Water-rich rock mass fracture precursor identification method based on lightweight neural network
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SANYA SCI & EDUCATION INNOVATION PARK WUHAN UNIV OF TECH
- Filing Date
- 2026-04-23
- Publication Date
- 2026-05-26
Smart Images
Figure CN122087731A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of safety monitoring technology for deep rock mass engineering. Background Technology
[0002] In the field of deep geotechnical engineering safety monitoring, acoustic emission technology, as a core means of characterizing the internal fracture dynamics of rock masses, directly affects the early warning effectiveness of major engineering disasters due to its monitoring accuracy and response speed. For water-rich rock masses (such as water-rich sandstone, rock strata with well-developed fractures and high water content), the internal fracture mechanism and signal propagation characteristics are more complex. The presence of water not only changes the mechanical properties of the rock mass but also causes multiple interferences to the generation, propagation, and reception of acoustic emission signals: on the one hand, fluctuations in pore water pressure may induce micro-fractures or generate fluid noise; on the other hand, the water medium causes significant attenuation and scattering of high-frequency signal components, further blurring the already weak characteristics of fracture precursor signals. Traditional acoustic emission analysis techniques for rock mass fractures mainly rely on energy statistical methods based on fixed thresholds or on manual interpretation of waveform characteristics based on expert experience. When applied to water-rich rock masses, the inherent limitations of such methods are further amplified: First, although threshold-based statistical models are simple to calculate, they are difficult to distinguish the nonlinear mixed patterns formed by the interweaving of rupture signals, fluid activity noise, and complex background noise in the time and frequency domain, leading to a simultaneous increase in the risk of missed and false alarms of instability precursors in water-rich rock masses; Second, although manual interpretation methods can utilize domain knowledge, they are highly subjective, inefficient, and unable to meet the real-time analysis needs of massive, high-noise, and feature-degraded monitoring data in water-rich environments, thus failing to meet the urgent needs of modern water-rich rock mass engineering (such as deep mines, submarine tunnels, and reservoir slopes) for automated, intelligent, and highly robust safety monitoring.
[0003] At its root, existing intelligent methods fail to fundamentally solve the four core challenges of acoustic emission monitoring in water-rich rock masses: accurate capture and enhancement of weak precursor signals under water-induced attenuation, extreme class imbalance and model bias under fluid noise interference, insufficient model generalization ability in complex water-rock coupled noise environments, and the contradiction between real-time requirements and model complexity in edge deployment scenarios. Therefore, a fracture precursor intelligent identification technology is needed that deeply integrates the physical mechanisms of acoustic emission in water-rich rock masses with the advantages of deep learning, possessing high precision, high robustness, strong anti-interference capabilities, and high engineering applicability, in order to achieve accurate early warning of instability disasters in water-rich rock masses.
[0004] The shortcomings of existing technology: 1. The problem of failing to perceive weak precursor signals under water-induced attenuation and interference: Traditional acoustic emission thresholding methods and general 1D-CNN models lack the ability to specifically perceive micro-fracture signals generated during the "bottleneck period" of water-rich rock mass fracturing. These signals are affected by the high-frequency filtering and scattering effects of pore water media, resulting in further energy attenuation, waveform distortion, and bandwidth broadening. They are easily submerged or even masked by strong environmental noise, equipment interference, and unique fluid noise (such as water flow sound and bubble bursting sound). This leads to an extremely high rate of missed detection of key precursor information of water-rich rock mass instability in existing systems, making it impossible to achieve effective early warning.
[0005] 2. Extreme Sample Imbalance and Model Recognition Bias Introduced by Fluid Noise: In acoustic emission data of water-rich rock masses, the number of effective micro- and macro-fracture event samples (positive samples) is much smaller than the massive background noise samples (negative samples). Furthermore, the noise contains a large number of non-fracture signals related to hydrological activities, leading to more severe class imbalance and feature confusion. Directly training the model on such datasets will cause its decision boundary to be severely biased towards the majority class (noise and fluid interference), resulting in "lazy" learning and even misclassifying some fluid noise as fracture events, thus seriously impairing the accuracy and reliability of identifying key fracture events.
