Gas-containing coal rock damage risk prediction method based on acoustic emission and deep residual attention network

Abstract

The invention discloses a gas-containing coal rock damage risk prediction method based on acoustic emission and a depth residual attention network, and belongs to the field of mine safety and geological disaster prediction.The method comprises the steps that firstly, original acoustic emission signals of a gas-containing coal rock combination in the whole uniaxial compression damage process are obtained; then, a ReliefF feature selection algorithm is adopted to objectively screen out main control factors from the multiple initial acoustic emission feature indexes; constructing an input feature vector and carrying out data preprocessing and sample division; then, a deep residual attention noise reduction network model is constructed and trained, the model sequentially extracts time sequence features and inhibits noise through a convolution sparse attention module, fuses time-frequency domain features through a time-frequency feature transmission module to capture non-stationary signal features, and finally outputs the damage risk level of the coal-rock combination through a multi-layer perceptron. According to the method, the objectivity and accuracy of prediction and the robustness in a noise environment are effectively improved, and refined quantitative identification of the damage risk level is realized.

Description

Technical Field

[0001] This invention belongs to the field of mine safety and geological disaster prediction technology, and in particular relates to a method for predicting the risk of damage to gas-bearing coal and rock based on acoustic emission and deep residual attention network. Background Technology

[0002] In the process of deep coal mining, gas-bearing coal-rock assemblages are prone to dynamic disasters such as coal and gas outbursts and rock bursts under complex geomechanical environments, seriously threatening mine safety. Currently, the prediction of such disasters mainly relies on monitoring and analyzing acoustic emission signals generated during the destruction of coal-rock assemblages. Existing technologies commonly employ prediction methods based on single empirical criteria, such as judging based on the intensity or energy of acoustic emission signals. This method can provide preliminary early warning of the destruction trend under specific conditions. In addition, there are evaluation methods based on multi-index weighting, such as the analytic hierarchy process (AHP) or fuzzy comprehensive evaluation, which introduce multiple acoustic emission characteristic parameters and assign corresponding weights to improve the comprehensiveness of the judgment. Other research attempts to use multi-source precursor information fusion technology to integrate acoustic emission with other monitoring data to achieve a more comprehensive characterization of the destruction process. These methods, in practice, help to identify and provide early warning of the destruction state of coal-rock masses to a certain extent.

[0003] However, existing technologies still have significant shortcomings: First, methods based on single empirical criteria cannot fully reflect the complex process of coal and rock failure controlled by multiple factors, ignoring the influence of other key characteristic parameters, which leads to deviations in prediction results in actual complex mining environments and insufficient effectiveness of early warning; Second, although methods based on multi-index weight allocation consider multiple factors, their weight settings often rely on prior experience, which is subjective and affects the objectivity and accuracy of prediction results; In addition, methods based on multi-source precursor information fusion are more theoretically sound, but in practical applications they often face technical challenges such as unstable data quality, high complexity of multi-source information fusion, and difficulty in effectively extracting precursor features of failure, which limit their engineering applicability and reliability. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention proposes a method for predicting the risk of damage to gas-bearing coal and rock based on acoustic emission and deep residual attention networks, thereby resolving the issues present in the existing technologies.

[0005] To achieve the above objectives, this invention provides a method for predicting the failure risk of gas-bearing coal and rock based on acoustic emission and deep residual attention networks, comprising: S1. Obtain the original acoustic emission signals of the gas-bearing coal-rock composite during the entire process of uniaxial compression failure; S2. Extract features from the original acoustic emission signal to obtain multiple initial acoustic emission feature indices, and use the ReliefF feature selection algorithm to select the main controlling factors from the multiple initial acoustic emission feature indices; S3. Construct an input feature vector based on the main controlling factors, and perform data preprocessing and sample partitioning on the input feature vector to obtain a training set and a test set; S4. Construct a deep residual attention denoising network model, and train the model using the training set; the deep residual attention denoising network model includes, in sequence: A convolutional sparse attention module is used to extract the temporal feature information of the main controlling factors during the damage process, and introduces a soft threshold to optimize the attention score to eliminate noise information; The time-frequency feature transfer module is used to learn the spatial structure of features and obtain non-stationary signal features by combining time-varying features and spectral features. A multilayer perceptron is used to establish a nonlinear relationship between mapping feature information and risk type, and output the damage risk level of gas-bearing coal-rock assemblages. S5. Evaluate the performance of the trained model using the test set and apply it to predict the failure risk level of actual gas-bearing coal-rock assemblages.

