Coal rock complex fracture acoustic emission characteristic precursor information identification and short-term early warning method

Abstract

The present application belongs to the technical field of coal mine dynamic disaster prevention and control, and provides a coal rock composite body fracture acoustic emission characteristic precursor information identification and short-term early warning method, which comprises: performing uniaxial compression test on a coal sample, applying an increasing axial load to the coal sample until the coal sample is damaged; collecting test data during the damage process of the coal sample, including acoustic emission signals and axial load; analyzing the acoustic emission signals by using ringing count and acoustic emission energy; performing correlation analysis of the acoustic emission signals and the coal sample fracture mechanism; constructing a coal sample fracture state identification model based on deep learning for identifying the crack distribution of the coal sample; and constructing a coal sample fracture early warning model based on dynamic characteristics for converting the output of the coal sample fracture state identification model into an early warning decision. The coal sample fracture state identification model proposed in the present application realizes efficient feature extraction of the coal sample fracture precursor acoustic emission signal under the application of deep separable convolution, attention mechanism and TCN time series convolution structure.

Description

Technical Field

[0001] This invention belongs to the field of coal mine dynamic disaster prevention and control technology, specifically involving a method for identifying precursor information of acoustic emission characteristics of coal-rock complex fractures and providing short-term early warning. Background Technology

[0002] Acoustic emission (AE) technology, as a direct monitoring method for coal and rock failure processes, can further address the problem of dynamic disasters in coal and rock formations. AE detection systems can acquire and interpret waveforms in the 10kHz-10MHz spectrum, making them currently the best monitoring system in terms of scale effect (temporal resolution) for identifying the development and eventual connection of micro-fractures. Through various characteristic parameters, such as ring count, energy entropy, and fractal dimension, it can comprehensively describe the entire process of fracture development (characteristic characterization capability), and can perform three-dimensional location calculations through an array system (spatial characterization capability). Therefore, AE technology can provide a full understanding of the catastrophic laws of rock masses and offer corresponding causal evidence. However, current engineering research on AE mainly focuses on various physical parameters, lacking research on the characteristic laws, physical mechanisms, and interactions that characterize fracture development and evolution within a large number of AE waveforms. This results in certain limitations in the engineering application of the established early warning models.

[0003] In recent years, scholars both domestically and internationally have conducted a series of experimental studies on the prevention and control of coal mine dynamic disasters, achieving significant progress in disaster mechanisms and related technologies. However, current research still has shortcomings in experimental data analysis, mainly in the following aspects: (1) Difficulties in extracting and analyzing precursor features of acoustic emission signals: The uniaxial compression failure test of coal samples releases a large amount of acoustic emission signals containing a wealth of damage evolution information. However, extracting physically meaningful damage features from these AE signals remains a challenge: how to select core parameters by characterizing the AE response throughout the entire compression test process while maintaining prediction accuracy, thus reducing computational load; and how to establish a quantitative analysis method to explore the temporal evolution characteristics of precursor signals. Based on this, this invention employs temporal density to dynamically quantize the AE signal sequence and identifies early warning thresholds through the detection of statistical outliers, providing data support for the subsequent establishment of an intelligent early warning model.

[0004] (2) The accuracy of coal sample damage early warning models: Based on the above research on the selection of precursor features, ensuring the accuracy of the early warning model is another key issue that needs to be addressed. Conventional techniques have limitations in capturing nonlinear damage information of coal bodies, while neural network models have strong fitting characteristics for nonlinearity and can dynamically identify the nonlinear relationship between AE features and damage states during fitting learning. This invention aims to study an identification model that combines neural network models with acoustic emission technology to predict coal sample damage. By training the prediction model's network, the spatial and temporal distribution of precursor signals is identified, optimizing the design of the entire early warning model and ultimately providing an accurate real-time early warning solution for the field, from the feature layer to the early warning layer. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a method for identifying precursor information of acoustic emission characteristics in coal-rock composite fractures and providing short-term early warning, thereby resolving the issues in the prior art. The technical solution adopted by this invention is as follows: A method for identifying precursor information and providing short-term early warning of acoustic emission characteristics in coal-rock complex fractures includes the following steps: Step 1: Conduct a uniaxial compression test on the coal sample. Apply gradually increasing axial loads to the coal sample until it fails. Collect test data during the failure process, including acoustic emission signals and axial loads. Step 2: Based on the experimental data, construct the stress-strain curve of the uniaxial compression test, and analyze the acoustic emission signal using ring counting and acoustic emission energy analysis; Step 3: Based on the temporal characteristics of the acoustic emission signal, conduct a correlation analysis between the acoustic emission signal and the coal sample fracture mechanism; Step 4: Construct a deep learning-based coal sample fracture state recognition model to identify the crack distribution in the coal sample; Step 5: Construct a coal sample fracture early warning model based on dynamic features to convert the output of the coal sample fracture state identification model into early warning decisions.

[0006] Furthermore, in step 3, the coal sample fracturing mechanism includes three stages: crack initiation, crack propagation, and final fracturing.

[0007] Furthermore, the abrupt change point of the ringing count is obtained using the cumulative sum method to reflect the stage transition of coal sample fracturing, as shown in the formula: In the formula, This is the current ringing count value, used to reflect the activity level of the acoustic emission signal. The baseline mean. To allow for fluctuation threshold, For the cumulative deviation, when Exceeding the set threshold At that time, it is determined that a sudden change has occurred in the ring count.