[0006] 3. Insufficient generalization ability of models in complex water-rock coupled noise environments: In real engineering scenarios such as water-rich mines, submarine tunnels, and saturated slopes, acoustic emission signals contain not only mechanical vibrations and electromagnetic interference, but also a large amount of non-stationary, non-Gaussian hydrological noise caused by pore water pressure fluctuations, fluid seepage, and phase changes (vaporization / condensation). Traditional feature engineering methods and models trained under dry or simple noise backgrounds exhibit a sharp decline in generalization performance under such complex water-rock coupled noise environments. While they perform well on laboratory or ideal datasets, their false alarm and false negative rates increase significantly when deployed in water-rich fields, and the model reliability fails to meet engineering requirements.
[0007] 4. The contradiction between model computational complexity and the real-time and edge computing requirements of water-rich engineering: Safety monitoring in water-rich rock mass engineering often faces more pressing disaster early warning windows, requiring systems with real-time response capabilities at the minute or even second level. Simultaneously, monitoring terminals often need to be deployed on edge devices with extremely limited computing, storage, power consumption, and communication bandwidth, such as underground, underwater, or remote slopes. However, high-performance deep learning models typically have high computational complexity and a large number of parameters, making efficient deployment and real-time inference difficult in resource-constrained embedded environments, severely limiting their large-scale and practical application in water-rich rock mass engineering. Summary of the Invention
[0008] This invention aims to provide a method for identifying precursors of fracture in water-rich rock masses based on lightweight neural networks.
[0009] The technical solution of the present invention is as follows: A method for identifying precursors of fracture in water-rich rock masses based on lightweight neural networks includes the following steps: S1: Collect pulse waves and interference waves from multiple points in the deep rock mass for preprocessing, and combine the pulse waves and interference waves from each point into data pairs. Construct a multi-dimensional waveform library based on the data pairs. The interference waves include mechanical vibration interference waves, electromagnetic interference waves, and environmental interference waves. S2: Construct a lightweight model for predicting rock fracture to obtain the ternary probability distribution of background noise, micro-fracture precursors, and macro-fracture; the lightweight model for predicting rock fracture includes a multi-scale fusion module, a depthwise separable convolution module, a channel attention weighting module, and a global context classifier. S21: Input the pulse wave and interference wave that make up the data pair into the multi-scale fusion module to obtain the fused feature map; the multi-scale fusion module includes a multi-scale parallel convolutional layer and a channel dimension splicing layer; S22: Input the fused feature map into the depthwise separable convolutional module to obtain a temporal feature map; the depthwise separable convolutional module includes a depthwise convolutional layer and a pointwise convolutional layer; S23: Input the temporal feature map into the channel attention weighting module to obtain the weighted feature map; the channel attention weighting module includes a global average pooling layer and two fully connected layers; S24: Input the weighted enhanced feature map into the global context classifier to obtain the ternary probability distribution of background noise, micro-fracture precursors, and macro-fractures; the global context classifier includes a global pooling layer, a fully connected layer, a Dropout layer, and an output layer; the output layer is constructed with the Softmax function as its core. S3: Input the pulse wave and interference wave that make up the data pair into the lightweight model for predicting rock fracture to train and obtain the optimized lightweight model for predicting rock fracture. S4: Input the pulse wave and interference wave that make up the data pair into the optimized lightweight model for predicting rock fracture to obtain the ternary probability distribution of background noise, micro-fracture precursors and macro-fracture.
[0010] Preferably, in step S21, the pulse wave and interference wave constituting the data pair are input into the multi-scale fusion module to obtain the fused feature map. The specific process is as follows: The pulse wave and interference wave that make up the data pair are input into each branch of the multi-scale parallel convolutional layer to extract features at different scales. The features at different scales are then input into the channel dimension splicing layer to obtain a fused feature map.
[0011] The formula for the branch convolution is as follows:
[0012] in, The label indicating the feature map of the branch output; Indicates the feature channel number; Indicates the output feature map elements; Represents the elements of the weight tensor; Indicates the input signal element; Indicates the kernel size; Indicates the internal index of the convolution kernel; Indicates the bias term; This represents the activation function.
