Step 6: Sequence information X of hidden statef,t Performing polymerization treatment;
[0017] Select the attention mechanism as the aggregation function, and use the aggregation function to transform the sequence information X of the hidden state f,t Compress to the context representation z and obtain each context vector z j ;
[0018] Step 7: Decode the context representation z to generate an output sequence of future trajectories;
[0019] The decoder selects a Bi-LSTM to summarize the entire encoded sequence X in the compressed context representation z. f,t , and discard irrelevant features while retaining important information, generating and outputting the future prediction sequence of the context representation z of the given input sequence Enable prediction of future impact risk areas;
[0020] Step 8: Visualization of time series data;
[0021] The time series data of microseismic clusters and future prediction sequences The data is sent to the display device, which selects the rainbow color card to visualize the movement trajectory of the microseismic cluster area. From red to purple, the time series of the data is represented. During the display process, arrows are selected to connect the time series data of the microseismic cluster area.
[0022] As a preferred embodiment, in step 2, a probability density estimation value is obtained according to formula (1);
[0023]
[0024] In the formula, represents the probability density estimate at position (x, y), n represents the number of data samples, and h x and h y Represents the bandwidth in the x and y directions respectively, x i and i is the i-th data point, K is the binary kernel function, w i Represents the weight of the i-th data point, and the sum of the weights is 1. In weighted kernel density estimation, the weight of each data point depends on its importance in the kernel density estimation process. For microseismic data, the microseismic energy in the microseismic data is used as the weight, which can better reflect the distribution of real microseismic data.
[0025] As a preference, the kernel function in step 2 is a Gaussian kernel function K(u), and the Gaussian kernel function K(u) is as shown in formula (2);
[0026]
[0027] As a preferred method, in step 3, the moving average method is used to smooth the data, and the moving average MA of the microseismic cluster area at time t is obtained according to formula (3): t , use median filtering to denoise the data, and obtain the filtered output value s according to formula (4) k ;
[0028]
[0029] In the formula, X t represents the central data of the microseismic cluster area, and m represents the moving average window size;
[0030] s k =median(MA t-(k-1) / 2 ,...,MA t-1 ,MA t ,MA t+1 ,...,MA t+(k-1) / 2 ) (4);
[0031] In the formula, (MA t-(k-1) / 2 ,...,MA t-1 ,MA t ,MA t+1 ,...,MA t+(k-1) / 2 ) is the moving average of the microseismic cluster area, k represents the window size, and median represents the function of calculating the median value.
[0032] As a preferred embodiment, in step three, the microseismic concentration area time series data C is obtained according to formula (5);
[0033] C = {S, T} (5);
[0034] Where, S = {s k ; k = 1, ..., T} represents the trajectory sequence, T = {t1, ..., t T} represents the time point, s k is the median filter value (x, y) after data processing.
[0035] As a preferred embodiment, in step 4, the input sequence X is obtained according to formula (6), and the output sequence Y is obtained according to formula (7);
[0036]
[0037] In the formula,
[0038]
[0039] In the formula,
[0040] As a preferred embodiment, in step five, the MLP module includes a time series mixing layer and a feature mixing layer, the time series mixing layer includes data transposition, time series MLP, residual connection and layer normalization, and the feature mixing layer includes data transposition, feature MLP, residual connection and layer normalization; the time series MLP includes 2 linear layers and a GeLU activation function; the feature MLP includes 3 linear layers and 2 GeLU activation functions; the encoder composed of the MLP module is used to process the input sequence X as follows:
[0041] S51: The time series mixing layer processes the input data through formula (8);
[0042] X *,t =Norm(X *,t +W1(GeLU(W1X *,t +b1))+b1) (8);
[0043] In the formula, X *,t Indicates that the temporal mixing layer acts on the data on the time axis, Norm(·) indicates layer normalization, W1 and b1 are the weight and bias in the linear layer in the temporal mixing layer, respectively;
[0044] S52: The feature mixing layer processes the input data through formula (9);
[0045] X f,* =Norm(X f,* +W2(GeLU(W2(GeLU(W2X f,* +b2))+b2))+b2) (9);
[0046] In the formula, X f,* It indicates that the feature mixing layer acts on the data in the feature domain, and W2 and b2 are the weight and bias in the linear layer in the feature mixing layer, respectively.
