Coal mine micro-seismic time sequence prediction method and system for simulating waveform forward modeling and seismic source inversion process
By combining the CNN-BiLSTM-Attention model with traditional inversion methods, a generative autoregressive neural network is used to perform forward modeling of simulated waveforms and source inversion. This solves the problems of wave field distortion and limited accuracy of traditional methods in heterogeneous structures, and achieves high-precision microseismic time series prediction in coal mines.
Patent Information
- Application Number
- CN202511330037.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-09-17
- Publication Date
- 2025-11-11
AI Technical Summary
Traditional coal mine microseismic monitoring methods struggle to characterize the wave field distortion effect of heterogeneous structures when simulating microseismic waveform signals. Furthermore, the accuracy of inversion methods is limited by velocity model errors and algorithm convergence, resulting in poor time series prediction performance.
The CNN-BiLSTM-Attention model is combined with traditional inversion methods. Generative autoregressive neural networks are used to perform forward modeling of simulated waveforms and source inversion. End-to-end training is performed using auxiliary training data and joint loss functions to extract deep spatiotemporal features of microseismic waveforms and predict source location and energy.
It has achieved high-precision microseismic time series prediction under heterogeneous geological conditions, improving the accuracy of microseismic risk modeling and real-time monitoring capabilities in coal mines.
Smart Images
Figure CN120928431A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of coal mine safety monitoring, and in particular to a method and system for predicting coal mine microseismic time series by simulating waveform forward modeling and source inversion processes. Background Technology
[0002] Coal mine microseismic monitoring, as a non-contact technology, uses a network of sensors deployed on the surface or underground to remotely and in real time sense the elastic wave signals generated by the rock mass fracture source. This allows for the determination of the location and parameter attributes of disasters such as rockbursts and roof collapses, enabling the modeling and risk prediction of the spatiotemporal evolution of microseismic events. It is a core means of early warning of disasters such as rockbursts and roof collapses.
[0003] Forward modeling and inversion modeling are two complementary geoscientific research methods. In coal mine microseismic monitoring, forward modeling refers to calculating and simulating the corresponding simulated microseismic waveform signal based on prior geological information and certain initial source conditions and seismic wave boundary conditions. Traditional forward modeling methods are based on the assumption of a homogeneous medium, which makes it difficult to characterize the wave field distortion effects of heterogeneous structures such as coal mine bedding and fracture zones. Inversion modeling, on the other hand, uses appropriate models and methods to infer the corresponding source conditions based on the monitored real microseismic waveform data and phenomena. The accuracy of traditional inversion methods is limited by the velocity model error and the convergence of the inversion algorithm.
[0004] As one of the main research tasks in machine learning, time series prediction is used to predict data for subsequent time steps by learning the spatiotemporal correlation of the preceding time steps of coal mine microseismic waveforms and source data. Recurrent neural networks such as LSTM and GRU can capture the non-steady-state abrupt changes of coal mine dynamic disasters due to their long-range dependency modeling capabilities, realize end-to-end model training and convenient deployment, avoid tedious feature engineering of the original data, and improve the spatiotemporal modeling effect of coal mine microseismic risk. Summary of the Invention
[0005] To achieve the above objectives, this invention provides a method for predicting the time series of microseismic events in coal mines by simulating the waveform forward modeling and source inversion process. The specific technical solution is as follows:
[0006] The collected raw microseismic waveforms, geological data of the working face, and auxiliary training data are processed to generate microseismic waveform data to be trained and target microseismic waveform data.
[0007] The process of inverting simulated source parameters from microseismic waveform signals to be trained involves receiving microseismic waveform data to be trained as input and outputting an intermediate vector representing the simulated source location and energy.
[0008] The process of forward modeling the simulated microseismic waveform signal from the simulated source parameters involves receiving the intermediate vector, predicting the target source vector, and finally outputting the simulated microseismic waveform data.
[0009] Furthermore, the auxiliary training data includes simulated source data for preceding and subsequent time steps;
[0010] The simulated seismic source data of the preceding time step is obtained from the microseismic waveform data to be trained using traditional inversion methods based on the geological data of the working face, including the simulated seismic source location and energy.