[0013] Preferably, in step S22, the fused feature map is input into the depthwise separable convolutional module to obtain a temporal feature map, and the specific process is as follows: The fused feature map is input into a deep convolutional layer to obtain local features, and the local features are then input into a pointwise convolutional layer for fusion to obtain a temporal feature map. The formula for the deep convolutional layer is as follows:
[0014] in, Represents the weights of the spatial filter; Indicates the input channel index; Indicates the kernel location index; Indicates the first The first input channel, the first The feature values output by the depthwise convolution at each time step; This represents the kernel size of a depthwise convolution; Indicates the first The first input channel, the first Input feature values at each time step; The formula for the pointwise convolutional layer is as follows:
[0015] in, Indicates the first The output channel, the first Pointwise convolution at each time step outputs feature values; Indicates the total number of input channels; Indicates the input channel index; Indicates the first The output channel, the first Pointwise convolution weights for each input channel; Indicates the first The bias term for each output channel.
[0016] Preferably, in step S23, the temporal feature map is input into the channel attention weighting module to obtain the weighted feature map, and the specific process is as follows: The temporal feature map is input into a global average pooling layer for compression to obtain a channel statistical vector. The channel statistical vector is then input into two fully connected layers to transform it into a non-linear channel correlation vector. The non-linear channel correlation vector is normalized to obtain a weight vector. The temporal feature map and the weight vector are multiplied step by step to obtain a weight-enhanced feature map.
[0017] Preferably, in step S24, the weighted enhanced feature map is input into the global context classifier to obtain the ternary probability distribution of background noise, micro-fracture precursors, and macro-fractures. The specific process is as follows: The weighted feature map is input into the global pooling layer for compression to obtain the pooling vector. The pooling vector is then input into the fully connected layer for nonlinear feature transformation to obtain high-level classification features. The high-level classification features are input into the Dropout layer for numerical regularization. Finally, the ternary probability distribution of background noise, micro-fracture precursors, and macro-fractures is obtained through the Softmax function in the output layer.
[0018] The high-level classification features are input into the Dropout layer for numerical regularization, and then the ternary probability distribution of background noise, micro-fracture precursors, and macro-fractures is obtained through the Softmax function in the output layer. The expression is as follows:
[0019] in, This represents the advanced classification features after nonlinear transformation and regularization. This represents a regularization technique to prevent overfitting. Represents the weight matrix; This represents the input feature vector after global pooling; This represents the bias vector of the fully connected layer; This represents the result after the linear transformation of the output layer. This represents the output weight matrix; This represents the bias vector of the output layer; This represents the final classification probability vector; This represents the activation function.
[0020] Compared with the prior art, the present invention has the following beneficial effects: 1. Breakthrough in the accuracy and sensitivity of weak fracture signal identification in water-rich environments: Based on the synergistic optimization of a multi-scale parallel sensing architecture and a channel-time dual attention mechanism, this invention achieves synchronous capture and adaptive enhancement of multi-resolution waveform features of water-attenuated acoustic emission signals, effectively separating fracture signals from fluid noise. This significantly improves the recall rate of precursor signals of micro-fractures in water-rich rock masses compared to traditional threshold discrimination methods and single convolutional neural network models, overcoming the industry challenge of high false negative rates for weak fracture signals in the context of strong water-rock coupling noise.
[0021] 2. Significant Upgrade in Early Warning and Model Generalization Capabilities Against Fluid Interference: By introducing a Focal Loss function optimized for water-rich environments and incorporating data augmentation strategies guided by hydrophysical mechanisms (such as simulating fluid noise superposition and water-induced signal attenuation), the system alleviates the severe class imbalance caused by the scarcity of precursor samples and hydrological noise interference. This endows the model with excellent generalization performance and robustness under different hydrological conditions, noise patterns, and full-energy spectrum fracture events. The model's accurate probabilistic distinction between background noise, fluid interference, micro-fracture precursors, and macro-fractures provides a more reliable quantitative decision-making basis for risk classification and management of water-rich rock mass engineering, promoting a fundamental paradigm shift in monitoring technology from "post-fracture identification" to "pre-fracture early warning."
[0022] 3. Significantly Enhanced Monitoring Efficiency and Lightweight Deployment Applicability in Water-Rich Engineering Scenarios: The end-to-end intelligent identification process replaces the traditional manual analysis of acoustic emission spectra, which relies on expert experience and is particularly difficult to handle complex water-rich signals. This reduces the analysis cycle from days to minutes, greatly improving monitoring timeliness. Simultaneously, the depthwise separable convolution and lightweight structural design, while ensuring prediction accuracy, reduce the model parameter scale and computational complexity by orders of magnitude. This provides key technical support for embedded real-time monitoring on resource-constrained edge computing devices (such as downhole microcontrollers and unmanned monitoring nodes), precisely meeting the urgent needs of water-rich rock mass engineering for intelligent, real-time, and highly reliable safety monitoring. Attached Figure Description
[0023] Figure 1 This is a flowchart of a method for identifying precursors of fracturing in water-rich rock masses based on a lightweight neural network, according to the present invention.