[0047] As a preferred embodiment, in step six, each context vector z is obtained by formula (10): j ;
[0048]
[0049] In the formula, e jt =Align(u j-1 ,h t ); Align represents the hidden state u of the decoder layer j-1 and the hidden state X of the encoder layer f,t Alignment calculation between α jt Indicates that the softmax operator is used to transform e jtThe calculated attention weight uses the softmax function to convert the vector e of length l jt Normalize to attention mask input time series;
[0050] As a preferred embodiment, in step seven, the future prediction sequence is obtained according to formula (11):
[0051]
[0052] In the formula, u j is the output of Bi-LSTM, u j =BiLSTM(μ t ,u t-1 θ D ); μ t is the input vector, Where u0 = tanh(W k X f,t +b k ), u0 is used for the initial decoder state, W k and b k They are the weights and biases of the trainable parameters of Bi-LSTM respectively; is the previous predicted state, where θ D is the trainable parameter of Bi-LSTM; W y and b y is to output u from Bi-LSTM j Mapping to the next predicted state The weights and biases.
[0053] In the present invention, the microseismic data set is first divided into time periods, which can better capture the trend of data changes over time. Smoothing and denoising the data can effectively improve the data quality and significantly reduce the impact of noise on data analysis and application. Introducing weighting of data points in the process of calculating probability density can more flexibly adapt to the importance of different data points, greatly improving the accuracy of density estimation. Smoothing and denoising the central data of the microseismic cluster area obtained can effectively reduce the noise and outliers caused by microseismic monitoring errors, microseismic monitoring equipment failures, etc., effectively improve the quality and reliability of the data, and can also reduce the volatility of the data, making the data easier to understand and analyze, and making the visualization results more intuitive. Using the Patch Embedding layer to process the input sequence X can effectively reduce the calculation time and better capture the local information in the physical indicator image. Converting the input sequence X into a token form can facilitate the subsequent network to better extract local information of the data. Using the aggregation function to encode the sequence information X of all hidden states encoded in the encoder f,tCompressing to the context representation z can more accurately select the relevant encoder hidden states at all time steps for prediction, which can significantly improve the prediction performance. By selecting Bi-LSTM for decoding, the entire encoded sequence X can be obtained through the context representation z f,t , and thus accurately obtain the future prediction sequence Selecting color changes to represent the changes in the data in time series makes it easier to identify and compare data at different time points, making the visualization results more intuitive, and thus intuitively showing the trend of data changes over time. Selecting arrows to connect the time series data of microseismic clusters can more clearly show the movement trajectory of the impact danger zone.
[0054] The present invention can realize the visualization display of the movement trajectory of the microseismic cluster area based on a continuous and noisy real microseismic data set. This method can provide a reliable technical reference for the spatial prediction and prevention of rock burst. BRIEF DESCRIPTION OF THE DRAWINGS
[0055] Figure 1 It is a schematic diagram of the process of the present invention;
[0056] Figure 2 is a flow chart of the microseismic data processing process in the present invention;
[0057] Figure 3 It is a schematic diagram of the median filtering process in the present invention;
[0058] Figure 4 is a schematic diagram of the before and after comparison of data smoothing and denoising processing in the present invention;
[0059] Figure 5 It is a schematic diagram of the visualization result of the movement trajectory of the microseismic concentration area in the present invention;
[0060] Figure 6 It is a flow chart of the process of trajectory prediction through microseismic clustering area time series data in the present invention;
[0061] Figure 7 It is the architecture diagram of the MLP model in the present invention;
[0062] Figure 8 is the architecture diagram of the sequential MLP in the present invention;
[0063] Fig. 9 It is the architecture diagram of the characteristic MLP in the present invention. DETAILED DESCRIPTION
[0064] The present invention will be further described below in conjunction with examples, but the present invention is not limited to the following examples.
[0065] like Figures 1 to 9As shown, the present invention provides a method for predicting and visualizing the trajectory of an impact danger zone based on deep learning, comprising the following steps:
[0066] Step 1: Segmentation of microseismic dataset;
[0067] like Figure 2 As shown, the original microseismic data set is segmented according to time periods to obtain multiple segmented microseismic data sets, which can better capture the trend of data changes over time;
[0068] Step 2: Feature extraction;
[0069] Since microseismic data is continuous and noisy, a two-dimensional weighted kernel density method is used to estimate the kernel density of each segmented microseismic data set, which can effectively extract features. Introducing weighting of data points in the process of calculating probability density can more flexibly adapt to the importance of different data points, greatly improving the accuracy of density estimation, and then accurately obtaining the probability density estimate of the data point, and then taking the data point with the largest probability density estimate as the center of the microseismic cluster area in this time period.