[0011] The simulated source data for the subsequent time step is obtained from the target microseismic waveform data using traditional inversion methods, which simulates the source location and energy.
[0012] Furthermore, the traditional inversion method includes using the P-wave first arrival method to determine the time difference of the P-wave arriving at each detector, combining the wave velocity field conditions to locate the three-dimensional coordinates of the source, and calculating the source energy through the duration of the seismic phase.
[0013] Furthermore, the process of inverting and simulating source parameters adopts the CNN-BiLSTM-Attention model, which includes a convolutional neural network (CNN) for extracting deep features of the microseismic waveform to be trained, a bidirectional long short-term memory network (BiLSTM) for learning the bidirectional time dependency of the deep feature sequence, and an attention mechanism responsible for calculating the similarity between the knowledge vector and the source vector to be trained, and outputting an intermediate vector.
[0014] Furthermore, the attention mechanism uses only vectors, including external knowledge vectors and internal knowledge vectors;
[0015] The external knowledge vector is the simulated seismic source data of the preceding time step;
[0016] The internal knowledge vector is the target seismic source vector output by the decoder forward modeling module.
[0017] Furthermore, the forward modeling process of the microseismic waveform signal adopts an LSTM-MLP model, which includes a Long Short-Term Memory (LSTM) network responsible for predicting the target source vector based on the intermediate vector output by the encoder inversion module, and a Multilayer Perceptron (MLP) responsible for mapping the target source vector into simulated microseismic waveform data.
[0018] Furthermore, the process of inverting and simulating source parameters and the process of forward simulating microseismic waveform signals constitute a generative autoregressive neural network as the backbone model, and the backbone model is trained using a joint loss function.
[0019] The joint loss function includes a first loss function and a second loss function;
[0020] The first loss function calculates the error between the target source vector predicted by the Long Short-Term Memory (LSTM) network and the simulated source data of the subsequent time steps during the forward modeling of the microseismic waveform signal.
[0021] The second loss function calculates the error between the simulated microseismic waveform output by the multilayer perceptron (MLP) and the target microseismic waveform data during the forward modeling of the microseismic waveform signal.
[0022] A coal mine microseismic time series prediction system simulating waveform forward modeling and source inversion process is provided. It is used to implement the above-mentioned coal mine microseismic time series prediction method simulating waveform forward modeling and source inversion process. The system includes hardware equipment and software system running on it.
[0023] The hardware device includes a processor, memory, and input and output interfaces.
[0024] The software system includes a data preprocessing module, an encoder inversion module, and a decoder forward modeling module:
[0025] The data preprocessing module is used to process the collected raw microseismic waveforms, working face geological data and auxiliary training data to generate microseismic waveform data to be trained and target microseismic waveform data.
[0026] The decoder inversion module is used to invert the simulated source parameters from the microseismic waveform signal to be trained. It receives the microseismic waveform data to be trained as input and outputs an intermediate vector representing the simulated source location and energy.
[0027] The decoder forward modeling module is used to perform forward modeling of the simulated microseismic waveform signal from the simulated source parameters, receive the intermediate vector, predict the target source vector, and finally output the simulated microseismic waveform data.
[0028] The video stream acquisition and standardization processing module is used to synchronously extract image frame sequences and corresponding first short-time audio sequences based on real-time video streams, and construct multimodal frame sequences.
[0029] Furthermore, the input interface is connected to a detector or microseismic monitoring system to receive real-time signals, the processor implements the prediction method by running a computer program, and the memory stores the computer program, coal mine geological data, and microseismic data.
[0030] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0031] This invention provides a method and system for predicting coal mine microseismic time series by simulating waveform forward modeling and source inversion processes. Its data preprocessing module can use the simulated source data of the preceding and following time steps obtained by traditional inversion methods for the auxiliary training of the generative autoregressive neural network, which helps the backbone model to achieve end-to-end training and convenient deployment. It is suitable for designing and developing practical coal mine microseismic real-time monitoring systems.
