Coal and gas outburst acoustic emission signal deep learning early warning method

Abstract

The present application relates to the technical field of deep learning and mine safety monitoring, and discloses a coal and gas outburst acoustic emission signal deep learning early warning method.The method comprises the following steps: collecting and preprocessing multi-channel acoustic emission signals; extracting local and global time sequence features through multi-scale one-dimensional convolution; modeling long-term dependence relationship by using a bidirectional gated recurrent unit; introducing a dynamic channel attention mechanism to adaptively weight key features; and outputting the state probability of the coal body through a space-time joint classifier, and triggering an early warning when the critical instability stage probability exceeds a threshold value.The system comprises a signal collection unit, a preprocessing unit, a multi-scale feature extraction unit, a time sequence modeling unit, an attention weighting unit, a classification early warning unit and an edge deployment communication unit.The present application realizes early, accurate and real-time early warning, which is superior to traditional methods.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of deep learning and mine safety monitoring, specifically involving a deep learning-based early warning method for acoustic emission signals of coal and gas outbursts. Background Technology

[0002] Coal and gas outbursts, as highly destructive dynamic hazards in coal mining, threaten underground operational safety and production efficiency. These hazards arise from the coupling effect of high ground stress, high gas pressure, and weakened coal structure, leading to rapid and severe instability of the coal and rock mass accompanied by a large outburst of gas. To control such hazards, existing monitoring systems generally employ acoustic emission (AE) technology, which indirectly reflects the internal damage evolution state by capturing elastic wave signals released from micro-fractures in the coal body during loading. However, traditional AE monitoring methods mainly rely on a single threshold for early warning judgment, making it difficult to effectively characterize the nonlinear and non-stationary precursor characteristics of the coal body throughout the entire process from compaction and elastic deformation to yielding and even macroscopic failure. This results in frequent missed or false alarms under complex operating conditions, failing to meet the requirements for high-reliability early warning.

[0003] Deep learning-based intelligent early warning methods have shown promise in the field of geotechnical engineering safety monitoring in recent years. These methods attempt to leverage the powerful learning capabilities of neural networks on time-series signals to automatically extract precursory patterns from raw or characterized adverse event (AE) data. However, existing research often focuses on a single signal dimension, neglecting the essential characteristics of the collaborative evolution of multiple physical quantities during coal seam failure. Most models lack coupling with coal and rock mechanics and the physical mechanisms of gas seepage, resulting in a "black box" nature in their decision-making process, making them difficult to trust and adopt in engineering practice. In high-risk scenarios such as severe disasters, if the model cannot reflect physical processes such as permeability evolution and gas desorption-diffusion-emergence, the interpretability and generalization ability of its early warning results will be limited.

[0004] Existing technologies for constructing AE signal-driven early warning models for coal and gas outbursts generally suffer from the following problems: the input feature dimension is singular, failing to integrate multimodal AE parameters such as amplitude, count, and energy to comprehensively characterize the damage state of the coal body at all stages; the model architecture is not optimized for the temporal nonlinearity and long-term / short-term dependence characteristics of AE signals, making it difficult to capture early and weak precursors; and it completely deviates from the physical evolution mechanism of coal and gas outbursts, failing to introduce constraints such as dynamic transformation probability and effective stress-permeability coupling relationship as in the SIR compartment model, resulting in models that, while capable of fitting, lack physical consistency.

[0005] The aforementioned problems are particularly prominent in complex application scenarios such as deep, high-gas mines, easily leading to delayed early warnings or frequent false alarms, thus hindering the practical implementation of intelligent monitoring technology. Therefore, there is an urgent need for a deep learning-based early warning method that integrates multimodal acoustic emission characteristics, embeds physical mechanism constraints, and possesses early nonlinear precursor recognition capabilities, in order to achieve accurate, reliable, and interpretable graded early warnings of coal and gas outburst risks. Summary of the Invention

[0006] This invention provides a deep learning-based early warning method for acoustic emission signals in coal and gas outbursts, aiming to solve the problems of missed and false alarms caused by relying on fixed thresholds to judge acoustic emission signals in existing technologies. This method constructs a deep neural network model based on the fusion of multi-scale temporal feature extraction and dynamic attention mechanisms to perform high-dimensional nonlinear feature modeling of acoustic emission signals generated by coal during loading, accurately identifying weak anomaly patterns in the precursory stage of coal fracturing, thereby achieving early, accurate, and real-time early warning of coal and gas outburst disasters.