[0070] Step 3: construct time series data of microseismic clustering areas;
[0071] In order to improve data quality and reduce the impact of noise on data analysis and application, the calculated data is smoothed and denoised using data filtering and denoising methods to obtain smoothed and denoised data, and the time series data of the microseismic cluster area is constructed with the smoothed and denoised data.
[0072] Smoothing and denoising the central data of the microseismic cluster area can reduce the noise and outliers caused by microseismic monitoring errors, microseismic monitoring equipment failures, etc., improve the quality and reliability of the data, and reduce the volatility of the data, making the data easier to understand and analyze, and the visualization results more intuitive.
[0073] Step 4: Data preprocessing;
[0074] like Figure 6As shown in the figure, in order to directly generate the output sequence of the future state from the given input sequence, the trajectory prediction problem is formulated as a supervised learning process of sequence data, and the entire trajectory (i.e., time series) is segmented into smaller, fixed-length sequences using data segmentation. Specifically, the time series data of the microseismic cluster area is segmented, and a sliding window is used to generate an input sequence consisting of l samples and an output sequence consisting of h samples. The sliding window size is set to 1, and the number of input sequences X and output sequences Y generated is n = T-(l+h)+1; in order to reduce the calculation time and better capture the local information in the physical indicator image, the Patch Embedding layer is used to process the input sequence X, and the Patch Embedding layer is used to segment the input data into small patches, and each small patch is embedded and converted into a token form, so that the subsequent network can better extract the local information of the data.
[0075] Step 5: Encode the input sequence X;
[0076] The encoder composed of MLP (feature perceptron) modules is used to process the input sequence X and output the sequence information X of the hidden state. f,t ;
[0077] Step 6: Sequence information X of hidden state f,t Performing polymerization treatment;
[0078] The attention mechanism is selected as the aggregation function, and the aggregation function is used to convert the sequence information X of all hidden states encoded in the encoder f,t Compress to the context representation z and obtain each context vector z j ;
[0079] The decoder will use this representation z to generate an output sequence of length h, with the goal of summarizing (downsampling) the entire encoded sequence X in the compressed context representation z f,t , while retaining important information, discarding irrelevant features. The main purpose of the attention mechanism is to generate prediction trajectories based on the decoder’s hidden state u j-1 For each context vector z j This operation is equivalent to assigning a probability of being noticed by the decoder to each context vector, which can more accurately select the relevant encoder hidden states at all time steps for prediction, thereby improving the prediction performance.
[0080] Step 7: Decode the context representation z to generate an output sequence of future trajectories;
[0081] The decoder D aims to generate future trajectories given the context representation z of the input sequence. Formally, the autoregressive decoder is trained to j And predict the next state given all previously predicted states Y Each context representation z is fed back into the decoder, which selects a Bi-LSTM (bidirectional long short-term neural network) to summarize the entire encoded sequence X in the compressed context representation z f,t , and discard irrelevant features while retaining important information, generate and output the output sequence of the future state of the context representation z of the given input sequence, and the output sequence of the future state is used as the future prediction sequence Enables prediction of future impact risk areas.
[0082] Step 8: Visualization of time series data;
[0083] Selecting color changes to represent the changes in data in time series makes it easier to identify and compare data at different time points, making the visualization results more intuitive. The selected colors should have sufficient contrast during the display process to ensure that users can clearly distinguish different data and time changes. At the same time, the selected colors should also be coordinated to avoid visual confusion and problems that affect data reading.
[0084] The time series data of microseismic clusters and future prediction sequences The data is sent to the display device, which selects the rainbow color card to visualize the movement trajectory of the microseismic clustering area. The color range from red to purple represents the time series of the data. In this way, the trend of data changes over time can be intuitively displayed. At the same time, in order to more clearly display the movement trajectory of the impact danger zone, during the display process, arrows are selected to connect the time series data of the microseismic clustering area.
[0085] Figure 5 The visualization results of the impact hazard zone trajectory of the Wenjiaopo 4106 working face from October 2022 to October 2023 are shown. The results show that the visualization method proposed in this application can more clearly display the movement trajectory of the microseismic aggregation area.
