Gas emission quantity prediction method based on time sequence self-attention network

Abstract

The invention discloses a gas emission quantity prediction method based on a time sequence self-attention network. The method comprises the following steps: screening main control factor indexes according to mine monitoring data; a time-frequency feature vector is extracted and reconstructed through Wigner-Ville distribution and a variational auto-encoder; calculating an attention weight by using a time sequence self-attention mechanism to strengthen time sequence feature representation; and performing multi-scale feature fusion and mapping through a parallel multilayer sensor network, and outputting a predicted value of the gas emission quantity. The method can effectively improve the prediction precision and generalization capability of the gas emission quantity.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of artificial intelligence and mining engineering, and in particular relates to a method for predicting gas outburst based on a temporal self-attention network. Background Technology

[0002] Excessive gas emissions are one of the main safety threats faced during the mining of high-gas and coal and gas outburst seams. Accurate prediction of gas emission is crucial for mine gas disaster monitoring, early warning, and prevention. As coal mining extends to deeper levels, gas geological conditions become increasingly complex. The gas migration and abnormal emission processes exhibit strong nonlinearity, strong coupling, and time-varying characteristics with mining disturbances, further increasing the difficulty of prediction.

[0003] In existing technologies, gas emission prediction methods are mainly divided into two categories: one is traditional methods based on mathematical statistics and empirical formulas (such as regression analysis and time series analysis). These methods are difficult to effectively capture the complex nonlinear relationship between gas emission and various influencing factors, and they have high requirements for data stability and insufficient generalization ability in the variable working environment of underground mines. The other category is methods based on shallow machine learning models (such as support vector machines and random forests). Although these methods improve nonlinear processing capabilities to some extent, their feature extraction depends on manual design, and the models have limited ability to capture long-term dependencies and dynamic time series patterns. In recent years, although some studies have attempted to introduce recurrent neural networks (RNNs) and their variants (such as LSTMs) for time series prediction, these models suffer from gradient vanishing / exploding problems, are insensitive to key abrupt changes in the sequence, and are difficult to compute in parallel. As a result, their prediction accuracy and real-time performance are still not ideal when processing long-sequence, high-frequency gas monitoring data.

[0004] Therefore, existing gas emission prediction technologies generally suffer from problems such as insufficient feature correlation identification, limited ability to extract time-frequency features, weak ability to model dynamic dependencies in long sequences, and poor model generalization performance, making it difficult to meet the actual needs of high-gas mines for high-precision, real-time prediction of emission volume. Summary of the Invention

[0005] This invention proposes a gas outflow prediction method based on a temporal self-attention network to solve the problems existing in the prior art.

[0006] To achieve the above objectives, this invention provides a gas outflow prediction method based on a temporal self-attention network, comprising the following steps: Based on mine monitoring data, the main controlling factors related to gas emission were selected; Based on the aforementioned key control factor indicators, a time-frequency feature vector is generated through time-frequency distribution transformation and feature reconstruction; Based on the time-frequency feature vector, the key features are enhanced by calculating the attention weights between features at different times in the time series, thereby obtaining the enhanced time-series feature representation. Based on the enhanced temporal feature representation, feature fusion and mapping are performed through a parallel multi-scale feature extraction network to output the predicted gas outburst.

[0007] Optionally, the step of generating the time-frequency feature vector includes: The time series composed of the main controlling factor indicators is subjected to Wigner-Ville distribution transformation to generate a time-frequency feature map; The time-frequency feature map is denoised and reconstructed using a variational autoencoder to generate the time-frequency feature vector.

[0008] Optionally, the steps for obtaining enhanced key features include: Based on the time-frequency feature vector, generate a query vector, a key vector, and a value vector; The attention weights are obtained by calculating the dot product of the transpose of the query vector and the key vector and applying a normalized exponential function. The enhanced temporal feature representation is obtained by weighting and summing the value vector according to the attention weights.

[0009] Optionally, the Wigner-Ville distribution transform is achieved by calculating the Fourier transform of the instantaneous autocorrelation function of the signal.

[0010] Optionally, the process of obtaining attention weights can be computed in parallel using a multi-head attention mechanism.

[0011] Optionally, the feature fusion and mapping via a parallel multi-scale feature extraction network includes: The enhanced temporal feature representation is subjected to dilated convolution operations through several parallel branches with different dilation rates to extract multi-scale features. The multi-scale features output from each of the branches are concatenated; Perform a fully connected mapping on the spliced ​​features to output the predicted gas outflow value.

[0012] Optionally, after outputting the predicted gas emission values, an online model update step is also included: The model parameters are adjusted based on the error between the predicted value and the actual monitored value.

[0013] Optionally, the steps for screening key control factor indicators include: Calculate the Pearson correlation coefficients between each influencing factor and the gas emission rate; Based on the comparison results between the correlation coefficient and the preset threshold, the main control factor indicators are selected.