[0007] This invention provides a deep learning-based early warning method for acoustic emission signals of coal and gas outbursts, comprising: The original acoustic emission signal sequence collected by the acoustic emission sensor during the uniaxial triaxial compression experiment of the coal body is obtained. The original acoustic emission signal sequence includes event count, ring count, energy, amplitude, rise time, duration and dominant frequency parameter. The original acoustic emission signal sequence is preprocessed, including timestamp alignment, noise filtering, signal normalization, and missing data imputation, to form standardized multi-channel time-series input data. The standardized multi-channel temporal input data is input into the multi-scale convolutional feature extraction module. The local and global temporal patterns of the signal are extracted synchronously by one-dimensional convolutional kernels with different receptive fields set in parallel, and a multi-scale feature map is generated. The multi-scale feature map is input into a bidirectional gated recurrent unit network to model the long-term dependency of the acoustic emission signal and output a hidden state sequence containing forward and backward temporal context information. The hidden state sequence is input into the dynamic channel attention module, which dynamically assigns weights to each feature channel according to the importance of each feature channel at different time steps, and generates a weighted feature representation. The weighted feature representation is input into a spatiotemporal joint classifier, which consists of a fully connected layer and a Softmax activation function. The classifier outputs the probability distribution of the current state of the coal body, including the stable loading stage, the micro-fracture accumulation stage, the critical instability stage, and the sudden fracture stage. When the probability of the critical instability stage is greater than a preset threshold, a coal and gas outburst early warning signal is triggered.

[0008] Preferably, the original acoustic emission signal sequence is preprocessed, including timestamp alignment, noise filtering, signal normalization, and missing data imputation, to form standardized multi-channel time-series input data, including: The original acoustic emission signal sequence is timestamped to eliminate inter-channel phase shift; The wavelet threshold denoising method is used to perform five-level wavelet decomposition on the signals of each channel, and the denoised signal is reconstructed after applying the improved SURE threshold rule to shrink the detail coefficients. Perform channel-independent Z-score normalization on the denoised signal of each channel; Missing sampling points due to sensor failure or communication interruption are filled using cubic spline interpolation; The processed multi-channel signals are organized in a sliding window manner, with a window length of 5000 sampling points and a step size of 1000 sampling points. Each window corresponds to a label, which is determined by whether a stress drop or a sudden increase in acoustic emission energy occurs within the subsequent 50 sampling points.

[0009] Preferably, the standardized multi-channel temporal input data is input to a multi-scale convolutional feature extraction module, and local and global temporal patterns of the signal are extracted simultaneously using one-dimensional convolutional kernels with different receptive fields set in parallel, generating a multi-scale feature map, including: The standardized multi-channel timing input data is input into three parallel branches respectively; The first branch uses a one-dimensional convolutional layer with a kernel size of 3 to capture high-frequency transient mutation features; The second branch uses a one-dimensional convolutional layer with a kernel size of 7 to extract mid-frequency local breakage patterns; The third branch uses a dilated convolutional layer with a dilation rate of 2 and a kernel size of 5 to detect low-frequency trend changes; The outputs of the three branches are concatenated along the channel dimension to form a fused multi-scale feature map.

[0010] Preferably, the multi-scale feature map is input into a bidirectional gated recurrent unit network to model the long-term dependency of the acoustic emission signal, and outputs a hidden state sequence containing forward and backward temporal context information, including: The multi-scale feature map is input into a bidirectional gated recurrent unit network containing a two-layer stacked structure, with 128 hidden units in each layer. The forward path processing evolves from the initial loading time to the current time. Backward path processing is the reverse evolution from the peak intensity moment back to the current moment; The hidden states of the forward path and the backward path at each time step are added together to generate a bidirectional context fusion representation as the hidden state sequence.