[0006] Preferably, the ReliefF feature selection algorithm uses Manhattan distance to calculate feature weights, and features with positive weight values ​​are determined as the controlling factors.

[0007] Preferably, the main control factors include amplitude, rise time, ring count, energy, root mean square, average signal level, number of impacts, impact rate, and peak frequency.

[0008] Preferably, the input feature vector is processed using a sliding window method to transform the time-series data into samples that the model can process, and the data within each window is normalized.

[0009] Preferably, the convolutional sparse attention module performs the following operations: Use the input data as query Q; A one-dimensional convolution operation is performed on the input data to generate a key K, which is shared with the value V; Calculate the attention score matrix S; Apply soft thresholding to the attention score matrix S, setting scores below a set threshold to zero; The calculation module outputs O.

[0010] Preferably, the soft thresholding process is implemented using the ReLU function, specifically as follows: S ij =ReLU( S ij- τ ), where τ is the set threshold.

[0011] Preferably, the time-frequency feature transfer module includes a multi-layer stacked structure of one-dimensional convolution, one-dimensional Fourier convolution, and residual connections, and weights the output features through a channel attention mechanism.

[0012] Preferably, the one-dimensional Fourier convolution transforms the input signal to the frequency domain through a fast Fourier transform, multiplies it with a trainable weight matrix, and then transforms it back to the time domain through an inverse Fourier transform.

[0013] Preferably, the output layer of the multilayer perceptron uses the Softmax activation function to output four nodes, which correspond to the four risk stages of coal-rock composite failure: initial compaction stage, linear elastic stage, plastic deformation stage, and post-peak stage.

[0014] Preferably, by constructing an experimental system that includes a mechanical servo loading system, an acoustic emission signal monitoring system, and a gas-solid coupling device, the stress and gas environment of deep coal and rock masses are simulated, and the original acoustic emission signals of the gas-bearing coal and rock assemblies during the entire process of uniaxial compression failure are collected.

[0015] Compared with the prior art, the present invention has the following advantages and technical effects: This invention employs the ReliefF feature selection algorithm to objectively and data-drivenly screen out the main controlling factors that significantly affect coal and rock damage from multiple initial acoustic emission feature indicators. This effectively avoids the subjectivity of manually setting feature weights, ensuring the scientific nature and representativeness of the model input features, thereby significantly improving the objectivity and accuracy of subsequent risk prediction.

[0016] The deep residual attention denoising network model constructed in this invention, especially the convolutional sparse attention module for extracting temporal features and eliminating noise information, and the time-frequency feature transfer module for fusing and learning the time-frequency domain features of non-stationary signals, enables the model to fully and resiliently extract key precursor features in the destruction process, significantly improving the stability and reliability of predictions in the complex acoustic environment of mines.

[0017] The multilayer perceptron at the end of the deep residual attention denoising network model constructed in this invention establishes a complex nonlinear relationship between mapping features and risk types, and finally outputs a specific level of damage risk. This enables refined and quantitative identification and classification of the risk status of coal-rock assemblies throughout their entire life cycle, from initial compaction to post-peak damage, providing a direct and clear decision-making basis for graded early warning and precise prevention and control of mine disasters. Attached Figure Description

[0018] 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 flowchart of a method for predicting the risk of damage to gas-bearing coal and rock based on acoustic emission and deep residual attention network, according to an embodiment of the present invention. Detailed Implementation

[0019] 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.

[0020] 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.

[0021] Example 1 like Figure 1 As shown, this embodiment provides a method for predicting the failure risk of gas-bearing coal and rock based on acoustic emission and deep residual attention networks, including: S1. Obtain the original acoustic emission signals of the gas-bearing coal-rock composite during the entire process of uniaxial compression failure; Furthermore, by constructing an experimental system that includes a mechanical servo loading system, an acoustic emission signal monitoring system, and a gas-solid coupling device, the stress and gas environment of deep coal and rock masses are simulated, and the original acoustic emission signals of the gas-bearing coal and rock assemblies during the entire process of uniaxial compression failure are collected.

[0022] Specifically, the experimental system for uniaxial compression of gas-bearing coal-rock assemblages includes: The RMT-150C rock mechanics servo control loading system is used to apply uniaxial compressive loads. In this embodiment, the displacement control mode is adopted and the loading rate is set to 0.005 mm / s.