[0008] Furthermore, step 4 includes: Step 4.1: Divide the acoustic emission signal into time window units, extract the acoustic emission energy and ring count of each time window unit, perform fast Fourier transform on the acoustic emission energy and ring count to generate the corresponding windowed spectrum image, and generate a time-series feature sequence. Step 4.2: Using ResNet18 as the basic backbone network, construct a parallel dual-branch model architecture, with the two branches being the acoustic emission signal image processing branch and the time-series signal processing branch; Step 4.3: Based on the acoustic emission signal image processing branch, extract the spatial features of the acoustic emission signal and output the image spatial feature vector; Step 4.4: Based on the time-series signal processing branch, extract the time-series dynamic features of the acoustic emission signal and output the time-series dynamic feature vector; Step 4.5: Concatenate the image spatial feature vector and the temporal dynamic feature vector along the channel dimension, output the multi-domain fusion feature vector to the fully connected layer, and obtain the final coal sample state classification result through the Softmax classifier.

[0009] Furthermore, step 4.3 includes: Step 4.31, input the windowed spectral image; Step 4.32: Perform initial feature extraction on the input windowed spectral image and output the initial feature map; Step 4.33: Perform dimensionality reduction on the initial feature map and output the dimensionality-reduced feature map; Step 4.34: Set at least 4 cascaded residual blocks in sequence, and embed a DSA module in each residual block to process the dimensionality-reduced feature map; Step 4.35, Global Average Pooling Processing: Through the global average pooling layer, global feature aggregation is performed on the feature map output by the last residual block to output the image space feature vector.

[0010] Furthermore, in step 4.34, the processing procedure of the DSA module includes: Perform pointwise convolution on the input feature map to compress the channel dimension to Local spatial information is obtained through DSC deep convolutional layers, and the dimension of the space is compressed to 1 through max pooling. The convolutional kernel restores the channel dimension to the number of channels in the input feature map, which is used to obtain the channel attention weights for each channel in the output feature map. Finally, the channel attention weights of each channel in the input feature map are compared. The product of these is used as the weighted output feature map; The input feature map is convolved to obtain a convolved feature map. Then, a transformation function is used to obtain channel attention weights, which are then multiplied element-wise with the original feature map to obtain a weighted channel attention feature map. Finally, a residual connection is used to concatenate the input feature map with the weighted channel attention feature map to obtain a new output feature map, represented as: .

[0011] Furthermore, in step 4.34, channel attention weights... The calculation formula is: In the formula, The input feature map; This is a pointwise convolution operation; yes Multiply by depthwise convolution; It is a space max pooling operation. Represent a The convolution is performed and includes a Sigmoid activation function.

[0012] Furthermore, step 4.4 includes: Input time series feature sequence The TCN module will input the time series feature sequence. After processing through causal convolutional layers, the receptive field is expanded in each causal convolutional layer using kernel dilation, as represented by: in, For expansion rate, The kernel size is [size]. For convolution kernel weights, For time step The convolution result; Causal convolutional layers use residual connections to process the input temporal feature sequences. The features are summed to obtain the output features. The time dimension of the output features is normalized by global average pooling to output a time-series dynamic feature vector.

[0013] Furthermore, step 5 includes: Step 5.1: Probabilize the output risk of the rupture state identification model to obtain the rupture probability value; Step 5.2: Reprocess the fracture probability value to construct a dynamic threshold model and a multi-index weighted fusion model, and determine the state of the coal sample based on the output of the multi-index weighted fusion model.

[0014] Further, step 5.1 includes: based on the coal sample fracture state identification model, in step 4.5, setting the output layer of the final fully connected layer to multiple neurons to correspond to the loading stage, elastoplastic stage, and fracture precursor stage; then setting a Softmax activation function after the output layer of the fully connected layer, and converting the original score of each stage into a probability distribution; the modified output vector is: In the formula Let the number of state categories be denoted by '[', ... As the probability of rupture Output, defined as: Step 5.2 includes: Step 5.21, construct the dynamic threshold model, which is expressed by the following formula: In the formula, , and These are the mean values ​​of the probability surge rate, energy surge rate, and ring count mutation point within the current statistical window, respectively. , and These are the standard deviations of the corresponding indicators; , and This is an empirical coefficient; This represents the probability increase rate. The rate of change of energy; This is a sudden change in the ring count; Step 5.22, construct a multi-indicator weighted fusion model, represented as: In the formula, , and These are the weighting coefficients; The threshold for joint decision-making; when the multi-indicator weighted fusion model... When the output exceeds the warning threshold, it is determined that the coal sample is at risk of rupture.

[0015] The present invention has the following beneficial effects: (1) In order to overcome the shortcomings of traditional statistical methods for analyzing nonlinear and temporal acoustic emission signals, this invention proposes a coal sample fracture state identification model based on ResNet18. With the application of depthwise separable convolution, attention mechanism and TCN temporal convolution structure, efficient feature extraction of acoustic emission signals preceding coal sample fracture is achieved. (2) This invention proposes an online early warning method that includes dynamic threshold determination and multi-indicator fusion weighting. It achieves three levels of early warning (low, medium, and high) by weighting and fusing multiple factors such as the probability of rapid increase rate, the energy of rapid increase rate, and the number of ringing abrupt changes, and by using continuous anomaly detection logic. (3) This invention has achieved a series of innovative results in the explanation of coal sample fracture mechanism, development of deep learning model and construction of early warning system. The proposed dynamic threshold early warning method based on multi-information fusion has value in engineering applications. Attached Figure Description

[0016] Figure 1 This is a flowchart of the present invention; Figure 2 Stress-strain curves for each coal sample specimen; Figure 3 Acoustic emission signals and stress diagrams for each coal sample specimen; Figure 4 Diagram of the coal sample fracture state identification model architecture; Figure 5 A schematic diagram of dilated causal convolution; Figure 6 This is a schematic diagram of the TCN residual structure; Figure 7 The time-series fracture probability of each coal sample specimen; Figure 8 The time window threshold trend for each coal sample specimen. Detailed Implementation

[0017] The following will be described in conjunction with embodiments of the present invention. Figures 1-8 The technical solutions in the embodiments of the present invention will be clearly and completely described. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Unless otherwise specified, the technical means used in the embodiments are conventional means well known to those skilled in the art.