[0024] Figure 2 This is a logical schematic diagram of a method for identifying precursors of fracturing in water-rich rock masses based on a lightweight neural network, according to the present invention.
[0025] Figure 3 This is a structural diagram of a method for identifying precursors of fracturing in water-rich rock masses based on a lightweight neural network, according to the present invention. Detailed Implementation
[0026] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments and accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0027] like Figure 1As shown, the present invention proposes a method for identifying precursors of fracturing in water-rich rock masses based on lightweight neural networks, comprising the following steps: S1: Collect pulse waves and interference waves from multiple points in the deep rock mass for preprocessing, and combine the pulse waves and interference waves from each point into data pairs. Construct a multi-dimensional waveform library based on the data pairs. The interference waves include mechanical vibration interference waves, electromagnetic interference waves, and environmental interference waves. The first step is to configure the sensor, selecting a wideband piezoelectric sensor such as... Figure 2 As shown, to match the high-frequency characteristics of deep rock fractures and avoid pulse feature loss due to insufficient bandwidth of low-frequency sensors, a grid-based deployment is adopted to ensure that the signal is simultaneously captured by at least three sensors, meeting positioning requirements. The sensors are tightly bonded to the rock mass using coupling agent and mechanically clamped to suppress signal attenuation and vibration noise. The multi-channel acquisition system is synchronized via GPS or trigger lines, with a timing error of <1μs, ensuring the accuracy of arrival time difference measurements. Next, the signal acquisition settings include a sampling rate set according to the Nyquist theorem, not less than twice the highest signal frequency. The single sampling length is 1024–4096 points, covering typical fracture durations with redundancy. An amplitude triggering mechanism is used, with a threshold of three times the standard deviation of background noise and 50 pre-trigger points to ensure complete capture of the pulse initiation segment. Next, signal labeling and classification are performed. Effective fracture signals are identified based on the following characteristics: Time domain: sudden pulse, rise time <10μs, rapid attenuation; Frequency domain: high frequency dominant (>100kHz), concentrated spectral peaks; Energy: ring count >20, energy significantly higher than background. Interference signals are categorized into mechanical vibration, electromagnetic interference, and environmental interference. Each signal is independently labeled by at least two experts, and discrepancies are reviewed to form a unified label file. Finally, data standardization and enhancement are performed by Z-score standardization of the samples, as shown below:
[0028] in, and These represent the mean and standard deviation of the training set, respectively. Represents the standardized eigenvalues; To address the issue of scarce samples, the following enhancement strategies are implemented: basic enhancement: time shifting, amplitude scaling; and advanced enhancement: superimposing real noise, tail truncation, and phase reversal. S2: Construct a lightweight model for predicting rock fracture to obtain the ternary probability distribution of background noise, micro-fracture precursors, and macro-fracture; the lightweight model for predicting rock fracture includes a multi-scale fusion module, a depthwise separable convolution module, a channel attention weighting module, and a global context classifier. S21: Input the pulse wave and interference wave that make up the data pair into the multi-scale fusion module to obtain the fused feature map; the multi-scale fusion module includes a multi-scale parallel convolutional layer and a channel dimension splicing layer; S22: Input the fused feature map into the depthwise separable convolutional module to obtain a temporal feature map; the depthwise separable convolutional module includes a depthwise convolutional layer and a pointwise convolutional layer; S23: Input the temporal feature map into the channel attention weighting module to obtain the weighted feature map; the channel attention weighting module includes a global average pooling layer and two fully connected layers; S24: Input the weighted enhanced feature map into the global context classifier to obtain the ternary probability distribution of background noise, micro-fracture precursors, and macro-fractures; the global context classifier includes a global pooling layer, a fully connected layer, a Dropout layer, and an output layer; the output layer is constructed with the Softmax function as its core. This step designs the architecture of a lightweight 1D-CNN model for predicting rock fracture based on acoustic emission signals, comprising four parts, such as... Figure 3 As shown, the first module is a multi-scale fusion module. By introducing multi-scale parallel convolutional layers, this invention significantly improves the model's initial perception capability of heterogeneous waveform features in acoustic emission signals, effectively solving the problem of feature scale uniformity in the shallow layers of traditional single-path convolution. The second module is a depthwise separable convolution module. This module first uses deep convolutional layers to perform independent spatial filtering on each input channel to focus on learning the local correlation of acoustic emission signals in the temporal dimension; then, it performs linear combinations between channels through pointwise convolution to construct new feature representations. This "decomposition-reconstruction" design concept effectively avoids overfitting risks while maintaining the model's feature extraction capabilities, enhances the model's