[0032] This invention provides a method and system for predicting the time series of microseismic events in coal mines by simulating the waveform forward modeling and source inversion process. Its encoder inversion module can extract the deep spatiotemporal characteristics of microseismic events from the original microseismic waveform data, and make fuller use of the microseismic waveform data through an external attention mechanism to simulate the source inversion process under the geological conditions of the working face.
[0033] This invention provides a method and system for predicting the time series of microseismic events in coal mines by simulating the waveform forward modeling and source inversion process. Its decoder forward modeling module can predict and simulate microseismic waveform signals based on the deep spatiotemporal characteristics of microseismic events, and make fuller use of the target microseismic waveform data by using a joint loss function to simulate the waveform forward modeling process under the geological conditions of the working face. Attached Figure Description
[0034] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0035] Figure 1 This is an overall flowchart of a coal mine microseismic time series prediction method based on the simulation of waveform forward modeling and source inversion process according to the present invention;
[0036] Figure 2 This is a flowchart of the coal mine microseismic time-series prediction method of the present invention;
[0037] Figure 3 This is a structural diagram of the generative autoregressive neural network model of the encoder-decoder architecture of the present invention;
[0038] Figure 4 This is a schematic diagram of the hardware and software system of the present invention;
[0039] Figure 5 This is a functional module diagram of a coal mine microseismic time series prediction system that simulates waveform forward modeling and source inversion processes according to the present invention. Detailed Implementation
[0040] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0041] Example 1
[0042] Please see Figure 1As shown, this embodiment provides a method for predicting the time series of microseismic events in coal mines by simulating the waveform forward modeling and source inversion process, including:
[0043] Step 1: Coal mine data acquisition and microseismic data processing, which includes raw microseismic waveform signal acquisition, working face geological data acquisition, and data preprocessing modules.
[0044] The acquisition of raw microseismic waveform signals in coal mine data mainly refers to the acquisition of raw microseismic waveform signals from the underground working face using a coal mine microseismic monitoring system within a preset time period, and the real-time acquisition of key parameters such as magnitude distribution, event frequency, waveform amplitude, frequency, and arrival time difference for subsequent feature calculations.
[0045] Here, magnitude distribution refers to obtaining the event magnitude M and classifying it into categories, recording the total number of categories L, and the category number l = 1, 2, ..., L, then the magnitude threshold M for each category is determined. l Event frequency refers to the total number N of microseismic events within a statistical time period T, and the number N of events corresponding to each magnitude level. l .
[0046] Among them, the magnitude of microseismic events can be divided into initial magnitude, average magnitude, and maximum magnitude.
[0047] Specifically, geological data of the working face also needs to be collected, including structural outline map of the mining area, topographic map, engineering geological profile map, comprehensive stratigraphic columnar section, coal and rock strata comparison map, etc., to describe the geological structure of the mining area, coal seam distribution and lithological characteristics, as well as shaft and tunnel construction calibration records, equipment installation and acceptance reports, and as-built drawings.
[0048] The data preprocessing module is responsible for cleaning the input raw microseismic waveform data, geological data of the working face, and auxiliary training data. Specifically, it includes missing value processing and error value repair, as well as training and test set partitioning, data standardization, and sliding window sample generation, so as to facilitate input into the neural network model.
[0049] The training set and the test set are divided in an 8:2 ratio to obtain microseismic waveform data to be trained and target microseismic waveform data.
[0050] The standardization method employs maximum-minimum normalization to linearly map the microseismic waveform data to be trained to [0,1], preserving the original distribution shape of the microseismic data and preventing the values from exceeding the range suitable for input into the neural network. Then, the normalization parameters of the training set data are used to fit the test set data to prevent leakage of the test set data and to maintain the original time order, ensuring that the test set fully represents future data.
[0051] Specifically, a fixed-length sliding window is used to extract sequence segments composed of original microseismic waveform data, forming an input-output sample pair sequence that can be trained by the model, consisting of input features and labels. The input sample set is the microseismic waveform data to be trained, and the output sample set is the target microseismic waveform data.