[0018] like Figure 1 This invention proposes a method for identifying precursor information of acoustic emission characteristics of coal-rock complex fractures and providing short-term early warning, comprising the following steps: Step 1: Conduct a uniaxial compression test on the coal sample. Apply gradually increasing axial loads to the coal sample until it fails. Collect test data during the failure process, including acoustic emission signals and axial loads. This invention uses uniaxial compression tests to obtain the response characteristics of coal samples under different stress states, including stress-strain curve types, peak intensity variation patterns, and failure mode categories, providing a theoretical basis for further understanding coal mechanical characteristics and early warning of coal-rock dynamic disasters. By combining acoustic emission technology to collect elastic wave signals generated during coal sample cracking, the relationship between acoustic emission characteristic parameters (energy, count, peak value, spectrum) and cracking stages can be explored, providing necessary information for characterizing the cracking evolution stages.

[0019] The coal sample of this invention is obtained from the undisturbed coal seam through core drilling, cutting, and surface finishing. It is preferably cylindrical with dimensions of Φ50mm × 100mm. The uniaxial compression test is performed using an electro-hydraulic rock servo pressure testing machine to compress the coal sample. Acoustic emission signals are acquired using an acoustic emission instrument.

[0020] During the uniaxial compression test, four RS-2A type resonant sensors were placed at 15 mm and 85 mm axial heights of the coal sample (avoiding the end friction interference area) to form upper and lower double detection planes. Within each detection plane, the four sensors were distributed at equal 90° intervals and fixed by adhesive adsorption converters. Signal pre-amplification was achieved through a pre-charge amplifier (gain 40dB).

[0021] Step 2: Based on the experimental data, construct the stress-strain curve of the uniaxial compression test, and analyze the acoustic emission signal using ring counting and acoustic emission energy analysis; Uniaxial compressive stress-strain curves can intuitively and comprehensively reflect the compressive mechanical behavior of coal samples. Their loading evolution process represents the entire mechanical evolution of the specimen from elastic deformation to irreversible deformation and instability failure. Based on the mechanical properties of actual coal samples and the stability characteristics of quasi-static test loads, a standard coal sample (Φ50×100mm) was subjected to a compression test under quasi-static conditions with a loading rate of 0.001mm / s. Combined with a high sampling rate (100Hz) compression data acquisition system, the entire uniaxial compression lifecycle test data of the coal sample can be obtained with relatively high accuracy. Figure 2 In specific experiments, the stress-strain curves show that the uniaxial compression process of each coal sample can be divided into four stages: linear elastic deformation stage (OA), nonlinear yielding stage (AB), plastic strengthening stage (BC), and post-peak strain softening stage (CD).

[0022] When analyzing acoustic emission signals, this invention uses ring count and acoustic emission energy as the main acoustic emission characteristic parameters; wherein: Ringing count refers to the number of times acoustic emission pulses exceed a given threshold within a certain time period; its purpose is to characterize the activity level of the acoustic emission signal. For example... Figure 3In the initial stage of the experiment, microcrack activity was not intense, and the number of ringings was small. When the applied stress increased to the accelerated growth stage, the number of acoustic emission signals increased significantly, the number of ringings surged and reached its maximum value, then decreased sharply upon sample failure. This indicates that the number of ringings can directly show the process of the sample from slow crack initiation to rapid crack propagation and even sample failure. Analyzing the cumulative trend curve of the ringing count can directly reflect the abrupt changes in energy during the loading process of the coal sample, especially the abrupt changes during the fracture process. Therefore, this invention uses the Cumulative Sum (CUSUM) method, i.e., the CUSUM algorithm, to obtain the instability value of the coal sample fracture state by finding the change points of the ringing count. The CUSUM algorithm accumulates the deviations of the ringing count sequence from the baseline mean, and judges a sudden change when the deviation exceeds a pre-specified value. The formula is: In the formula, This is the current ringing count, which directly reflects the activity level of the acoustic emission signal. The baseline mean is typically selected from the average ring counts at the initial stage of coal sample loading (i.e., when the stress is below 30% of the peak value) to ensure the stability of the baseline data. A tolerance threshold is set to avoid interference from small random fluctuations, making mutation detection more robust. For the cumulative deviation, when Exceeding the set threshold (usually taken) ,in When the standard deviation of the ring count during the baseline phase is used, a sudden change in the ring count is determined.

[0023] Acoustic emission energy is a marker of crack propagation and concentrated energy release in coal samples during failure. Experimental results show that acoustic emission energy is low and fluctuates little during the initial and steady-state phases. However, during the acceleration phase, as numerous cracks propagate and penetrate the coal sample, energy release rapidly increases, peaking during the unstable failure phase, exhibiting typical main fracture signals. This abrupt change in energy release not only indicates the rapid increase in internal cracks but also demonstrates that after the coal sample forms numerous microcracks over a long period, it ultimately completes the failure process through a concentrated release of large amounts of energy. Step 3: Based on the temporal characteristics of the acoustic emission signal, conduct a correlation analysis between the acoustic emission signal and the coal sample fracture mechanism; The time-domain characteristic parameters of acoustic emission signals include amplitude, frequency, time domain, and frequency domain. Coal sample fracturing can be divided into three main stages: crack initiation, crack propagation, and final fracturing. In the crack initiation stage, the initial formation of microcracks results in acoustic emission signals exhibiting low amplitude and high frequency characteristics, reflecting the initial dynamic process of crack formation within local stress concentration areas. In the crack propagation stage, the interconnection and coordinated propagation between cracks lead to an increase in signal amplitude while the frequency characteristics shift towards lower frequencies; that is, during fracturing, the energy dissipation channel gradually tends towards the direction of macroscopic crack propagation. In the final fracturing stage, due to the occurrence of macroscopic fracture, the acoustic emission signal clearly exhibits low frequency and high energy, providing a certain intuitive standard for determining the instability process of the specimen.