[0014] Compared with the prior art, the present invention has the following advantages and technical effects: This application provides a gas emission prediction method based on a temporal self-attention network. Through a master factor screening module, Pearson correlation analysis is used to accurately identify key influencing factors significantly related to gas emission, effectively reducing data dimensionality and removing redundant information. Through a time-frequency feature extraction and reconstruction module, combining Wigner-Ville distribution and variational autoencoder, high-resolution time-frequency features are obtained and their robustness is enhanced. Through a temporal self-attention feature enhancement module, long-term dependencies and key mutation patterns in the gas emission sequence are dynamically captured. Through a parallel multilayer perceptron feature enhancement module, multi-scale features are integrated, comprehensively improving the model's nonlinear fitting capability. Attached Figure Description

[0015] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a schematic diagram of the method flow according to an embodiment of the present invention. Detailed Implementation

[0016] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0017] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0018] Example 1 like Figure 1 As shown, this embodiment provides a gas outflow prediction method based on a temporal self-attention network, including the following steps: Based on mine monitoring data, the main controlling factors related to gas emission were selected; Based on the main control factor indicators, a time-frequency feature vector is generated through time-frequency distribution transformation and feature reconstruction; Based on the time-frequency feature vector, attention weights between features at different times in the time series are calculated to enhance key features, thereby obtaining an enhanced time-series feature representation. Based on the enhanced temporal feature representation, feature fusion and mapping are performed through a parallel multi-scale feature extraction network to output the predicted gas outburst.

[0019] Specifically, the following steps are included: Step 1: Data Collection and Feature Filtering; Furthermore, the steps for screening the key control factor indicators include: Calculate the Pearson correlation coefficients between each influencing factor and the gas emission rate; Based on the comparison results between the correlation coefficient and the preset threshold, the main control factor indicators are selected.

[0020] Specifically, several time-series parameters affecting gas emission are collected from the mine safety monitoring system, including but not limited to: coal seam depth, air volume, return air gas concentration, extraction volume, extraction rate, and daily output.

[0021] The Pearson correlation analysis algorithm was used to calculate the correlation coefficients between various influencing factors and gas emission, identifying the main controlling factors highly correlated with the target variable (gas emission). This was used to construct the model input feature set. Redundant features (indicators with correlation coefficients below 0.3) were removed through quantitative screening, improving the model's input effectiveness and convergence speed. The Pearson correlation coefficient calculation formula is as follows: In the formula, X i Y represents the specific value of a single influencing factor, such as the extraction volume at time i or the daily output at time i. i This represents the measured value of gas emission at the corresponding time, where n is the total number of samples in the "candidate index data - measured value of gas emission".

[0022] Step 2: Time-frequency feature extraction and data reconstruction; Furthermore, the steps for generating time-frequency feature vectors include: The time series composed of the main controlling factor indicators is subjected to Wigner-Ville distribution transformation to generate a time-frequency feature map; The time-frequency feature map is denoised and reconstructed using a variational autoencoder to generate the time-frequency feature vector.

[0023] The Wigner-Ville distribution transform is achieved by calculating the Fourier transform of the instantaneous autocorrelation function of the signal.

[0024] Specifically, gas emission volume exhibits a complex variation pattern over time, showing both short-term high-frequency fluctuations and long-term low-frequency trends. To accurately extract the frequency characteristics of gas emission volume changes over time, a Wigner-Ville distribution (WVD) transformation is performed on the time-series signals of the selected main control factors to obtain joint time-frequency features. These features can simultaneously reflect the impact of changes in influencing factors at different time points and frequencies on gas emission volume. To reduce interference from cross-terms generated during the WVD transformation (avoiding spurious features affecting the accuracy of gas emission volume prediction) and enhance the representation capability of time-frequency features, a variational autoencoder (VAE) structure is introduced to reconstruct the time-frequency features. The optimized features can more accurately map the correlation between influencing factors and gas emission volume.

[0025] Input signals of influencing factors (x[n] represents the time series signal of a certain controlling influencing factor after screening in step one), its autocorrelation and Wigner-Ville distribution The definition is as follows: In the formula: It is the time series signal of the main influencing factors input; n is the time index; k reflects the degree of correlation of the influencing factor data at different times; yes The complex conjugate; ω is the kernel of the Fourier transform, used to convert time-domain signals into frequency-domain signals and extract frequency features; j is the imaginary unit (j2=-1); m is the time lag, capturing the changing patterns of influencing factors under different time intervals; ω is the frequency, reflecting the speed of change of influencing factors.