generalization ability under limited sample conditions, and provides a necessary lightweight foundation for embedded deployment in engineering settings. Next is the channel attention weighting module. This module first performs global average pooling on the input feature map to compress global spatial information and generate channel statistical description vectors. Then, through two fully connected layers containing dimensionality reduction and restoration, it learns the non-linear dependencies between channels and outputs normalized channel attention weights. The output feature values are then constrained by the Sigmoid activation function. Finally, channel-wise weighting is used to enhance key features and suppress redundant features. Finally, a global context classifier is used, and the feature vector is then non-linearly transformed through fully connected layers and regularized using a Dropout layer to suppress overfitting. Finally, the output layer uses the Softmax function to map high-level features into a ternary probability distribution of background noise, micro-fracture precursors, and macro-fractures, thus completing end-to-end rock mass fracture identification.
[0029] This prediction model consists of four core processing layers: To address the issue of traditional single-path convolution's limited feature perception in shallow network layers, a multi-scale parallel input layer is used to enhance the perception of heterogeneous waveform features. The formula for the branch convolution is as follows:
[0030] in, The label indicating the feature map of the branch output; Indicates the feature channel number; Indicates the output feature map elements; Represents the elements of the weight tensor; Indicates the input signal element; Indicates the kernel size; Indicates the internal index of the convolution kernel; Indicates the bias term; This represents the activation function.
[0031] This module performs spatial filtering through depthwise convolution, followed by pointwise convolution to achieve channel fusion. This "decomposition-reconstruction" design significantly improves the model's generalization ability and deployment convenience while maintaining performance. The formula for the depthwise convolutional layer is as follows:
[0032] in, Represents the weights of the spatial filter; Indicates the input channel index; Indicates the kernel location index; Indicates the first The first input channel, the first The feature values output by the depthwise convolution at each time step; This represents the kernel size of a depthwise convolution; Indicates the first The first input channel, the first Input feature values at each time step; The formula for the pointwise convolutional layer is as follows:
[0033] in, Indicates the first The output channel, the first Pointwise convolution at each time step outputs feature values; Indicates the total number of input channels; Indicates the input channel index; Indicates the first The output channel, the first Pointwise convolution weights for each input channel; Indicates the first The bias term for each output channel.
[0034] This layer performs global average pooling on the input feature map to obtain channel statistics, and then generates normalized attention weights through a fully connected layer with a bottleneck structure and a sigmoid function. Finally, key feature enhancement and redundant feature suppression are achieved through channel-wise weighting.
[0035] Finally, the global context classifier performs nonlinear transformation and regularization on the features through fully connected layers and Dropout layers, and uses the Softmax function to output the ternary probability distribution of background noise, micro-fracture precursors and macro-fractures, thus completing the end-to-end identification of rock mass fracture status.
[0036]
[0037] in, This represents the advanced classification features after nonlinear transformation and regularization. This represents a regularization technique to prevent overfitting. Represents the weight matrix; This represents the input feature vector after global pooling; This represents the bias vector of the fully connected layer; This represents the result after the linear transformation of the output layer. This represents the output weight matrix; This represents the bias vector of the output layer; This represents the final classification probability vector; Indicates the activation function; S3: Input the pulse wave and interference wave that make up the data pair into the lightweight model for predicting rock fracture to train and obtain the optimized lightweight model for predicting rock fracture. This step aims to overcome a key bottleneck in model training—the severe imbalance in class distribution caused by the extreme scarcity of precursor samples for rock mass fracturing. In the label preparation stage, a standardized labeling system is established through expert annotation. To address sample imbalance, a composite objective function integrating Focal Loss and mean squared error regression loss is designed. Adjustable focusing parameters suppress the gradient contribution of simple samples, and energy regression is used to guide the model in learning the physical nature of the fracturing process. At the data level, a hybrid enhancement pipeline based on physical mechanisms is constructed. Three types of operators—elastic time warp, pulse superposition, and mode mixing—are used to simulate wave velocity inhomogeneity, multi-source fracturing superposition, and composite fracturing modes, respectively, systematically improving the model's adaptability to complex acoustic scenarios. At the structural optimization level, stochastic depth regularization and channel-time dual attention mechanisms are introduced. Through random path discarding and feature dimension recalibration, the overfitting tendency of the model to specific noise patterns is effectively curbed, comprehensively improving the discriminative and generalization performance of feature representation.