[0024] Specifically: Crack initiation stage (energy threshold <50mV·ms): Acoustic emission signal exhibits low amplitude ( High frequency (centroid frequency > 150kHz) characteristics, with RMS value maintained below 15mV, corresponding to the elastic deformation of micro-defects inside the coal body; Crack propagation stage (50-200 mV·ms): Signal amplitude significantly increased ( The ring count increased exponentially. The peak frequency shifts to the mid-frequency domain (80-120kHz), reflecting the fractal propagation characteristics of the crack; Macroscopic fracture stage (>200mV·ms): Typical high-energy events occur, with a duration exceeding 5ms. Energy and duration show a significant positive correlation, characterizing the formation of through fractures.

[0025] Where ASL represents the average signal level, The coefficient of determination is given. It is worth noting that the energy of the sampling events in the macroscopic fracture segment changes abruptly, and the logarithmic coordinates of the event duration and event energy show a linear change (slope of 1.32±0.15), which satisfies the Griffith crack propagation theory; low-frequency events (<50kHz) are highly correlated with the slip of the bedding plane.

[0026] Step 4: Construct a deep learning-based coal sample fracture state recognition model to identify the crack distribution in the coal sample; The overall coal sample fracture state identification model is as follows: Figure 4 As shown, it uses ResNet18 as its base network. A key feature of ResNet is the use of "identity mapping connections." By adding identity mapping connections to the CNN, the output of each layer in the CNN can be well controlled through residual learning, avoiding the gradient vanishing phenomenon during network learning, thereby enabling more effective extraction of signal features.

[0027] Overall, this invention inputs the ring count and acoustic emission energy of the acoustic emission signal into the image processing branch via a windowed image obtained through Fast Fourier Transform (FFT). The input samples are initially processed through a 7×7 convolutional layer. Then, the features are input into four residual blocks for feature extraction. A DSA module is embedded in each residual block. This DSA module uses depthwise separable convolution to obtain an attention map composed of channel weights, which describe the interdependencies between channels. By multiplying the features point-by-point with the attention map, the network can adaptively emphasize information objects in the image. Then, the TCN module extracts the temporal dependency features of the corresponding input sequence through a series of 1D convolutional layers and dilated convolutions. Finally, the image feature vectors from the DSA and the temporal feature vectors from the TCN are fused to output the final coal sample state identification result.

[0028] Specifically, step 4, the processing steps of the coal sample fracture state identification model, includes: Step 4.1: Divide the acoustic emission signal into a time-series window unit of 256 sampling points, extract the acoustic emission energy and ring count of each time-series window unit, perform Fast Fourier Transform (FFT) on the acoustic emission energy and ring count to generate the corresponding windowed spectrum image, and generate a time-series feature sequence; combine the stress-strain curve of the coal sample with the fracture evolution stage, divide all acoustic emission signal samples into three categories: loading stage samples, elastoplastic stage samples, and fracture precursor stage samples. Use the acoustic emission signals of different coal sample specimens as training datasets and validation datasets, respectively.

[0029] In practice, among the 9442 time-series feature samples, there were 2096 samples in the loading stage, 5952 samples in the elastoplastic stage, and 1394 samples in the pre-rupture stage. Specimens 1-3 were used as the training dataset, and specimen 4 was used as the validation dataset.

[0030] Step 4.2: Using ResNet18 as the basic backbone network, construct a parallel dual-branch model architecture, with the two branches being the acoustic emission signal image processing branch and the time-series signal processing branch; Step 4.3: Based on the acoustic emission signal image processing branch, extract the spatial features of the acoustic emission signal and output the image spatial feature vector; Step 4.4: Based on the time-series signal processing branch, extract the time-series dynamic features of the acoustic emission signal and output the time-series dynamic feature vector; Step 4.5: Concatenate the image spatial feature vector and the temporal dynamic feature vector in the channel dimension, output the multi-domain fusion feature vector to the fully connected layer, and obtain the final coal sample state classification result through the Softmax classifier. This ensures that different feature dimensions are complementary, maintaining both the frequency domain features and the dynamic change features of the signal.

[0031] Specifically, step 4.3 includes: Step 4.31: Input acoustic emission energy and ringing count windowed spectrum images with dimensions (B, 224, 224, 3), where B is the batch size; Step 4.32, Initial convolutional layer processing: Set a 7×7 convolutional layer, a BN layer and a ReLU activation function to perform initial feature extraction on the input windowed spectral image, and output an initial feature map with dimensions (B, 112, 112, 64). Step 4.33, Max Pooling Layer Dimensionality Reduction: Set a 3×3 max pooling layer to reduce the dimensionality of the initial feature map, and output a feature map with dimensions (B, 56, 56, 64). Step 4.34, Residual Block Feature Extraction: Set at least 4 concatenated residual blocks in sequence, and embed a DSA module (Depthly Separable Convolutional Attention Module) into each residual block to process the feature map output in Step 4.33; Among them, the feature map of the output dimension (B,56,56,64) of residual block 1; the feature map of the output dimension (B,28,28,128) of residual block 2; the feature map of the output dimension (B,14,14,256) of residual block 3; and the feature map of the output dimension (B,7,7,512) of residual block 4. Step 4.35, Global Average Pooling Processing: Through the global average pooling layer, global feature aggregation is performed on the feature map output by the last residual block 4, and the output image space feature vector has a dimension of (B, 512).