[0026] The variational autoencoder consists of a convolutional neural network, ensuring that the reconstructed features retain key information affecting gas emission rates. The latent vector z and the objective function L are represented as follows: In the formula: μ represents the mean, reflecting the overall trend of the reconstructed features; σ represents the standard deviation, reflecting the dispersion of the reconstructed features; ε represents the random noise sampled from the standard normal distribution; It represents the time-frequency characteristics of influencing factors at a certain moment; The reconstructed time-frequency characteristics of the main influencing factors; n is the total number of inputs; R_Loss represents the reconstruction loss, measuring the difference between the decoder output and the original factors affecting gas emission variation; K_Loss represents the KL divergence loss, avoiding prediction deviations in gas emission due to abnormal feature distribution; μ i and σ i Let represent the mean and standard deviation of the i-th latent variable, respectively.

[0027] Step 3: Constructing the temporal self-attention module; Furthermore, the steps for obtaining enhanced key features include: Based on the time-frequency feature vector, generate a query vector, a key vector, and a value vector; The attention weights are obtained by calculating the dot product of the transpose of the query vector and the key vector and applying a normalized exponential function. The enhanced temporal feature representation is obtained by weighting and summing the value vector according to the attention weights.

[0028] The process of obtaining attention weights employs a multi-head attention mechanism for parallel computation.

[0029] Specifically, the changes in gas emission volume have significant temporal dependence. To capture this temporal dependence, the main influencing factor data after screening in step one and reconstruction in step two are input into the Temporal Self-Attention (TSA) module to learn the global dependence between features in the time dimension and enhance the feature capture of key periods affecting gas emission volume.

[0030] The data after correlation analysis is input into the Temporal Self-Attention (TSA) module to learn the global dependencies between features in the temporal dimension.

[0031] First, we define the input features X (X is a multi-dimensional time series matrix, where each row corresponds to a monitoring time and each column corresponds to a reconstructed feature of a major influencing factor) of gas emission influencing factors. We then perform a convolution operation on the input time series data to obtain the query (Q), key (K), and value (V) as shown below: In the formula, Q is used to locate the historical influencing factors that need to be focused on at the current moment; K is used to match Q and calculate the correlation degree of influencing factors at different times; V is used to output the corresponding feature information based on the correlation degree.

[0032] By combining a batch normalization layer and a softmax function, Q' and K' are obtained, and the attention weights between Q' and K' are calculated. The attention weights S and V are multiplied to obtain the output activation weight w. The obtained activation weights are then used to enhance the key features of the reconstructed data in step two. The relevant calculations are shown below: In the formula: This represents a dot product operation; the value in row t and column s of S represents the weight of the influencing factor features at time s on the prediction of gas emission at time t, and w is the incentive weight. Based on S, V is weighted to strengthen the features of key time periods and weaken the features of irrelevant time periods, thereby improving the effectiveness of the features in predicting gas emission.

[0033] Furthermore, feature fusion and mapping via parallel multi-scale feature extraction networks include: The enhanced temporal feature representation is subjected to dilated convolution operations through several parallel branches with different dilation rates to extract multi-scale features. The multi-scale features output from each of the branches are concatenated; Perform a fully connected mapping on the spliced ​​features to output the predicted gas outflow value.

[0034] This can be achieved through the following steps: Step 4: Design of the feature enhancement module for the multilayer perceptron; The multiplicative index obtained by multiplying the activation weight w from step three and the output W from step two is input into the multilayer perceptron (MLP) feature enhancement module to extract local and global features at different scales. The module consists of several parallel branches, each containing dilated convolutional layers, batch normalization layers, and nonlinear activation layers. The outputs of several branches are concatenated and fused with a fully connected layer to form the enhanced feature F. out : F out =Concat (f1(X), f2(X), f3(X), f4(X)); Where f i (X) represents the dilated convolution feature extraction function of the i-th branch.

[0035] Step 5: Output layer construction and model training; Fusion feature F out Through the Flatten layer and the fully connected layer, the one-dimensional feature vector is mapped to the gas emission prediction value y, realizing the final mapping of "influencing factor features → gas emission".

[0036] The Adam optimizer is used for training, and the loss function is mean squared error (MSE).

[0037] in, Let be the measured gas emission at time i. Let be the predicted gas outflow at time i.

[0038] Furthermore, after outputting the predicted gas emission value, the model is updated online. The model parameters are adjusted based on the error between the predicted value and the actual monitored value.