[0052] In some embodiments, such as Figure 2 and Figure 3 As shown, a flowchart of a method for predicting the time series of microseismic events in coal mines by simulating the forward modeling of waveforms and the inversion of seismic sources is presented.
[0053] The data preprocessing module, after processing the original microseismic waveform data w, generates both the input-output sample pair sequence s and auxiliary training data g and x.
[0054] Wherein, for the input-output sample pair sequence s=[s0,s1,s2,…,s…] t-1 ,s t ,…,s n-1 ,s n If we use data from the past W steps to predict data from the future H steps, then the input-output pair generated by the sliding window is:
[0055] X t =[s t-W ,s t-W+1 ,...,s t-1 ];
[0056] Y t =[y t ,y t+1 ,...,y t+H-1 ];
[0057] The sliding window size W determines the length of historical data contained in each sample, and the prediction step size H controls the number of future data points that need to be predicted.
[0058] In the auxiliary training data, g represents the simulated source data for the preceding time step, which is the simulated source location and energy obtained from the microseismic waveform data sin=[s0,s1,…,st-1,…,sn-1] using the traditional inversion method based on the geological data of the working face; x represents the simulated source data for the subsequent time step, which is the simulated source location and energy obtained from the target microseismic waveform data sout=[s1,s2,…,st,…,sn] using the traditional inversion method.
[0059] Traditional inversion methods include using the P-wave first arrival method to determine the time difference of P-wave arrival at each detector, locating the three-dimensional coordinates of the source in combination with specific wave velocity field conditions, and calculating the source energy through the duration of the seismic phase.
[0060] In some embodiments, a method for predicting the time series of microseismic events in coal mines by simulating the waveform forward modeling and source inversion process is provided, including a generative autoregressive neural network model structure diagram of an encoder-decoder architecture.
[0061] Specifically, the encoder inversion module is responsible for simulating the source location and energy from the Sin inversion of the microseismic waveform data to be trained. This module adopts a CNN-BiLSTM-Attention composite deep learning architecture to realize multivariate time series prediction of coal mine microseismic events.
[0062] Among them, CNN, or Convolutional Neural Network, extracts deep features such as fluctuation trends and periodic patterns from the micro-seismic waveform time series data to be trained through one-dimensional convolutional kernels. It adopts multi-scale convolutional kernels in parallel operation to enhance its adaptability to non-stationary signals. Pooling layers compress feature dimensions to improve noise resistance. The convolution calculation formula used in the CNN layer is as follows:
[0063] c t =f(W conv *x t +b conv );
[0064] Among them, W conv The weights are the kernel weights, * indicates the convolution operation, b conv As the bias, f(·) is the ReLU activation function, x t Let c be the input vector. t This is the output vector.
[0065] BiLSTM, or Bidirectional Long Short-Term Memory Network, captures the bidirectional dependencies between microseismic time-series data processed by CNN. Its forward LSTM learns historical information, and its backward LSTM learns future information. Finally, through splicing or weighted fusion, it outputs bidirectional data, overcoming the limitations of the traditional unidirectional information flow of LSTM, thus effectively identifying the spatiotemporal evolution of rock mass fracturing events. The gating equations used in the LSTM layers are as follows:
[0066]
[0067] h t =o t ⊙tanh(C t );
[0068] Among them, W f W i W o as well as Let b be the weights of the forget gate, input gate, output gate, and candidate state, respectively, and let b be their corresponding bias vectors. f b i b o , The activation function σ is the Sigmoid function, and the Hadamard product ⊙ is multiplied element by element, [x t ,h t-1 [] represents the splicing operation, x t h is the input vector t In hidden state, For candidate cell states, C t For cell states, with respect to the preceding time step, they are h t-1 and C t-1 .
[0069] In this embodiment, the Attention module used is a multi-head attention mechanism. It calculates the similarity between the internal and external knowledge vectors and the source vector to be trained using two attention mechanisms, respectively. After normalization, the attention weights are obtained. Then, the outputs of the two attention mechanisms are concatenated. The final intermediate vector z output by the module represents the simulated source location and energy. The calculation formula of the attention mechanism is as follows:
[0070]
[0071] Wherein, the input matrix X and the trainable parameter matrix W Q W K W V Performing linear transformations respectively yields Q, K, and V, which are the query vector, key vector, and value vector, respectively. k It is the dimension of the key vector, used as a scaling factor to stabilize the calculation of the attention score.