[0032] Specifically, in step 4.34, the processing procedure of the DSA module includes: The input feature map is subjected to pointwise convolution to fuse the features of each channel, compressing the channel dimension to [value missing]. Local spatial information is obtained through DSC deep convolutional layers, and the dimension of the space is compressed to 1 through max pooling. The convolutional kernel restores the channel dimension to the number of channels in the input feature map, thereby obtaining the specific channel attention weights for each channel. Finally, the product of this product and the channel attention weight of each channel in the input features is the weighted feature. The formula for calculating the channel attention weights obtained through the convolution kernel is: (4-6) In the formula, For input features. For pointwise convolution operations, it reduces the input dimension from... Reduced to . yes Multiply by depthwise convolution. It is a space max pooling operation. Represent a The convolution is performed and includes a Sigmoid activation function, which ensures that the channel dimension of the output feature is the same as that of the input channel.

[0033] The DSA module is applied to the residual block of ResNet18 to perform convolution on the input feature map to obtain a more abstract feature map. Then, channel attention weights are obtained through transformation function operations, and then multiplied element-wise with the feature map to obtain weighted channel attention features. Finally, the original input features are connected with the weighted features through residual connections to obtain a new output feature map. The whole process is represented as follows: The processing flow of the DSC depthwise convolutional layer is as follows: The DSC deep convolutional layer consists of two parts: depthwise convolution (Conv-dw) and pointwise convolution (Conv-pw). It primarily separates the feature extraction functions between and within channels, thereby reducing the parameters and computational cost of the convolutional network and improving its classification performance.

[0034] Given input features ,in and For spatial dimensions, The number of channels, with a kernel size of From standard convolution, we can obtain the output. ,in and For spatial dimensions, This represents the number of output channels.

[0035] To reduce the computational parameters in standard convolution, depthwise separable convolution first convolves the input features with a depthwise convolution, which can be formulated as: in, For input, Use a depthwise convolution kernel. Then, add a kernel with... Pointwise convolution of the convolution kernel is applied to the output of depthwise convolution, which can be formulated as follows: In the formula, The output of depthwise convolution. for convolution, This is the output of the depthwise separable convolution. The computational cost of depthwise separable convolution is estimated as follows: Considering that the standard convolutional kernel size is The estimated computational workload is as follows: The ratio of the two is: In practical applications, the number of output channels and the kernel size are typically much larger than 1 (i.e., Therefore, depthwise separable convolution can effectively compress the number of parameters and computational cost in convolution while achieving the same convolution effect as standard convolution.

[0036] Specifically, step 4.4, the processing procedure of the TCN module includes: Considering that the data collected by the acoustic emission experiment is a continuous time series with high complexity and large data volume, a TCN module was added to extract the temporal features of the samples. It uses convolutional layers instead of recurrent layers to handle sequence dependencies.

[0037] like Figure 5 Given an input sequence The TCN module processes the input sequence through causal convolutional layers. These convolutions ensure that the output at each time step depends only on the current and previous time sequences, thus preserving the causality of temporal dependencies. Subsequently, within each convolutional layer, the receptive field is expanded using kernel dilation, enhancing the ability to capture temporal dependencies. This is represented as: in, For expansion rate, The kernel size is [size]. For convolution kernel weights, For time step The convolution result.

[0038] The convolutional layer sums the input features through residual connections to obtain the output features. Global average pooling then normalizes the temporal dimension of the output features, forming a temporal feature vector. The basic structure is as follows: Figure 6 As shown, TCN, through the use of causal convolution, dilated convolution, and residual connection operations, not only captures long-term and short-term dependencies in time series, but also ensures efficient computation and a stable model.

[0039] This invention designs a coal sample fracture state recognition model based on ResNet network, which integrates deep separable convolution, temporal convolution, and self-attention mechanism. The model uses deep separable convolution to reduce the feature learning process and simplify the model structure; it uses self-attention mechanism to quickly extract local features, which improves the model's feature extraction capability while accelerating the calculation speed; and it uses temporal convolution network to obtain the long-term sequence dependence of acoustic emission signal, which can more effectively extract damage features at each stage.

[0040] This invention focuses on a ResNet model improved based on a deep separable attention mechanism and temporal convolution. It investigates how the deep separable convolution module helps the model extract features at multiple scales and improves the calculation method of attention weights in the self-attention module. A temporal convolution structure suitable for acoustic emission signals is designed, and a multi-domain feature fusion network suitable for acoustic emission signals is constructed. A feature pyramid structure facilitates closer information interaction between different scales. The above method can more accurately identify the crack distribution in coal sample images under different fracture states, possesses good versatility, and provides a feasible method for early warning of coal and rock dynamic disasters. Furthermore, this invention enriches the application of deep learning in the field of coal and rock dynamics and lays the foundation for solving the problem of constructing intelligent monitoring methods under the conditions of multi-physics field coupling.

[0041] Step 5: Construct a coal sample fracture early warning model based on dynamic features to convert the output of the coal sample fracture state identification model into early warning decisions.