[0038] Based on the independent interpretation and cross-validation results of acoustic emission signal waveforms by no fewer than two experts in the field, the samples were divided into three categories: background noise (0), micro-fracture precursors (1), and macro-fractures (2).
[0039] The loss function is designed as a weighted binary cross-entropy function, as follows:
[0040] in, Represents the category index (0: background noise, 1: micro-fracture precursor, 2: macro-fracture). Indicates the true label of the sample; Indicates the model's predicted probability; Indicates the category weight coefficient; Indicates the focus parameter; This represents the loss function.
[0041] The channel-time dual attention mechanism to prevent overfitting is formulated as follows:
[0042] in, This represents the channel statistics vector output by the extrusion step; Represents the dimensionality reduction weight matrix; Represents the dimension reduction bias vector; Represents the ReLU activation function; Represents the weight matrix for increasing dimensionality; Represents the up-dimensional bias vector; This represents the Sigmoid activation function; Indicates channel attention weights; Indicates time attention weights; Represents a 1-dimensional transformation function; Indicates a splicing operation; This represents the feature obtained by global average pooling over the time dimension. This represents the feature obtained by global max pooling in the time dimension.
[0043] The formula for data augmentation guided by physical mechanisms is as follows:
[0044] in, This represents two different input samples randomly selected from the same training batch; Represents the mixing coefficient; This represents the new sample generated after mixing; This indicates an enhanced signal after time warping; Represents the time warp function; Represents the original timing signal; This represents the new sample after the shift enhancement; Indicates a time shift operation; Indicates a physical disturbance signal; Indicates the time offset.
[0045] S4: Input the pulse wave and interference wave that make up the data pair into the optimized lightweight model for predicting rock fracture to obtain the ternary probability distribution of background noise, micro-fracture precursors and macro-fracture.
[0046] Although preferred embodiments of this application have been described, those skilled in the art, once they learn the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application. Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if these modifications and variations of this application fall within the scope of the claims of this application and their equivalents, this application also intends to include these modifications and variations.
Claims
1. A method for identifying precursors of fracture in water-rich rock masses based on lightweight neural networks, characterized in that, include: S1: Collect pulse waves and interference waves from multiple points in the deep rock mass for preprocessing, and combine the pulse waves and interference waves from each point into data pairs. Construct a multi-dimensional waveform library based on the data pairs. The interference waves include mechanical vibration interference waves, electromagnetic interference waves, and environmental interference waves. S2: Construct a lightweight model for predicting rock fracture to obtain the ternary probability distribution of background noise, micro-fracture precursors, and macro-fracture; the lightweight model for predicting rock fracture includes a multi-scale fusion module, a depthwise separable convolution module, a channel attention weighting module, and a global context classifier. S21: Input the pulse wave and interference wave that make up the data pair into the multi-scale fusion module to obtain the fused feature map; the multi-scale fusion module includes a multi-scale parallel convolutional layer and a channel dimension splicing layer; S22: Input the fused feature map into the depthwise separable convolutional module to obtain a temporal feature map; the depthwise separable convolutional module includes a depthwise convolutional layer and a pointwise convolutional layer; S23: Input the temporal feature map into the channel attention weighting module to obtain the weighted feature map; the channel attention weighting module includes a global average pooling layer and two fully connected layers; S24: Input the weighted enhanced feature map into the global context classifier to obtain the ternary probability distribution of background noise, micro-fracture precursors, and macro-fractures; the global context classifier includes a global pooling layer, a fully connected layer, a Dropout layer, and an output layer; the output layer is constructed with the Softmax function as its core. S3: Input the pulse wave and interference wave that make up the data pair into the lightweight model for predicting rock fracture to train and obtain the optimized lightweight model for predicting rock fracture. S4: Input the pulse wave and interference wave that make up the data pair into the optimized lightweight model for predicting rock fracture to obtain the ternary probability distribution of background noise, micro-fracture precursors and macro-fracture.