[0023] The acoustic emission signal monitoring system employs two acoustic emission sensors with a frequency range of 100-900 kHz, symmetrically arranged at both ends of the specimen. A data acquisition threshold of 60 dB is set to filter out background noise and spurious signals generated by mechanical vibration.

[0024] A gas-solid coupling device and an exhaust / gas supply system are used to create a gas-containing environment. The gas pressure is set to 2 MPa, and the coal sample is placed in this pressure environment for 24 hours to ensure saturation.

[0025] A total of 18 standard "rock-coal-rock" composite specimens, each with a height of 100 mm and a diameter of 50 mm, were prepared according to three height-to-diameter ratios of 1:1:1, 1:2:1, and 1:3:1. The original acoustic emission signals generated by the specimens during the entire uniaxial compression failure process were collected using the aforementioned system.

[0026] S2. Extract features from the original acoustic emission signal to obtain multiple initial acoustic emission feature indices, and use the ReliefF feature selection algorithm to select the main controlling factors from the multiple initial acoustic emission feature indices; Furthermore, the ReliefF feature selection algorithm uses Manhattan distance to calculate feature weights, and features with positive weight values ​​are identified as the controlling factors.

[0027] Furthermore, the main control factors include amplitude, rise time, ring count, energy, root mean square, average signal level, number of impacts, impact rate, and peak frequency.

[0028] The ReliefF algorithm evaluates the importance of a feature by calculating its weight. A positive weight value indicates that the feature can effectively distinguish sample instances, while a negative weight value indicates that the feature is irrelevant or redundant.

[0029] This embodiment can remove redundant and irrelevant features, reduce data dimensionality, and improve the efficiency of subsequent model processing.

[0030] Specifically, 12 initial characteristic indicators are extracted from the acquired raw acoustic emission signals, including amplitude, rise time, ring count, rise count, energy, RMS (root mean square), ASL (average signal level), number of impacts, impact rate, peak frequency, center frequency, and duration.

[0031] The ReliefF feature selection algorithm was used to analyze the correlation between these feature indices and the stress changes in the specimens, in order to identify the main controlling factors affecting the damage and failure of the coal-rock assemblages. The ReliefF algorithm evaluates the importance of each feature by calculating its weight, and features with weight values ​​close to or less than zero are considered redundant or irrelevant and are discarded.

[0032] The ReliefF algorithm uses Manhattan distance to calculate weights, generating positive and negative weight values. Positive weight values ​​indicate that the feature effectively distinguishes samples, while negative weight values ​​indicate that the feature is irrelevant or redundant. The weight values ​​obtained by the ReliefF algorithm can be derived using the formula: ; In the formula, U [ F ] i Indicates the first i After the nth iteration, the th FThe weights of each feature (such as amplitude, energy, etc.) are initialized to 0; Q This represents a sample randomly drawn from the dataset; Hj Indicates from and Q Searching for similar samples k The th nearest neighbor sample j One (referred to as "guessing the nearest neighbor"); Mj ( C ) indicates from each different class C Searching for in the samples respectively k The th nearest neighbor sample j One (referred to as "mistaken neighbor"); diff ( F , Q , Hj ) indicates a sample Q With sample Hj In feature F F The distance on is usually calculated using the formula | value ( Q , F )- value ( Hj , F )∣; n It is the number of random samples; P ( class ( Q )) is a sample Q The prior probability of belonging to a category.

[0033] In practice, the calculated weight of "duration" was close to zero, so it was removed. Although the weight of "center frequency" was slightly higher than the threshold, it was coupled with "peak frequency" in terms of physical mechanism, and it was not as significant as the latter in reflecting the energy accumulation and sudden release phenomenon of coal and rock destruction. To avoid feature redundancy, it was also removed in the end.

[0034] Finally, the main controlling factors were selected to form the input feature vector of the subsequent model, including: amplitude, rise time, ring count, energy, RMS, ASL, number of impacts, impact rate, peak frequency, etc.

[0035] S3. Construct an input feature vector based on the main controlling factors, and perform data preprocessing and sample partitioning on the input feature vector to obtain a training set and a test set; Furthermore, the input feature vector is processed using a sliding window method to transform the time-series data into samples that the model can process, and the data within each window is normalized.

[0036] The time series data of the main controlling factors obtained in S2 are processed using the sliding window method to construct the training and testing samples required for the model. In this embodiment, the time window size n is set, and the sliding step m is adjusted according to the amount of data.