[0042] In coal mine production environments, the reliable operation of coal sample fracture early warning systems is crucial for mine safety and the development of emergency response plans. The ResDSATCNet model proposed in this invention extracts key features from acoustic emission signals and mechanical information to accurately classify the loading stages of coal samples at different fracture stages. However, single-state identification results cannot meet the requirements for timely and accurate early warnings in actual coal mine environments. Therefore, realizing a method chain that transforms the output of the state classification model into early warning decisions is key to dynamic safety monitoring in coal mines.

[0043] Step 5, based on the coal sample fracture state identification model proposed in Step 4, mainly focuses on post-processing the model's output probability information, realizing a dynamic threshold judgment and multi-information fusion early warning mechanism. This early warning mechanism can promptly detect abnormal precursor information during the coal sample loading process and can issue early warnings for low, medium, and high-level abnormal state information by statistically analyzing the number of consecutive abnormalities, thus achieving sensitivity and targeting in early warning. Specifically, deep learning technology is used to classify the state images during the coal sample fracture process. Convolutional neural networks are used to extract image features, and fully connected layers output the discrimination results for each state category. The results are then visualized in the form of image states, showing different states of the coal sample, such as normal loading, elastoplastic stage, and signs of fracture. However, to implement the early warning system, it is still necessary to probabilize the output of the coal sample fracture state identification model to obtain the "fracture probability." The continuous time-series information provides a quantitative basis for subsequent dynamic threshold determination and online early warning.

[0044] The aforementioned multi-level alarm strategy can avoid false alarms caused by data anomalies at a certain moment, and can also prevent critical rupture events in a timely manner. The network rupture probability is weighted and fused by combining the extracted results of abnormal acoustic emission energy growth rate and ringing count mutation threshold to achieve joint application of multiple physical quantities.

[0045] Specifically, step 5 includes: Step 5.1: Probabilize the output risk of the rupture state identification model to obtain the rupture probability value.

[0046] Based on the original coal sample fracture state identification model, in step 4.5, the output layer of the final fully connected layer is configured with multiple neurons to correspond to different state types (loading stage, elastoplastic stage, and pre-fracture stage). Then, a Softmax activation function is added after this output layer, converting the original score for each type into a probability distribution; the modified output vector is: In the formula This represents the number of state categories. To adapt to the early warning system, the probability representing the "precursor stage of rupture" is... As the probability of rupture The output, the model's predicted probability value output once per time period, reflects the fracturing probability of the coal sample within that time period. It can be defined as follows: like Figure 7 By analyzing the characteristics of fracture probability changes across multiple time windows, the system revealed the stage evolution law of coal sample deformation and failure. Experimental data show that: The fracture probability curve distribution during the loading stage is at a low point of 0-20%, indicating that the state of the specimen is relatively stable during this stage. When the external load continues to act on the specimen, there is no obvious increase in destructive force, and there is no accumulation of damage area on the specimen. The mechanical response is characterized by elastic deformation.

[0047] The fracture probability curve gradually increases during the elastoplastic stage, rising from approximately 20% to around 60%. At this point, a small number of microcracks appear and propagate in the specimen, accumulating damage and forming cumulative damage, leading to a gradual increase in the fracture probability. For example, the fracture probability curve of specimen 1 shows several sudden increases in the 15-25 minute stage at the beginning of loading, which is the mechanical response law of the specimen undergoing elastic deformation to plastic deformation.

[0048] The probability curve of coal sample rupture during the pre-rupture stage shows a sharp increase, rapidly approaching 100%, indicating that the cracks in the coal sample extend and penetrate rapidly, and the damage evolution is accelerating. The exponential growth characteristic of the rupture probability directly shows the critical state before and after the instability and failure of the coal sample, providing a reference for the identification of early warning signs of engineering disasters.

[0049] Step 5.2 involves reprocessing the fracture probability values ​​predicted by the fracture state identification model to construct a dynamic threshold model and a multi-index weighted fusion model. The coal sample state is then determined based on the output of the multi-index weighted fusion model. Specifically, the correlation between the probability curve and energy and ring count is determined based on the correlation of time series data, thus constructing the dynamic threshold model. The fracture probability values ​​output by the fracture state identification model are introduced and linearly aggregated with the acoustic emission energy surge rate and the ring count mutation index to construct the multi-index weighted fusion model. Three consecutive outliers are used for determination, i.e., the grade response.

[0050] Specifically, step 5.2 includes: Step 5.21, Construct a dynamic threshold model: To improve the reliability and anti-interference capability of early warning systems, this invention designs a dynamic threshold model. This model mainly includes the following three key indicators: (1) Probability surge rate This refers to the rate of change of the "rupture probability value" of the model within several time windows. By calculating the average and standard deviation of the probabilities for different time windows, an empirical critical value is selected to determine the occurrence of anomalies.

[0051] (2) Rate of change of energy When energy increases rapidly, it often indicates an intensification of crack activity inside the coal sample.

[0052] (3) Sudden change in ring count A sudden increase in the ringing count is also a significant characteristic before coal sample fracturing.

[0053] The dynamic threshold model is expressed by the following formula: In the formula, , and These are the mean values ​​of the probability surge rate, energy surge rate, and ring count mutation point within the current statistical window, respectively. , and These are the standard deviations of the corresponding indicators; , and These are empirical coefficients, determined through fitting historical data. Parameter settings are shown in Table 1.

[0054] Table 1: Parameter Settings

[0055] Step 5.22, construct a multi-indicator weighted fusion model: To comprehensively reflect information from multiple indicators, this invention constructs a multi-indicator weighted fusion model, which is defined as follows: In the formula, , and The weight coefficients for each indicator are obtained through principal component analysis, as detailed in Table 2. This is the threshold for determining joint decision-making.