[0011] Preferably, the hidden state sequence is input to a dynamic channel attention module, which dynamically assigns weights based on the importance of each feature channel at different time steps to generate a weighted feature representation, including: The hidden state sequence is subjected to global average pooling in the time dimension to obtain the statistical response vector of each channel; The statistical response vector is input into a two-layer fully connected network. The first layer has 1 / 4 of the total number of channels and the activation function is ReLU. The second layer has the same number of neurons as the total number of channels and the activation function is Sigmoid. The output is a channel weight vector. The channel weight vector is multiplied channel by channel by the hidden state sequence to obtain the weighted feature representation.

[0012] Preferably, the weighted feature representation is input into a spatiotemporal joint classifier, which consists of a fully connected layer and a Softmax activation function, and outputs the probability distribution of the current state of the coal body, including: Max pooling is performed on the weighted feature representation in the time dimension to obtain a feature vector of fixed length; The feature vector is passed sequentially through a three-layer fully connected network with 256, 128 and 4 neurons in each layer, respectively. Apply the Softmax activation function to the output of the last layer to generate probability distributions for the four states.

[0013] Preferably, when the probability of the critical instability stage is greater than a preset threshold, a coal and gas outburst early warning signal is triggered. The preset threshold is 0.85, which is determined by maximizing the early warning accuracy and minimizing the false alarm rate on the validation set. The false alarm penalty coefficient is 3 times the false alarm penalty coefficient.

[0014] Preferably, in the multi-scale convolutional feature extraction module, each branch of the convolutional layer is followed by a batch normalization layer and a ReLU activation function.

[0015] Preferably, the wavelet thresholding denoising method uses the db4 wavelet basis function to perform a five-level decomposition of the original signal.

[0016] Preferably, the method is deployed on a mine edge computing node, which is equipped with a neural network inference acceleration chip, supports INT8 quantization model operation, and has a single inference latency of no more than 20 milliseconds.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention constructs a deep learning-based early warning framework capable of deeply integrating the temporal evolution laws of multidimensional acoustic emission parameters. Through a multi-scale convolution mechanism, it captures different frequency band features throughout the entire process from instantaneous micro-fractures to macroscopic instability; through bidirectional gated recurrent units, it models the long-range dependencies in the coal body damage accumulation process; and through a dynamic channel attention mechanism, it adaptively strengthens the feature channels most discriminative for critical precursors while suppressing environmental noise and irrelevant interference. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the overall technical solution architecture of the present invention; Figure 2 This is a schematic diagram of the core principle framework of the fusion of multi-scale temporal feature extraction and dynamic attention mechanism in this invention; Figure 3 This is a flowchart illustrating the logical flow of the preprocessing of the original acoustic emission signal and the construction of standardized multi-channel timing input in this invention. Figure 4 This is a diagram of the joint feature processing framework of multi-scale convolutional feature extraction, bidirectional temporal modeling, and dynamic channel attention weighting in this invention. Figure 5 This is a logical framework diagram of the state discrimination and early warning decision of the spatiotemporal joint classifier in this invention; Figure 6 This is a schematic diagram of the multi-level interaction relationship and data flow between edge deployment nodes and ground monitoring centers in this invention. Detailed Implementation

[0019] refer to Figures 1 to 6 This invention provides a deep learning-based early warning method for acoustic emission signals in coal and gas outbursts. By constructing an end-to-end trainable deep neural network model, it performs high-dimensional nonlinear time-series modeling of multi-channel acoustic emission signals generated by coal during loading, thereby accurately identifying weak abnormal patterns in the precursory stage of coal fracturing and achieving early, accurate, and real-time early warning of coal and gas outburst disasters. The following description, in conjunction with the steps listed in the invention description, will illustrate this embodiment.

[0020] The method includes the following steps: obtaining the original acoustic emission signal sequence collected by an acoustic emission sensor during the uniaxial triaxial compression experiment of the coal body; The original acoustic emission signal sequence is preprocessed to form standardized multi-channel timing input data; The standardized multi-channel temporal input data is input into the multi-scale convolutional feature extraction module to generate a multi-scale feature map. The multi-scale feature map is input into a bidirectional gated recurrent unit network, and the output is a hidden state sequence containing forward and backward temporal context information; The hidden state sequence is input into the dynamic channel attention module to generate a weighted feature representation; the weighted feature representation is input into the spatiotemporal joint classifier to output the probability distribution of the current state of the coal body. When the probability of the critical instability stage is greater than a preset threshold, a coal and gas outburst early warning signal is triggered.