[0072] The external knowledge vector used by the encoder inversion module to assist training is the preprocessed simulated seismic source data g of the preceding time step, and the internal knowledge vector is the target seismic source vector u of the decoder forward modeling module. It is used alone for the attention mechanism only when geological data of the working face or simulated seismic source data of the preceding time step is missing.
[0073] Specifically, the decoder forward modeling module is responsible for simulating the process of forward modeling the microseismic waveform data y from the intermediate vector z of the simulated source location and energy. This module uses an LSTM-MLP model, where the LSTM model is responsible for predicting the target source vector u, and the MLP model is responsible for further mapping u to the simulated microseismic waveform data y.
[0074] Among them, the MLP fully connected computation formula replaces the convolution operation with the fully connected matrix multiplication on the basis of the CNN convolution computation formula. Compared with the CNN that only performs local connections within the window, global interaction can model global complex relationships. MLP performs high-order feature combination on the LSTM output u through multiple fully connected layers, and then uses activation functions such as ReLU to enhance nonlinear expression ability, and finally outputs the regression result of y.
[0075] Step 2: Training the generative autoregressive neural network model, which includes the encoder inversion module and the decoder forward module.
[0076] A generative autoregressive neural network consisting of an encoder inversion module and a decoder forward module is used as the backbone model.
[0077] The backbone model takes the microseismic waveform to be trained as input, the target source vector u output by the LSTM model of the decoder forward modeling module and the simulated source data x of the subsequent time step as input to the first loss function, and calculates the first loss function value Loss1; the simulated microseismic waveform data y output by the MLP model of the decoder forward modeling module and the target microseismic waveform data sou are taken as input to the second loss function, and the second loss function value Loss2 is calculated.
[0078] The first loss function value Loss1 and the second loss function value Loss2 together form the joint loss function Loss. total The specific process formula is as follows:
[0079] Loss total = (1-λ)Loss1+λLoss2;
[0080] The default adjustment parameter λ = 0.5.
[0081] Specifically, Loss1 uses the mean squared error (MSE) to correct outliers in the simulated source data x that deviate too much from subsequent time steps in the target source vector u. Loss2 uses the mean absolute error (MAE) because the simulated microseismic waveform data y can only approximate the shape of the target microseismic waveform data Sout, and needs to be robust to outliers. The expressions for each loss function are as follows:
[0082]
[0083] Among them, y t For the t-th target value, Let t be the predicted value, and N be the number of samples.
[0084] In some embodiments, such as Figure 4 As shown, a schematic diagram of a hardware and software system for predicting the time series of microseismic events in coal mines by simulating the forward modeling of waveforms and the inversion of seismic sources is presented.
[0085] The hardware includes a processor, a memory, and input and output interfaces. The input interface is mainly connected to the real-time signal output interface of the detector and the microseismic monitoring system. The memory stores computer programs, database data, and documents. The processor is responsible for running the computer programs stored in the memory, thereby realizing the coal mine microseismic time series prediction function of the above-mentioned simulated waveform forward modeling and source inversion process.
[0086] Step 3: Deployment of the trained model and construction of the software system, which includes data preprocessing module, encoder inversion module, decoder forward modeling module, and implementation of microseismic monitoring, prediction, data analysis and visualization functions of the software system.
[0087] The software system mainly includes computer programs and a database system. The data preprocessing module, encoder inversion module, and decoder forward modeling module run as computer programs in the processor. The processor also calls various coal mine data and microseismic data stored in the memory.