2. The method for identifying precursors of fracturing in water-rich rock masses based on lightweight neural networks according to claim 1, characterized in that, In step S21, the pulse wave and interference wave that make up the data pair are input into the multi-scale fusion module to obtain the fused feature map. The specific process is as follows: The pulse wave and interference wave that make up the data pair are input into each branch of the multi-scale parallel convolutional layer to extract features at different scales. The features at different scales are then input into the channel dimension splicing layer to obtain a fused feature map.
3. The method for identifying precursors of fracturing in water-rich rock masses based on lightweight neural networks according to claim 2, characterized in that, The formula for the branch convolution is as follows: ， in, The label indicating the feature map of the branch output; Indicates the feature channel number; Indicates the output feature map elements; Represents the elements of the weight tensor; Indicates the input signal element; Indicates the kernel size; Indicates the internal index of the convolution kernel; Indicates the bias term; This represents the activation function.
4. The method for identifying precursors of fracturing in water-rich rock masses based on lightweight neural networks according to claim 1, characterized in that, In step S22, the fused feature map is input into the depthwise separable convolutional module to obtain the temporal feature map. The specific process is as follows: The fused feature map is input into a deep convolutional layer to obtain local features, and the local features are then input into a pointwise convolutional layer for fusion to obtain a temporal feature map. The formula for the deep convolutional layer is as follows: ， in, Represents the weights of the spatial filter; Indicates the input channel index; Indicates the kernel location index; Indicates the first The first input channel, the first The feature values output by the depthwise convolution at each time step; This represents the kernel size of a depthwise convolution; Indicates the first The first input channel, the first Input feature values at each time step; The formula for the pointwise convolutional layer is as follows: ， in, Indicates the first The output channel, the first Pointwise convolution at each time step outputs feature values; Indicates the total number of input channels; Indicates the input channel index; Indicates the first The output channel, the first Pointwise convolution weights for each input channel; Indicates the first The bias term for each output channel.
5. The method for identifying precursors of fracturing in water-rich rock masses based on lightweight neural networks according to claim 1, characterized in that, In step S23, the temporal feature map is input into the channel attention weighting module to obtain the weighted enhanced feature map. The specific process is as follows: The temporal feature map is input into a global average pooling layer for compression to obtain a channel statistical vector. The channel statistical vector is then input into two fully connected layers to transform it into a non-linear channel correlation vector. The non-linear channel correlation vector is normalized to obtain a weight vector. The temporal feature map and the weight vector are multiplied step by step to obtain a weight-enhanced feature map.
6. The method for identifying precursors of fracturing in water-rich rock masses based on lightweight neural networks according to claim 1, characterized in that, In step S24, the weighted enhanced feature map is input into the global context classifier to obtain the ternary probability distribution of background noise, micro-fracture precursors, and macro-fractures. The specific process is as follows: The weighted feature map is input into the global pooling layer for compression to obtain the pooling vector. The pooling vector is then input into the fully connected layer for nonlinear feature transformation to obtain high-level classification features. The high-level classification features are input into the Dropout layer for numerical regularization. Finally, the ternary probability distribution of background noise, micro-fracture precursors, and macro-fractures is obtained through the Softmax function in the output layer.
7. The method for identifying precursors of fracturing in water-rich rock masses based on lightweight neural networks according to claim 6, characterized in that, The high-level classification features are input into the Dropout layer for numerical regularization, and then the ternary probability distribution of background noise, micro-fracture precursors, and macro-fractures is obtained through the Softmax function in the output layer. The expression is as follows: ， ， ， in, This represents the advanced classification features after nonlinear transformation and regularization. This represents a regularization technique to prevent overfitting. Represents the weight matrix; This represents the input feature vector after global pooling; This represents the bias vector of the fully connected layer; This represents the result after the linear transformation of the output layer. This represents the output weight matrix; This represents the bias vector of the output layer; This represents the final classification probability vector; This represents the activation function.