[0056] Table 2: Weighting coefficients of multi-index principal component analysis

[0057] when When the output exceeds the warning threshold, the system will initially consider the coal sample to have a risk of fracturing and trigger a warning signal. Simultaneously, to prevent false alarms caused by occasional data fluctuations, this invention introduces a mechanism for determining the number of consecutive anomalies: (1) When a single anomaly occurs (i.e., the current time window data exceeds the threshold), the system triggers a low-risk warning, which corresponds to the end of the elastoplastic stage when the probability value first exceeds the 55% threshold in the experiment; (2) If there are two consecutive abnormalities, the warning level will be upgraded to a medium risk warning, which corresponds to the continuous increase of red data points in the experiment; (3) If there are three or more consecutive abnormalities, the system will determine it as a high-risk warning, which corresponds to the rupture precursor stage in the experiment where the probability value rises sharply.

[0058] The results are as follows Figure 8 As shown, under laboratory conditions, the test was divided into a loading stage, an elastic-plastic stage, and a pre-breakage stage. 55% was used as the primary judgment threshold, with red dots representing "greater than the threshold (triggered warning)" and blue dots representing "less than the threshold (no warning triggered)," which accurately reflects the test phenomena. The specimen in the figure is the test coal sample.

[0059] (1) Test analysis of specimen 1: The test results for specimen 1 show that the probability of the loaded portion fluctuates below 5%, ranging from 1% to 10%. The small fluctuation of the blue dots indicates that the probability determination for this portion is effective, consistent with the model's performance in the initial stage. The probability of the elastoplastic stage increases in a wave-like manner, reaching 55%. The probability of the pre-failure stage rises rapidly from 55% to 98%, with the red dots showing an exponential increase. This aligns with the model's ability to "determine and warn of pre-failure signs," and matches the characteristics of pre-failure signs in this stage.

[0060] (2) Test analysis of specimen 2: In the test results of specimen 2, the probability value oscillated between 10% and 20% after the specimen entered the loading stage. After the specimen entered the elastoplastic stage, it gradually increased in a "stepwise" manner: there were four consecutive stepwise increases in the 1500-3000 time window, a plateau period of 52%-54% in the 3800-4000 time window, and continued to rise after breaking the threshold in the 4000 time window. The specimen reached a red high-risk warning after exceeding the threshold three times in a row using continuous anomaly judgment.

[0061] (3) Test analysis of specimen 3: The probability value of specimen No. 3 during the loading stage was basically between 8% and 18%, and the model was in a non-triggered state. The probability value of the elastoplastic stage fluctuated and increased until it exceeded 55% in the 1200-time window, after which it dropped rapidly. According to the continuous anomaly logic, when the threshold was exceeded for the first time, the model reached the pre-rupture stage in the 1300-time window, and the probability value increased significantly from 55% to 99%. The red dot increased significantly, and the evolution process of "stability - development - early warning" was fully demonstrated, which can prove the model logic of judging according to the threshold criterion.

[0062] (4) Test analysis of specimen 4: Analysis of specimen 4 data shows that the probability of loading oscillates around 10%, and the model continuously outputs a no-trigger signal. The probability of the elastoplastic stage gradually increases, reaching 55% within approximately a 900-second time window. The probability of the pre-fracture stage rises sharply to 99%, and the changes in each stage are accurately mapped to the mechanical state within the model framework.

[0063] In four comparative experiments, 55% of the parameters were used as key model parameters, correctly defining the relationship between the experimental process and the determination of coal sample fracture state. The average deviation between the occurrence of continuous abnormal events and the time of experimental crack in the measured model was 3.2%, which reduced the misjudgment rate by 28.7% compared to a single threshold model. This is because the model incorporates the concept of continuous anomalies into the judgment, providing both a judgment element with an upward probability of fracture precursor periods and suppressing the occurrence of random anomalies, thus improving the probability of effective judgment. This invention is applicable to various scenarios such as mineral geological exploration services utilizing high technology, computer-aided monitoring and early warning, and detection of new coal and rock materials.

[0064] The above embodiments are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Any modifications, alterations, alterations, or substitutions made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for identifying precursor information and providing short-term early warning of acoustic emission characteristics in coal-rock complex fractures, characterized in that, Includes the following steps: Step 1: Conduct a uniaxial compression test on the coal sample. Apply gradually increasing axial loads to the coal sample until it fails. Collect test data during the failure process, including acoustic emission signals and axial loads. Step 2: Based on the experimental data, construct the stress-strain curve of the uniaxial compression test, and analyze the acoustic emission signal using ring counting and acoustic emission energy analysis; Step 3: Based on the temporal characteristics of the acoustic emission signal, conduct a correlation analysis between the acoustic emission signal and the coal sample fracture mechanism; Step 4: Construct a deep learning-based coal sample fracture state recognition model to identify the crack distribution in the coal sample; Step 5: Construct a coal sample fracture early warning model based on dynamic features to convert the output of the coal sample fracture state identification model into early warning decisions.

2. The method for identifying precursor information and providing short-term early warning of acoustic emission characteristics of coal-rock complex fractures according to claim 1, characterized in that, In step 3, the coal sample fracturing mechanism includes three stages: crack initiation, crack propagation, and final fracturing.

3. The method for identifying precursor information and providing short-term early warning of acoustic emission characteristics of coal-rock composite fractures according to claim 2, characterized in that, The abrupt change point of the ringing count is obtained using the cumulative sum method to reflect the stage transition of coal sample fracturing. The formula is as follows: In the formula, This is the current ringing count value, used to reflect the activity level of the acoustic emission signal. The baseline mean. To allow for fluctuation threshold, For the cumulative deviation, when Exceeding the set threshold At that time, it is determined that a sudden change has occurred in the ring count.