[0021] In step S1, the raw acoustic emission signal sequence of the coal body is acquired by an acoustic emission sensor during the uniaxial compression experiment. The acoustic emission sensor is a piezoelectric ceramic broadband acoustic emission probe with a frequency response range of 20 kHz to 1500 kHz and a sampling frequency of not less than 5 MHz. The sensor is mounted on the radial and axial surfaces of the coal sample specimen and is tightly attached to the specimen using a magnetic coupling agent to ensure signal transmission efficiency. Each sensor is independently connected to a preamplifier with a gain of 40 dB and a bandwidth covering 20 kHz to 1500 kHz. The amplified analog signal is converted into a digital signal by an analog-to-digital converter at a sampling rate of 5 MHz, and a timestamp is recorded synchronously.

[0022] The raw acoustic emission signal sequence includes event counts, ring counts, energy, amplitude, rise time, duration, and main frequency parameters. Each parameter constitutes an independent data channel, forming a 7-channel raw timing signal. All channel data employs a hardware-level time synchronization mechanism during acquisition to ensure that the time alignment error between channels is no greater than one sampling period. The data acquisition system incorporates a high-precision crystal oscillator as its clock source, guaranteeing time stability during long-term operation. The raw signal is temporarily stored in a local high-speed cache as a binary stream, awaiting subsequent preprocessing calls.

[0023] In step S2, the original acoustic emission signal sequence is preprocessed, including timestamp alignment, noise filtering, signal normalization, and missing data imputation, forming standardized multi-channel timing input data. Timestamp alignment is performed based on a globally unified clock, and all channel data are strictly aligned according to the sampling point index to eliminate phase shifts caused by transmission delays or processing differences. Noise filtering uses a wavelet thresholding method, selecting the db4 wavelet basis function to perform a five-level decomposition of the original signal to obtain approximate coefficients and five-level detail coefficients. An improved SURE thresholding rule is applied to each level of detail coefficients for shrinkage processing. This thresholding rule adaptively calculates the threshold based on the noise variance of each level, avoiding excessive suppression of weak signals by traditional fixed thresholds. The reconstruction process uses inverse wavelet transform to restore the processed coefficients to the denoised signal.

[0024] Signal normalization employs channel-independent Z-score standardization, which involves subtracting the mean of each channel from the training set's data and dividing by its standard deviation, resulting in a mean of 0 and a variance of 1 for each channel's data distribution. This eliminates the impact of dimensional differences on model training. Missing data imputation addresses data loss due to momentary sensor malfunctions or communication interruptions by using cubic spline interpolation based on adjacent valid sampling points, ensuring the continuity of the input sequence.

[0025] The preprocessed data is organized in a sliding window format, with a window length of 5000 sampling points and a step size of 1000 sampling points. Each window corresponds to a label, determined by whether a sudden drop in stress or a sudden increase in acoustic emission energy occurs within the next 50 sampling points: if such an event occurs, it is marked as a critical instability stage; otherwise, it is marked as one of the following stages based on the current loading stage: stable loading stage, micro-fracture accumulation stage, or sudden fracture stage. The label generation logic is executed by an independent state determination module, which makes its judgment based on the joint trend of the axial stress curve and the acoustic emission cumulative energy curve fed back by the pressure machine.

[0026] In step S3, the standardized multi-channel temporal input data is input to the multi-scale convolutional feature extraction module. The module uses parallel one-dimensional convolutional kernels with different receptive fields to simultaneously extract local and global temporal patterns from the signal, generating a multi-scale feature map. The multi-scale convolutional feature extraction module contains three parallel branches. The first branch uses a one-dimensional convolutional layer with a kernel size of 3, a stride of 1, and a "same" padding method to capture high-frequency transient abrupt changes, such as the sharp acoustic emission pulses released instantaneously by microcracks. The second branch uses a one-dimensional convolutional layer with a kernel size of 7, a stride of 1, and a "same" padding method to extract local fracture patterns in the mid-frequency band, such as composite signals formed by the coordinated expansion of multiple microcracks. The third branch uses a dilated convolutional layer with a dilation rate of 2 and a kernel size of 5, with an equivalent receptive field of 9, to expand the receptive field to perceive low-frequency trend changes, such as the slow increase in acoustic emission activity caused by damage accumulation.