[0086] As a preferred method, in step 2, the probability density estimate is obtained according to formula (1), and finally, the probability density estimate is obtained by selecting the microseismic data after daily aggregation: The maximum position (x, y) is taken as the center of the microseismic clustering area of the microseismic data of that day;
[0087]
[0088] In the formula, represents the probability density estimate at position (x, y), n represents the number of data samples, and h xand h y Represents the bandwidth in the x and y directions respectively, x i and i is the i-th data point, K is the binary kernel function, w i Represents the weight of the i-th data point, and the sum of the weights is 1.
[0089] In weighted kernel density estimation, the weight of each data point depends on its importance in the kernel density estimation process. For microseismic data, the microseismic energy in the microseismic data is used as the weight, which can better reflect the distribution of real microseismic data.
[0090] As a preference, the kernel function in step 2 is a Gaussian kernel function K(u), and the Gaussian kernel function K(u) is as shown in formula (2);
[0091]
[0092] As one of the most commonly used kernel functions, the Gaussian kernel function has good properties in continuity and smoothness, and can make a relatively smooth estimate of the probability density.
[0093] The moving average method is a commonly used data smoothing method. It is a method that gradually moves along the time series and calculates the time series average containing a certain number of items in sequence to reflect the long-term trend. The principle of this method is to slide a fixed-size window in sequence and average the data in the window. By calculating the average value, the data becomes more stable and the long-term trend can be highlighted. The moving average method has been widely used in statistics and time series analysis, including the simple moving average method, the weighted moving average method, and the trend moving average method. In view of the fact that the development trend of microseismic data has not changed much, the simple moving average method was selected. Specifically, in step three, the moving average method is used to smooth the data, and the moving average MA of the microseismic cluster area at time t is obtained according to formula (3): t ;
[0094]
[0095] In the formula, X t represents the central data of the microseismic cluster area, and m represents the moving average window size;
[0096] Median filtering is a common digital signal processing method used to remove noise from images, audio, or other digital signals. This method uses the middle value of the data in the window to replace the original data value to reduce the impact of noise, help smooth the signal and retain the details in the signal, such as Figure 3 Specifically, median filtering is used to denoise the data, and the filtered output value s is obtained according to formula (4): k ;
[0097] s k =median(MA t-(k-1) / 2 ,...,MA t-1 ,MA t ,MA t+1 ,...,MA t+(k-1) / 2 ) (4);
[0098] In the formula, (MA t-(k-1) / 2 ,...,MA t-1 ,MA t ,MA t+1 ,...,MA t+(k-1) / 2 ) is the moving average of the microseismic cluster area, k represents the window size, and median represents the function of calculating the median value. Figure 4 In order to compare the visualization results of the microseismic clustering area data before and after smoothing and denoising, the results show that the visualization effect of the data after denoising and smoothing is better and can better reflect the movement trajectory of the microseismic clustering area.
[0099] As a preferred embodiment, in step three, the microseismic concentration area time series data C is obtained according to formula (5);
[0100] C = {S, T} (5);
[0101] Where, S = {s k ; k = 1, ..., T} represents the trajectory sequence, T = {t1, ..., t T} represents the time point, s k is the median filter value (x, y) after data processing.
[0102] As a preferred embodiment, in step 4, the input sequence X is obtained according to formula (6), and the output sequence Y is obtained according to formula (7);
[0103]
[0104] In the formula,
[0105]
[0106] In the formula,
[0107] like Figure 7 As shown, in step 5, the MLP module includes a time series mixing layer and a feature mixing layer. The time series mixing layer includes data transposition, time series MLP, residual connection and layer normalization. The feature mixing layer includes data transposition, feature MLP, residual connection and layer normalization. The time series MLP includes 2 linear layers and a GeLU activation function, as shown in Figure 8As shown; the feature MLP contains 3 linear layers and 2 GeLU activation functions, as shown Fig. 9 As shown;
[0108] The encoder composed of MLP modules processes the input sequence X as follows:
[0109] S51: The input sequence X outputs tokens through the PatchEmbedding layer, which can be regarded as data with time dimension and feature dimension. In order to alleviate the gradient vanishing problem, simplify network training and promote network learning effect, the data is processed by the time series MLP and then connected through residual connection. In order to make the neural network more robust to the distribution changes of input data, improve the training stability of the model, accelerate convergence and prevent overfitting, the data after residual connection is layer normalized, and the time series mixing layer processes the input data through formula (8);
[0110] X *,t =Norm(X *,t +W1(GeLU(W1X *,t +b1))+b1) (8);
[0111] In the formula, X *,t Indicates that the temporal mixing layer acts on the data on the time axis, Norm(·) indicates layer normalization, W1 and b1 are the weight and bias in the linear layer in the temporal mixing layer, respectively;
[0112] S52: In order to analyze the correlation of data in different feature domains, the output of the time series mixing layer is transposed before entering the feature mixing layer. The transposed data is connected using residual connection after passing through the feature mixing layer, and then processed by layer normalization. The result is transposed again and then output. The feature mixing layer processes the input data according to formula (9);
[0113] X f,* =Norm(X f,* +W2(GeLU(W2(GeLU(W2X f,* +b2))+b2))+b2) (9);
[0114] In the formula, X f,* It indicates that the feature mixing layer acts on the data in the feature domain, and W2 and b2 are the weight and bias in the linear layer in the feature mixing layer, respectively.