[0037] The data within each window is normalized and scaled to the [0,1] interval. The calculation formula is as follows: ; in, X The data is to be normalized; X max and X min These are the maximum and minimum values ​​of the feature within the window, respectively.

[0038] S4. Construct a deep residual attention denoising network model, and train the model using the training set; the deep residual attention denoising network model includes, in sequence: A convolutional sparse attention module is used to extract the temporal feature information of the main controlling factors during the damage process, and introduces a soft threshold to optimize the attention score to eliminate noise information; Furthermore, the convolutional sparse attention module extracts key temporal features of the controlling factors during the damage process and suppresses noise. Specific operations include: (1) Input data X Directly used as a query Q .

[0039] (2) To X Perform a one-dimensional convolution operation (kernel size is 1) to generate keys. K and with value V Sharing (i.e.) V = K ).

[0040] (3) Calculate the attention score matrix S : in d k It is the dimension of the key.

[0041] (4) Attention score matrix S Apply soft thresholding: S ij =ReLU( S ij - τ ),in τ This is the set threshold. This operation sets scores below the threshold to zero, enhancing the sparsity of the representation and thus filtering out noise.

[0042] (5) Calculation module output O: O = S ^× V .

[0043] The convolutional sparse attention module smooths and sparsifies the attention score matrix using the Softmax function, and combines the ReLU function with soft thresholding to suppress irrelevant information and enhance the model's robustness to noise.

[0044] The convolutional sparse attention module in this embodiment generates keys and values ​​through convolution operations, enhancing the ability to extract local features, and improves the noise resistance of the model through soft thresholding.

[0045] The time-frequency feature transfer module is used to learn the spatial structure of features and obtain non-stationary signal features by combining time-varying features and spectral features. Furthermore, the time-frequency feature transfer module fuses the time-domain and frequency-domain features of the signal to capture the complex characteristics of non-stationary signals and improve the robustness of the model. Specific operation: The module contains a three-layer stacked structure of "Conv1D + Fourier Convolution 1D + Residual Connection".

[0046] Fourier convolutional layer: converts the input signal X Transform to the frequency domain using Fast Fourier Transform. X fft , and the trainable weight matrix W After multiplication Y fft = X fft W Then, it is transformed back to the time domain by inverse Fourier transform. Y N .

[0047] Residual connection: The output of Conv1D is added to the output of Fourier convolution 1D to alleviate gradient vanishing.

[0048] Channel attention mechanism: Features are obtained from the output of the stacked modules through global average pooling. G Then, the channel attention weights are calculated through two fully connected layers (using ReLU and Sigmoid activation functions). A Finally, it is multiplied by the original features and weighted. To highlight the characteristics of important passages.

[0049] The time-frequency feature transfer module in this embodiment uses Fourier convolution to convert the signal to the frequency domain for processing, and combines it with the channel attention mechanism to extract key features, so as to achieve complementarity of information in different domains and enhance the robustness of the model.

[0050] A multilayer perceptron is used to establish a nonlinear relationship between mapping feature information and risk type, and output the damage risk level of gas-bearing coal-rock assemblages. Furthermore, the output layer of the multilayer perceptron uses the Softmax activation function to output four nodes, which correspond to the four risk stages of coal-rock composite failure: initial compaction stage, linear elastic stage, plastic deformation stage, and post-peak stage.

[0051] Specifically, this embodiment maps the high-order features extracted by the preceding modules to the final damage risk level. The specific structure consists of multiple fully connected layers, with the final output layer using the Softmax activation function to output four nodes, corresponding to the four risk stages of coal-rock composite failure (initial compaction stage, linear elastic stage, plastic deformation stage, and post-peak stage).

[0052] The model was trained using Python 3.9 and the Keras framework, employing the Adam optimizer and using classification cross-entropy as the loss function.

[0053] S5. Evaluate the performance of the trained model using the test set and apply it to predict the failure risk level of actual gas-bearing coal-rock assemblages.

[0054] Specifically, the predictive performance of the model is quantified by metrics such as accuracy, precision, and recall, and the generalization ability and stability of the model are verified by cross-domain experiments and simulated noise experiments.

[0055] The model prediction results of this embodiment are shown in Table 1.