[0027] The inputs to all three branches are 7-channel standardized temporal data. Each branch's convolutional layer is followed by a batch normalization layer and a ReLU activation function to enhance non-linear expressiveness and accelerate convergence. The outputs of the three branches are concatenated along the channel dimension to form a fused multi-scale feature map, the number of which is the sum of the output channels of each branch. Assuming each branch outputs 64 channels, the fused feature map has 192 channels. This fused feature map retains the original temporal dimension, providing high-dimensional input for subsequent temporal modeling.

[0028] In step S4, the multi-scale feature map is input into a bidirectional gated recurrent unit network to model the long-term dependency of the acoustic emission signal and output a hidden state sequence containing forward and backward temporal context information. The bidirectional gated recurrent unit network contains a two-layer stacked structure, with 128 hidden units in each layer. The forward path processing follows the temporal evolution from the initial loading time to the current time, capturing the cumulative effect of coal body damage as the loading process progresses; the backward path processing follows the reverse evolution from the peak intensity time back to the current time, using future information to assist in the judgment of the current state.

[0029] The bidirectional structure is implemented through two independent gated recurrent units. The forward unit processes the input sequence in chronological order, while the backward unit processes the same input sequence in reverse chronological order. The hidden states of both units are summed point-by-point at each time step to generate a bidirectional context fusion representation. This bidirectional context fusion representation incorporates both historical dependencies and future trends, overcoming the lag in response to critical precursor signals in unidirectional models. The network output is a sequence of hidden states with the same number of time steps as the input, and the hidden state dimension at each time step is 128.

[0030] In step S5, the hidden state sequence is input to the dynamic channel attention module. This module dynamically assigns weights to each feature channel based on its importance at different time steps, generating a weighted feature representation. The dynamic channel attention module first performs global average pooling on the feature sequence output by the bidirectional gated recurrent unit network in the time dimension to obtain the statistical response vector for each channel, the dimension of which is equal to the hidden state dimension. Then, a nonlinear transformation is performed on this statistical response vector through two fully connected layers: the first layer has 32 neurons (1 / 4 of the total number of channels) and uses ReLU as the activation function; the second layer has 128 neurons (equal to the total number of channels) and uses Sigmoid as the activation function.

[0031] The resulting weight vector reflects the contribution of each channel to the current overall state judgment. Finally, the resulting weight vector is multiplied with the original feature sequence channel by channel to complete the feature recalibration. This process strengthens the feature channels that are most discriminative to critical instability (such as energy surge channels or dominant frequency drift channels), while suppressing noise channels that are more susceptible to environmental interference, thus improving the robustness of the model.

[0032] In step S6, the weighted feature representation is input to a spatiotemporal joint classifier, which consists of a fully connected layer and a Softmax activation function, and outputs the probability distribution of the current state of the coal body. The input of the spatiotemporal joint classifier is the max pooling result of the weighted feature sequence output by the dynamic channel attention module in the time dimension. This operation extracts the strongest response of each channel within the entire time window, forming a fixed-length feature vector.

[0033] The feature vectors are fed into a three-layer fully connected network: the first layer has 256 neurons, the second layer has 128 neurons, and the third layer has 4 neurons, corresponding to the stable loading stage, the micro-fracture accumulation stage, the critical instability stage, and the sudden fracture stage, respectively.

[0034] The final layer uses the Softmax function to output the probability distribution of the four states, ensuring that the sum of the probabilities is 1. The model training uses the cross-entropy loss function, with AdamW as the optimizer, an initial learning rate of 0.001, and a weight decay coefficient of 0.01. A transfer learning strategy is employed during training: first, pre-training is performed on a large number of laboratory single-triaxial compressed acoustic emission datasets to learn general coal seam fracture characteristics; then, fine-tuning is performed on a small amount of downhole measured data. During fine-tuning, the first two convolutional parameters of the multi-scale feature extraction unit are frozen, and only the weights of the higher-level networks are updated to adapt to complex field conditions.

[0035] In step S7, when the probability of the critical instability stage exceeds a preset threshold, a coal and gas outburst early warning signal is triggered. The preset threshold is 0.85, which is determined by a weighted sum of maximizing the early warning accuracy and minimizing the false alarm rate on the validation set. The weighting coefficient is set according to the tolerance for false alarms in the coal mine safety regulations, and the penalty coefficient for false alarms is three times that for false alarms, reflecting the safety principle of "better to have a false alarm than a false alarm".