[0088] Example 2
[0089] This embodiment, based on Embodiment 1, provides a coal mine microseismic time series prediction system that simulates waveform forward modeling and source inversion processes, such as... Figure 5 As shown, it includes a data preprocessing module, an encoder inversion module, and a decoder forward modeling module;
[0090] The data preprocessing module is responsible for cleaning the collected raw microseismic waveforms, geological data of the working face, and auxiliary training data; dividing the training set and test set; standardizing; generating samples using a sliding window; and obtaining the microseismic waveform data to be trained and the target microseismic waveform data, so as to facilitate input into the neural network model.
[0091] The auxiliary training data includes simulated seismic source data for preceding and subsequent time steps. Based on the geological data of the working face, the simulated seismic source data for the preceding time step is the simulated seismic source location and energy obtained from the microseismic waveform data to be trained using traditional inversion methods; the simulated seismic source data for the subsequent time step is the simulated seismic source location and energy obtained from the target microseismic waveform data using traditional inversion methods.
[0092] The traditional inversion method includes using the P-wave first arrival method to determine the time difference of the P-wave arriving at each detector, combining specific wave velocity field conditions to locate the three-dimensional coordinates of the source, and calculating the source energy through the duration of the seismic phase.
[0093] The encoder inversion module is responsible for simulating the process of inverting the simulated earthquake source location and energy from the microseismic waveform signal to be trained. This module adopts the CNN-BiLSTM-Attention model. CNN is responsible for extracting the deep features of the microseismic waveform to be trained, BiLSTM is responsible for learning long-term time series features along the bidirectional path, and the Attention module is responsible for calculating the similarity between the internal and external knowledge vectors and the earthquake source vector to be trained. After normalization, the attention weights are obtained. The intermediate vector finally output by the module represents the simulated earthquake source location and energy.
[0094] The external knowledge vector is the preprocessed simulated seismic source data of the preceding time step, and the internal knowledge vector is the target seismic source vector of the decoder forward modeling module. When the geological data of the working face or the simulated seismic source data of the preceding time step is missing, only the internal knowledge vector is used.
[0095] The decoder forward modeling module is responsible for simulating the process of forward modeling the simulated microseismic waveform signal from the simulated source location and energy. This module adopts the LSTM-MLP model. The LSTM model is responsible for predicting the target source vector based on the intermediate vector output by the encoder inversion module, and the MLP model is responsible for mapping the target source vector to simulated microseismic waveform data.
[0096] The generative autoregressive neural network composed of the encoder inversion module and the decoder forward model is used as the backbone model; the source vector to be trained corresponds to the simulated source data of the subsequent time step, and the target source vector corresponds to the simulated source data of the preceding time step.
[0097] The backbone model takes the microseismic waveform to be trained as input, the target source vector output by the LSTM model of the decoder forward modeling module and the simulated source data of the subsequent time step as input to the first loss function, and calculates the value of the first loss function; the simulated microseismic waveform output by the MLP model of the decoder forward modeling module and the target microseismic waveform data are taken as input to the second loss function, and the value of the second loss function is calculated.
[0098] The first loss function value and the second loss function value together form the joint loss function.
[0099] The parts of the technical solutions provided in the embodiments of this application that are consistent with the implementation principles of corresponding technical solutions in the prior art have not been described in detail to avoid excessive elaboration.
[0100] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the invention. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A method for predicting the time series of microseismic events in coal mines by simulating waveform forward modeling and source inversion processes, characterized in that, include: The collected raw microseismic waveforms, geological data of the working face, and auxiliary training data are processed to generate microseismic waveform data to be trained and target microseismic waveform data. The process of inverting simulated source parameters from microseismic waveform signals to be trained involves receiving microseismic waveform data to be trained as input and outputting an intermediate vector representing the simulated source location and energy. The process of forward modeling the simulated microseismic waveform signal from the simulated source parameters involves receiving the intermediate vector, predicting the target source vector, and finally outputting the simulated microseismic waveform data.
2. The method for predicting the time series of microseismic events in coal mines by simulating waveform forward modeling and source inversion processes according to claim 1, characterized in that, The auxiliary training data includes simulated seismic source data for preceding and subsequent time steps; The simulated seismic source data of the preceding time step is obtained from the microseismic waveform data to be trained using traditional inversion methods based on the geological data of the working face, including the simulated seismic source location and energy. The simulated source data for the subsequent time step is obtained from the target microseismic waveform data using traditional inversion methods, which simulates the source location and energy.