[0056] Table 1

[0057] Experiments show that this embodiment has excellent generalization ability and stability on experimental data with different height-to-diameter ratio combinations, and can still maintain high prediction accuracy in noisy environments, providing a reliable technical means for the prevention and control of gas dynamic disasters in deep coal mines.

[0058] The beneficial effects of this embodiment: This embodiment provides a method for predicting the failure risk of gas-bearing coal and rock formations based on acoustic emission and a deep residual attention network. The ReliefF algorithm is used to screen out the key acoustic emission controlling factors affecting coal and rock failure, effectively reducing data dimensionality and highlighting core features. The constructed deep residual attention denoising network model, particularly its convolutional sparse attention module and time-frequency feature transfer module, can fully extract the temporal and time-frequency features during the failure process and effectively suppress noise interference, thereby achieving accurate and stable prediction of the failure risk level of gas-bearing coal and rock formations. This embodiment uses the deep residual attention denoising network to output four levels of risk probabilities—compaction, elasticity, plasticity, and post-peak—in end-to-end, enabling real-time quantification of the failure stages of formations with different height ratios.

[0059] 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 method for predicting the failure risk of gas-bearing coal and rock based on acoustic emission and deep residual attention networks, characterized in that, Includes the following steps: S1. Obtain the original acoustic emission signals of the gas-bearing coal-rock composite during the entire process of uniaxial compression failure; S2. Extract features from the original acoustic emission signal to obtain multiple initial acoustic emission feature indices, and use the ReliefF feature selection algorithm to select the main controlling factors from the multiple initial acoustic emission feature indices; S3. Construct an input feature vector based on the main controlling factors, and perform data preprocessing and sample partitioning on the input feature vector to obtain a training set and a test set; S4. Construct a deep residual attention denoising network model and train the model using the training set; The deep residual attention denoising network model includes, in sequence: A convolutional sparse attention module is used to extract the temporal feature information of the main controlling factors during the damage process, and introduces a soft threshold to optimize the attention score to eliminate noise information; The time-frequency feature transfer module is used to learn the spatial structure of features and obtain non-stationary signal features by combining time-varying features and spectral features. A multilayer perceptron is used to establish a nonlinear relationship between mapping feature information and risk type, and output the damage risk level of gas-bearing coal-rock assemblages. S5. Evaluate the performance of the trained model using the test set and apply it to predict the failure risk level of actual gas-bearing coal-rock assemblages.

2. The method according to claim 1, characterized in that, The ReliefF feature selection algorithm uses Manhattan distance to calculate feature weights, and features with positive weights are identified as the controlling factors.

3. The method according to claim 2, characterized in that, The main control factors include amplitude, rise time, ring count, energy, root mean square, average signal level, number of impacts, impact rate, and peak frequency.

4. The method according to claim 1, characterized in that, The input feature vector is processed using a sliding window method to transform the time series data into samples that the model can process, and the data within each window is normalized.

5. The method according to claim 1, characterized in that, The convolutional sparse attention module performs the following operations: Use the input data as query Q; A one-dimensional convolution operation is performed on the input data to generate a key K, which is shared with the value V; Calculate the attention score matrix S; Apply soft thresholding to the attention score matrix S, setting scores below a set threshold to zero; The calculation module outputs O.

6. The method according to claim 5, characterized in that, The soft thresholding process is implemented using the ReLU function, specifically as follows: S ij =ReLU( S ij - τ ), where τ is the set threshold.

7. The method according to claim 1, characterized in that, The time-frequency feature transfer module includes a multi-layer stacked structure of one-dimensional convolution, one-dimensional Fourier convolution, and residual connections, and weights the output features through a channel attention mechanism.

8. The method according to claim 7, characterized in that, The one-dimensional Fourier convolution transforms the input signal to the frequency domain through a fast Fourier transform, multiplies it with a trainable weight matrix, and then transforms it back to the time domain through an inverse Fourier transform.

9. The method according to claim 1, characterized in that, The output layer of the multilayer perceptron uses the Softmax activation function to output four nodes, which correspond to the four risk stages of coal-rock composite failure: initial compaction stage, linear elastic stage, plastic deformation stage, and post-peak stage.

10. The method according to claim 1, characterized in that, By constructing an experimental system that includes a mechanical servo loading system, an acoustic emission signal monitoring system, and a gas-solid coupling device, the stress and gas environment of deep coal and rock masses are simulated, and the original acoustic emission signals of the gas-bearing coal and rock assemblies are collected during the entire process of uniaxial compression failure.

Images