[0036] Upon generation of the warning signal, a local audible and visual alarm device is immediately triggered via an interrupt mechanism, and the signal is uploaded to the ground monitoring center via industrial Ethernet with the highest priority. The entire method is deployed on a mine edge computing node equipped with a neural network inference acceleration chip, supporting INT8 quantization model operation with a single inference latency of no more than 20 milliseconds, meeting real-time warning requirements. The edge node has a built-in watchdog timer and a dual-power redundancy module to ensure at least 30 minutes of local warning capability in the event of a power outage or communication interruption, and automatically retransmits warning records after connection is restored.

[0037] This method enables a shift from passive response to proactive prediction, providing reliable technical support for safe coal mine production.

[0038] At the system level, this invention provides a deep learning-based early warning system for acoustic emission signals in coal and gas outbursts. This system includes an acoustic emission signal acquisition unit, a data preprocessing unit, a multi-scale feature extraction unit, a temporal context modeling unit, a dynamic attention weighting unit, a state classification and early warning decision unit, and an edge deployment and communication unit. The acoustic emission signal acquisition unit consists of multiple piezoelectric ceramic broadband acoustic emission probes, a preamplifier, an analog-to-digital converter, and a synchronous clock module, responsible for high-fidelity acquisition of multi-channel raw signals. The data preprocessing unit performs timestamp alignment, wavelet denoising, Z-score normalization, and missing value interpolation to generate standardized time-series data. The multi-scale feature extraction unit, temporal context modeling unit, dynamic attention weighting unit, and state classification and early warning decision unit together constitute an end-to-end deep neural network model, deployed in the inference engine of edge computing nodes. The edge deployment and communication unit is responsible for model operation, early warning triggering, and data uploading, and has automatic retransmission functions after power failure and communication recovery. All units are interconnected through an internal bus to form a closed-loop early warning system, meeting the high reliability requirements of complex mining environments.

Claims

1. A deep learning-based early warning method for acoustic emission signals of coal and gas outbursts, characterized in that, include: The original acoustic emission signal sequence collected by the acoustic emission sensor during the uniaxial triaxial compression experiment of the coal body is obtained. The original acoustic emission signal sequence includes event count, ring count, energy, amplitude, rise time, duration and dominant frequency parameter. The original acoustic emission signal sequence is preprocessed, including timestamp alignment, noise filtering, signal normalization, and missing data imputation, to form standardized multi-channel time-series input data. The standardized multi-channel temporal input data is input into the multi-scale convolutional feature extraction module. The local and global temporal patterns of the signal are extracted synchronously by one-dimensional convolutional kernels with different receptive fields set in parallel, and a multi-scale feature map is generated. The multi-scale feature map is input into a bidirectional gated recurrent unit network to model the long-term dependency of the acoustic emission signal and output a hidden state sequence containing forward and backward temporal context information. The hidden state sequence is input into the dynamic channel attention module, which dynamically assigns weights to each feature channel according to the importance of each feature channel at different time steps, and generates a weighted feature representation. The weighted feature representation is input into a spatiotemporal joint classifier, which consists of a fully connected layer and a Softmax activation function. The classifier outputs the probability distribution of the current state of the coal body, including the stable loading stage, the micro-fracture accumulation stage, the critical instability stage, and the sudden fracture stage. When the probability of the critical instability stage is greater than a preset threshold, a coal and gas outburst early warning signal is triggered.

2. The deep learning early warning method for acoustic emission signals of coal and gas outbursts according to claim 1, characterized in that, The original acoustic emission signal sequence is preprocessed, including timestamp alignment, noise filtering, signal normalization, and missing data imputation, to form standardized multi-channel time-series input data, including: The original acoustic emission signal sequence is timestamped to eliminate inter-channel phase shift; The wavelet threshold denoising method is used to perform five-level wavelet decomposition on the signals of each channel, and the denoised signal is reconstructed after applying the improved SURE threshold rule to shrink the detail coefficients. Perform channel-independent Z-score normalization on the denoised signal of each channel; Missing sampling points due to sensor failure or communication interruption are filled using cubic spline interpolation; The processed multi-channel signals are organized in a sliding window manner, with a window length of 5000 sampling points and a step size of 1000 sampling points. Each window corresponds to a label, which is determined by whether a stress drop or a sudden increase in acoustic emission energy occurs within the subsequent 50 sampling points.