3. The method for predicting the time series of microseismic events in coal mines by simulating waveform forward modeling and source inversion processes according to claim 2, characterized in that, The traditional inversion method includes using the P-wave first arrival method to determine the time difference of the P-wave arriving at each detector, combining the wave velocity field conditions to locate the three-dimensional coordinates of the source, and calculating the source energy through the phase duration.
4. The method for predicting the time series of microseismic events in coal mines by simulating waveform forward modeling and source inversion processes according to claim 1, characterized in that, The process of inverting and simulating source parameters adopts the CNN-BiLSTM-Attention model, which includes a convolutional neural network (CNN) to extract deep features of the microseismic waveform to be trained, a bidirectional long short-term memory network (BiLSTM) to learn the bidirectional time dependency of the deep feature sequence, and an attention mechanism responsible for calculating the similarity between the knowledge vector and the source vector to be trained, and outputting an intermediate vector.
5. The method for predicting the time series of microseismic events in coal mines by simulating waveform forward modeling and source inversion processes according to claim 4, characterized in that, The attention mechanism described above only uses vectors, including external knowledge vectors and internal knowledge vectors. The external knowledge vector is the simulated seismic source data of the preceding time step; The internal knowledge vector is the target seismic source vector output by the decoder forward modeling module.
6. The method for predicting the time series of microseismic events in coal mines by simulating waveform forward modeling and source inversion processes according to claim 1, characterized in that, The forward modeling process for simulating microseismic waveform signals employs an LSTM-MLP model, which includes a Long Short-Term Memory (LSTM) network responsible for predicting the target source vector based on the intermediate vector output by the encoder inversion module, and a Multilayer Perceptron (MLP) responsible for mapping the target source vector into simulated microseismic waveform data.
7. A method for predicting the time series of microseismic events in coal mines by simulating waveform forward modeling and source inversion processes according to any one of claims 1-6, characterized in that, The process of inverting and simulating source parameters and the process of forward simulating microseismic waveform signals constitute a generative autoregressive neural network as the backbone model, and the backbone model is trained using a joint loss function. The joint loss function includes a first loss function and a second loss function; The first loss function calculates the error between the target source vector predicted by the Long Short-Term Memory (LSTM) network and the simulated source data of the subsequent time steps during the forward modeling of the microseismic waveform signal. The second loss function calculates the error between the simulated microseismic waveform output by the multilayer perceptron (MLP) and the target microseismic waveform data during the forward modeling of the microseismic waveform signal.
8. A coal mine microseismic time series prediction system simulating waveform forward modeling and source inversion processes, used to implement the coal mine microseismic time series prediction method simulating waveform forward modeling and source inversion processes as described in any one of claims 1-9, characterized in that, The system includes hardware devices and software systems running on them; The hardware device includes a processor, memory, and input and output interfaces; The software system is a software system for implementing the coal mine microseismic time series prediction method for simulating waveform forward modeling and source inversion process as described in any one of claims 1-9, including a data preprocessing module, an encoder inversion module, and a decoder forward modeling module; The data preprocessing module is used to process the collected raw microseismic waveforms, working face geological data and auxiliary training data to generate microseismic waveform data to be trained and target microseismic waveform data. The decoder inversion module is used to invert the simulated source parameters from the microseismic waveform signal to be trained. It receives the microseismic waveform data to be trained as input and outputs an intermediate vector representing the simulated source location and energy. The decoder forward modeling module is used to perform forward modeling of the simulated microseismic waveform signal from the simulated source parameters, receive the intermediate vector, predict the target source vector, and finally output the simulated microseismic waveform data.
9. A coal mine microseismic time series prediction system for simulating waveform forward modeling and source inversion processes according to claim 8, characterized in that, The input interface is connected to a detector or microseismic monitoring system to receive real-time signals. The processor implements the prediction method by running a computer program. The memory stores the computer program, coal mine geological data, and microseismic data.