[0115] As a preferred embodiment, in step 7, consider the hidden state X of the encoder f,* The sequence and the hidden state u of the decoder network j-1 , then each context vector z j It can be calculated as the weighted sum of the hidden states. Specifically, each context vector z is obtained by formula (10):j ;
[0116]
[0117] In the formula, e jt =Align(u j-1 ,h t ); Align represents the hidden state u of the decoder layer j-1 and the hidden state X of the encoder layer f,t Alignment calculation between α jt Indicates that the softmax operator is used to transform e jt The calculated attention weight uses the softmax function to convert the vector e of length l jt Normalized to the attention mask input time series; z j is the final state of the attention layer;
[0118] As a preferred embodiment, in step seven, the future prediction sequence is obtained according to formula (11): In this process, the length of the output sequence obtained is h;
[0119]
[0120] In the formula, u j is the output of Bi-LSTM, u j =BiLSTM(μ t ,u t-1 θ D ); μ t is the input vector, Where u0 = tanh(W k X f,t +b k ), u0 is used for the initial decoder state, W k and b k They are the weights and biases of the trainable parameters of Bi-LSTM respectively; is the previous predicted state, where is the first prediction sequence; θ D is the trainable parameter of Bi-LSTM; W y and b y is to output u from Bi-LSTM j Mapping to the next predicted state The weights and biases.
Claims
1. A deep learning-based method for predicting the trajectory of impact danger zones. It is characterized in that The steps include: Step 1: Segmentation of microseismic dataset; The original microseismic data set is segmented according to time periods to obtain multiple segmented microseismic data sets; Step 2: Feature extraction; For each segmented microseismic data set, a two-dimensional weighted kernel density method is used to estimate the kernel density. The weighting of data points is introduced in the process of calculating the probability density to obtain the probability density estimate of the data point. The data point with the largest probability density estimate is then taken as the center of the microseismic cluster area in this time period. Step 3: construct time series data of microseismic clustering areas; The calculated data are then smoothed and denoised using data filtering and denoising methods to obtain smoothed and denoised data, and the time series data of the microseismic cluster area are constructed using the smoothed and denoised data. Step 4: Data preprocessing; The time series data of the microseismic cluster area is segmented, and a sliding window is used to generate an input sequence consisting of l samples and an output sequence consisting of h samples. The sliding window size is set to 1, and the number of generated input sequences X and output sequences Y is n = T-(l+h)+1; the input sequence X is processed by the Patch Embedding layer, and the input data is segmented into small patches using the Patch Embedding layer, and each small patch is embedded and converted into a token form; Step 5: Encode the input sequence X; The encoder composed of MLP modules is used to process the input sequence X and output the sequence information X of the hidden state. f,t ; Step 6: Sequence information X of hidden state f,t Performing polymerization treatment; Select the attention mechanism as the aggregation function, and use the aggregation function to transform the sequence information X of the hidden state f,t Compress to the context representation z and obtain each context vector z j ; Step 7: Decode the context representation z to generate an output sequence of future trajectories; The decoder selects a Bi-LSTM to summarize the entire encoded sequence X in the compressed context representation z. f,t , and discard irrelevant features while retaining important information, generating and outputting the future prediction sequence of the context representation z of the given input sequence Enable prediction of future impact risk areas; Step 8: Visualization of time series data; The time series data of microseismic clusters and future prediction sequences The data is sent to the display device, which selects the rainbow color card to visualize the movement trajectory of the microseismic cluster area. From red to purple, the temporal nature of the data is represented. During the display process, arrows are selected to connect the temporal data of the microseismic cluster area.