4. The method for identifying precursor information and providing short-term early warning of acoustic emission characteristics of coal-rock complex fractures according to claim 1, characterized in that, Step 4 includes: Step 4.1: Divide the acoustic emission signal into time window units, extract the acoustic emission energy and ring count of each time window unit, perform fast Fourier transform on the acoustic emission energy and ring count to generate the corresponding windowed spectrum image, and generate a time-series feature sequence. Step 4.2: Using ResNet18 as the basic backbone network, construct a parallel dual-branch model architecture, with the two branches being the acoustic emission signal image processing branch and the time-series signal processing branch; Step 4.3: Based on the acoustic emission signal image processing branch, extract the spatial features of the acoustic emission signal and output the image spatial feature vector; Step 4.4: Based on the time-series signal processing branch, extract the time-series dynamic features of the acoustic emission signal and output the time-series dynamic feature vector; Step 4.5: Concatenate the image spatial feature vector and the temporal dynamic feature vector along the channel dimension, output the multi-domain fusion feature vector to the fully connected layer, and obtain the final coal sample state classification result through the Softmax classifier.

5. The method for identifying precursor information and providing short-term early warning of acoustic emission characteristics of coal-rock complex fractures according to claim 4, characterized in that, Step 4.3 includes: Step 4.31, input the windowed spectral image; Step 4.32: Perform initial feature extraction on the input windowed spectral image and output the initial feature map; Step 4.33: Perform dimensionality reduction on the initial feature map and output the dimensionality-reduced feature map; Step 4.34: Set at least 4 cascaded residual blocks in sequence, and embed a DSA module in each residual block to process the dimensionality-reduced feature map; Step 4.35, Global Average Pooling Processing: Through the global average pooling layer, global feature aggregation is performed on the feature map output by the last residual block to output the image space feature vector.

6. The method for identifying precursor information and providing short-term early warning of acoustic emission characteristics of coal-rock composite fractures according to claim 5, characterized in that, In step 4.34, the processing procedure of the DSA module includes: Perform pointwise convolution on the input feature map to compress the channel dimension to Local spatial information is obtained through DSC deep convolutional layers, and the dimension of the space is compressed to 1 through max pooling. The convolutional kernel restores the channel dimension to the number of channels in the input feature map, which is used to obtain the channel attention weights for each channel in the output feature map. Finally, the channel attention weights of each channel in the input feature map are compared. The product of these is used as the weighted output feature map; The input feature map is convolved to obtain a convolved feature map. Then, a transformation function is used to obtain channel attention weights, which are then multiplied element-wise with the input feature map to obtain a weighted channel attention feature map. Finally, a residual connection is used to concatenate the input feature map with the weighted channel attention feature map to obtain a new output feature map, represented as: 。 7. The method for identifying precursor information and providing short-term early warning of acoustic emission characteristics of coal-rock composite fractures according to claim 6, characterized in that, In step 4.34, channel attention weights The calculation formula is: In the formula, The input feature map; This is a pointwise convolution operation; yes Multiply by depthwise convolution; It is a space max pooling operation. Represent a The convolution is performed and includes a Sigmoid activation function.

8. The method for identifying precursor information and providing short-term early warning of acoustic emission characteristics of coal-rock composite fractures according to claim 4, characterized in that, Step 4.4 includes: Input time series feature sequence The TCN module will input the time series feature sequence. After processing through causal convolutional layers, the receptive field is expanded in each causal convolutional layer using kernel dilation, as represented by: in, For expansion rate, The kernel size is [size]. For convolution kernel weights, For time steps The convolution result; Causal convolutional layers use residual connections to process the input temporal feature sequences. The features are summed to obtain the output features. The time dimension of the output features is normalized by global average pooling to output a time-series dynamic feature vector.

9. The method for identifying precursor information and providing short-term early warning of acoustic emission characteristics of coal-rock complex fractures according to claim 4, characterized in that, Step 5 includes: Step 5.1: Probabilize the output risk of the rupture state identification model to obtain the rupture probability value; Step 5.2: Reprocess the fracture probability value to construct a dynamic threshold model and a multi-index weighted fusion model, and determine the state of the coal sample based on the output of the multi-index weighted fusion model.

10. The method for identifying precursor information and providing short-term early warning of acoustic emission characteristics of coal-rock complex fractures according to claim 9, characterized in that, Step 5.1 includes: Based on the coal sample fracture state identification model, in step 4.5, the output layer of the final fully connected layer is set to multiple neurons to correspond to the loading stage, elastoplastic stage, and pre-fracture stage; then, a Softmax activation function is set after the output layer of the fully connected layer, and the original score of each stage is converted into a probability distribution; the modified output vector is: In the formula Let the number of state categories be denoted by '[', ... As the probability of rupture Output, defined as: Step 5.2 includes: Step 5.21, construct the dynamic threshold model, which is expressed by the following formula: In the formula, , and These are the mean values ​​of the probability surge rate, energy surge rate, and ring count mutation point within the current statistical window, respectively. , and These are the standard deviations of the corresponding indicators; , and This is an empirical coefficient; This represents the probability increase rate. The rate of change of energy; This is a sudden change in the ring count; Step 5.22, construct a multi-indicator weighted fusion model, represented as: In the formula, , and These are the weighting coefficients; The threshold for joint decision-making; when the multi-indicator weighted fusion model... When the output exceeds the warning threshold, it is determined that the coal sample is at risk of rupture.

Images