3. The deep learning early warning method for acoustic emission signals of coal and gas outbursts according to claim 2, characterized in that, The standardized multi-channel temporal input data is input into a multi-scale convolutional feature extraction module. Local and global temporal patterns are simultaneously extracted from the signal using one-dimensional convolutional kernels with different receptive fields set in parallel, generating a multi-scale feature map, including: The standardized multi-channel timing input data is input into three parallel branches respectively; The first branch uses a one-dimensional convolutional layer with a kernel size of 3 to capture high-frequency transient mutation features; The second branch uses a one-dimensional convolutional layer with a kernel size of 7 to extract mid-frequency local breakage patterns; The third branch uses a dilated convolutional layer with a dilation rate of 2 and a kernel size of 5 to detect low-frequency trend changes; The outputs of the three branches are concatenated along the channel dimension to form a fused multi-scale feature map.

4. The deep learning early warning method for acoustic emission signals of coal and gas outbursts according to claim 3, characterized in that, The multi-scale feature map is input into a bidirectional gated recurrent unit network to model the long-term dependency of the acoustic emission signal, and outputs a hidden state sequence containing forward and backward temporal context information, including: The multi-scale feature map is input into a bidirectional gated recurrent unit network containing a two-layer stacked structure, with 128 hidden units in each layer. The forward path processing evolves from the initial loading time to the current time. Backward path processing is the reverse evolution from the peak intensity moment back to the current moment; The hidden states of the forward path and the backward path at each time step are added together to generate a bidirectional context fusion representation as the hidden state sequence.

5. The deep learning early warning method for acoustic emission signals of coal and gas outbursts according to claim 4, characterized in that, The hidden state sequence is input into a dynamic channel attention module, which dynamically assigns weights to each feature channel based on its importance at different time steps, generating a weighted feature representation, including: The hidden state sequence is subjected to global average pooling in the time dimension to obtain the statistical response vector of each channel; The statistical response vector is input into a two-layer fully connected network. The first layer has 1 / 4 of the total number of channels and the activation function is ReLU. The second layer has the same number of neurons as the total number of channels and the activation function is Sigmoid. The output is a channel weight vector. The channel weight vector is multiplied channel by channel by the hidden state sequence to obtain the weighted feature representation.

6. The deep learning early warning method for acoustic emission signals of coal and gas outbursts according to claim 5, characterized in that, The weighted feature representation is input into a spatiotemporal joint classifier, which consists of a fully connected layer and a Softmax activation function. The classifier outputs the probability distribution of the current state of the coal body, including: Max pooling is performed on the weighted feature representation in the time dimension to obtain a feature vector of fixed length; The feature vector is passed sequentially through a three-layer fully connected network with 256, 128 and 4 neurons in each layer, respectively. Apply the Softmax activation function to the output of the last layer to generate probability distributions for the four states.

7. The deep learning early warning method for acoustic emission signals of coal and gas outbursts according to claim 6, characterized in that, When the probability of the critical instability stage is greater than a preset threshold, a coal and gas outburst early warning signal is triggered. The preset threshold is 0.85, which is determined by maximizing the early warning accuracy and minimizing the false alarm rate on the validation set. The false alarm penalty coefficient is 3 times the false alarm penalty coefficient.

8. The deep learning early warning method for acoustic emission signals of coal and gas outbursts according to claim 7, characterized in that, In the multi-scale convolutional feature extraction module, each branch of the convolutional layer is followed by a batch normalization layer and a ReLU activation function.

9. The deep learning early warning method for acoustic emission signals of coal and gas outbursts according to claim 8, characterized in that, The wavelet threshold denoising method uses the db4 wavelet basis function to perform a five-level decomposition on the original signal.

10. The deep learning early warning method for acoustic emission signals of coal and gas outbursts according to claim 9, characterized in that, The method is deployed on a mine edge computing node equipped with a neural network inference acceleration chip, which supports the operation of INT8 quantized models and has a single inference latency of no more than 20 milliseconds.

Images