[0039] This embodiment also proposes a gas outflow prediction system based on a temporal self-attention network, comprising four main modules: ① a main control factor screening module; ② a time-frequency feature extraction and reconstruction module; ③ a temporal self-attention (TSA) feature enhancement module; and ④ a parallel multilayer perceptron (MLP) feature enhancement and output module. Specific technical measures, structural features, and functions are as follows: Main control factor screening module: Based on the original time series data of gas emission influencing factors obtained by the downhole online sensor group, establish the main control factor index; Time-frequency feature extraction and reconstruction module: Based on the main control factor index, the time-frequency feature map is obtained through Wigner-Ville distribution transformation, and a variational autoencoder is introduced to denoise and reconstruct the time-frequency feature map to obtain a high-resolution time-frequency feature vector; Temporal Self-Attention (TSA) Feature Enhancement Module: Based on the time-frequency feature vector, key dynamic dependency features are extracted through the temporal self-attention mechanism to obtain the enhanced temporal feature representation; Parallel Multilayer Perceptron (MLP) Feature Enhancement and Output Module: Based on the enhanced temporal feature representation, nonlinear mapping is performed through a parallel hollow multilayer perceptron network to output the gas emission prediction value, and to continuously correct and optimize the prediction results.

[0040] The methods for establishing the main control factor indicators include: Based on the raw monitoring data collected by downhole sensors, the correlation coefficients between each influencing factor and the gas emission rate were calculated using the Pearson correlation analysis method. Based on the set correlation threshold (|r| ≥ 0.5), the main controlling factors that are highly correlated with gas emission volume are screened out, and a simplified input feature set is constructed.

[0041] Furthermore, the method for establishing the time-frequency feature vector includes: Based on the time series composed of the main control factor indicators, a joint time-frequency domain analysis was performed using the Wigner-Ville distribution to generate a high-resolution time-frequency feature map. Based on the variational autoencoder structure, the time-frequency feature map is encoded and decoded to remove cross-term interference and compress it into a low-dimensional latent feature vector, which serves as the input to the subsequent model.

[0042] Methods for extracting enhanced temporal feature representations include: Based on the time-frequency feature vector, a time-series self-attention module is constructed. By generating Query, Key, and Value vectors and calculating attention weights, the long-term and short-term dependencies in the time series are dynamically captured. Based on the multi-head attention mechanism, key features at different time scales are extracted in parallel to improve the model's ability to perceive sudden gas outburst events.

[0043] The method for outputting the predicted gas emission value includes: Based on the enhanced temporal feature representation, it is input into a parallel hollow multilayer perceptron network, and multi-scale nonlinear features are extracted through fully connected branches with different expansion rates. The feature fusion layer concatenates the outputs of each branch and maps them to the final output layer to obtain the predicted gas emission rate for the next moment. The model parameters are then corrected by comparing them with the actual monitoring values, thus enabling online updating and optimization of the model.

[0044] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A gas emission amount prediction method based on a timing self-attention network, characterized by, The method comprises the following steps: According to the mine monitoring data, the main control factor index related to the gas emission amount is screened out; According to the main control factor index, the time-frequency feature vector is generated through time-frequency distribution transformation and feature reconstruction; According to the time-frequency feature vector, the key features are strengthened by calculating the attention weights between features at different times in the time sequence, and the strengthened time sequence feature representation is obtained; According to the strengthened time sequence feature representation, the feature fusion and mapping are performed through the parallel multi-scale feature extraction network, and the gas emission amount prediction value is output.

2. The method of claim 1, wherein, The step of generating the time-frequency feature vector comprises: The Wigner-Ville distribution transformation is performed on the time sequence composed of the main control factor index to generate a time-frequency feature map; The variational autoencoder is used to denoise and reconstruct the time-frequency feature map to generate the time-frequency feature vector.

3. The method of claim 1, wherein, The step of obtaining the strengthened key features comprises: According to the time-frequency feature vector, a query vector, a key vector and a value vector are generated; The dot product of the transpose of the query vector and the key vector is calculated, and a normalization exponential function is applied to obtain the attention weight; According to the attention weight, the value vector is weighted and summed to obtain the strengthened time sequence feature representation.

4. The method of claim 2, wherein, The Wigner-Ville distribution transformation is realized by calculating the Fourier transform of the signal instantaneous autocorrelation function.

5. The method of claim 3, wherein, The process of obtaining the attention weight adopts a multi-head attention mechanism for parallel calculation.

6. The method of claim 1, wherein, The feature fusion and mapping through the parallel multi-scale feature extraction network comprises: The strengthened time sequence feature representation is subjected to a hole convolution operation through a plurality of parallel branches with different expansion rates to extract multi-scale features; The multi-scale features output by each branch are spliced; The spliced features are fully connected and mapped to output the gas emission amount prediction value.

7. The method of claim 1, wherein, After outputting the gas emission amount prediction value, the model online updating step is further included: According to the error between the prediction value and the actual monitoring value, the model parameters are corrected.

8. The method of claim 1, wherein, The step of screening out the main control factor index comprises: The Pearson correlation coefficient between each influencing factor and the gas emission amount is calculated; According to the comparison result of the correlation coefficient and the preset threshold, the main control factor index is screened out.

Images