2. According to claim 1, a method for visualizing trajectory prediction of impact danger zone based on deep learning, It is characterized in that In step 2, the probability density estimate is obtained according to formula (1); In the formula, represents the probability density estimate at position (x, y), n represents the number of data samples, and h x and h y Represents the bandwidth in the x and y directions respectively, x i and i is the i-th data point, K is the binary kernel function, w i Represents the weight of the i-th data point, and the sum of the weights is 1.
3. According to claim 2, a method for predicting and visualizing the impact danger zone trajectory based on deep learning, It is characterized in that The kernel function in step 2 is a Gaussian kernel function K(u), and the Gaussian kernel function K(u) is shown in formula (2); 4. According to claim 3, a method for predicting and visualizing the trajectory of an impact danger zone based on deep learning, It is characterized in that In step 3, the moving average method is used to smooth the data, and the moving average MA of the microseismic cluster area at time t is obtained according to formula (3): t , use median filtering to denoise the data, and obtain the filtered output value s according to formula (4) k ; Where, X t represents the central data of the microseismic cluster area, and m represents the moving average window size; S k =median(MA t-(k-1) / 2 ,...,ME t-1 ,ME t ,ME t+1 ,...,ME t+(k-1) / 2 ) (4); In the formula, (MA t-(k-1) / 2 ,...,MA t-1 ,MA t ,MA t+1 ,...,MA t+(k-1) / 2 ) is the moving average of the microseismic cluster area, k represents the window size, and median represents the function of calculating the median value.
5. According to claim 4, a method for predicting and visualizing the trajectory of an impact danger zone based on deep learning, It is characterized in that In step 3, the time series data C of the microseismic clustering area is obtained according to formula (5); C = {S, T} (5); Where, S = {s k ; k = 1, ..., T} represents the trajectory sequence, T = {t 1 ,...,t T } represents the time point, s k is the median filter value (x, y) after data processing.
6. According to claim 5, a method for predicting and visualizing the trajectory of an impact danger zone based on deep learning, It is characterized in that In step 4, the input sequence X is obtained according to formula (6), and the output sequence Y is obtained according to formula (7); In the formula, In the formula, 7. According to claim 1, a method for predicting and visualizing the trajectory of an impact danger zone based on deep learning, It is characterized in that In step 5, the MLP module includes a time series mixing layer and a feature mixing layer. The time series mixing layer includes data transposition, time series MLP, residual connection and layer normalization. The feature mixing layer includes data transposition, feature MLP, residual connection and layer normalization. The time series MLP includes 2 linear layers and a GeLU activation function. The feature MLP includes 3 linear layers and 2 GeLU activation functions. The encoder composed of the MLP module is used to process the input sequence X as follows: S51: The time series mixing layer processes the input data through formula (8); X *,t =Norm(X *,t +W 1 (GeLU(W 1 X *,t +b 1 ))+b 1 ) (8); Where, X *,t Indicates that the temporal mixing layer acts on the data on the time axis, Norm(·) indicates layer normalization, and W 1 and b 1 They are the weights and biases in the linear layer in the temporal mixing layer; S52: The feature mixing layer processes the input data through formula (9); X f,* =Norm(X f,* +W 2 (GeLU(W 2 (GeLU(W 2 X f,* +b 2 ))+b 2 ))+b 2 ) (9); Where, X f,* Indicates that the feature mixing layer acts on the data in the feature domain, W 2 and b 2 are the weights and biases in the linear layer in the feature mixing layer, respectively.
8. According to claim 1, a method for predicting and visualizing impact danger zone trajectory based on deep learning, It is characterized in that In step 6, each context vector z is obtained by formula (10) j ; In the formula, e jt =Align(u j-1 ,h t ); Align represents the hidden state u of the decoder layer j-1 and the hidden state X of the encoder layer f,t Alignment calculation between α jt Indicates that the softmax operator is used to transform e jt The calculated attention weight uses the softmax function to convert the vector e of length l jt Normalize the input time series to the attention mask.
9. The method for predicting and visualizing the impact danger zone trajectory based on deep learning according to claim 1, It is characterized in that In step 7, the future prediction sequence is obtained according to formula (11): In the formula, u j is the output of Bi-LSTM, u j =BiLSTM(μ t ,u t-1 θ D ); μ t is the input vector, Among them, u 0 =tanh(W k X f,t +b k ),u 0 For the initial decoder state, W k and b k They are the weights and biases of the trainable parameters of Bi-LSTM respectively; is the previous predicted state, where θ D is the trainable parameter of Bi-LSTM; W y and b y is to output u from Bi-LSTM j Mapping to the next predicted state The weights and biases.
