A lightweight model construction method for microseismic signal classification

Abstract

The application provides a lightweight model construction method for microseismic signal classification, belongs to the technical field of microseismic signal analysis, and acquires microfracture signals collected by a microseismic monitoring system and carries out pretreatment to obtain a microseismic signal classification data set; an initial lightweight microseismic signal classification model is constructed, including an input layer, a lightweight feature extraction backbone network and a classifier connected in sequence, the lightweight feature extraction backbone network includes a down-sampling unit, three feature extraction units alternately stacked by enhanced shuffle network unit modules and double-attention adaptive residual shrinkage modules and a hollow convolution pyramid pooling unit connected in sequence; the initial lightweight microseismic signal classification model is subjected to model training iteration by using a training set until convergence; the trained lightweight microseismic signal classification model is verified by using a verification set, model parameters are adjusted and retrained; the lightweight microseismic signal classification model is tested by using a test set, and evaluation indexes are counted and evaluated.

Description

Technical Field

[0001] This invention relates to the field of microseismic signal analysis technology, and in particular to a lightweight model construction method for microseismic signal classification. Background Technology

[0002] Microseismic monitoring technology, as a three-dimensional, real-time method for monitoring rock mass stability, has been widely used in early warning of disasters in deep underground engineering. However, in practical engineering applications, traditional microseismic signal processing methods mainly rely on manual analysis, which is inefficient and highly dependent on the judgment of experienced engineers, making it difficult to meet the growing demands for real-time and automation. Therefore, developing methods that can automatically identify and classify microseismic signals has become a current research focus.

[0003] In recent years, significant progress has been made in machine learning-based microseismic signal classification methods. Traditional machine learning methods typically require complex feature engineering, heavily relying on the experience of domain experts and advanced signal processing techniques. However, manually extracted features often fail to fully reflect the essential characteristics of the original signal. In contrast, deep learning models can achieve end-to-end processing from feature extraction to signal classification, among which convolutional neural networks have attracted widespread attention due to their powerful feature representation capabilities.

[0004] While the above methods perform well in laboratory environments or on specific datasets, they still face multiple challenges in real-world engineering deployments:

[0005] Lightweight Model Requirements: Due to the harsh environments often present at deep underground engineering sites, such as high temperatures and weak network coverage, and the need to process a large amount of microseismic data in real time, strict requirements are placed on the computational efficiency of the model. As the number of layers and channels increases in traditional deep convolutional neural networks, the amount of floating-point operations also increases, which seriously affects the running efficiency of the model.

[0006] Insufficient noise robustness: There are various types of interference noise at the construction site (such as blasting vibration, mechanical vibration, electromagnetic interference, etc.), resulting in a low signal-to-noise ratio of the collected microseismic signals, which seriously affects the recognition accuracy of the model. Most existing models have not been optimized for the noise patterns unique to microseismic signals and lack effective denoising mechanisms. Summary of the Invention

[0007] To address the problems existing in the prior art, this invention provides a lightweight model construction method for microseismic signal classification. While ensuring the accuracy of microseismic signal classification, it solves the problems of high model complexity, large computational resource consumption, and severe noise interference in the prior art.

[0008] This invention provides a lightweight model construction method for microseismic signal classification, specifically including the following steps:

[0009] The micro-fracture signals collected by the microseismic monitoring system are acquired and preprocessed to obtain a microseismic signal classification dataset. The microseismic signal classification dataset is then divided into a training set, a validation set, and a test set according to a preset ratio.

[0010] An initial lightweight microseismic signal classification model was constructed, and the hyperparameters for training the lightweight microseismic signal classification model were set. The initial lightweight model consists of an input layer, a lightweight feature extraction backbone network, and a classifier connected in sequence. The lightweight feature extraction backbone network consists of a downsampling unit, three feature extraction units consisting of an enhanced shuffling network unit module and a dual attention adaptive residual shrinking module stacked alternately, and a dilated convolution pyramid pooling unit connected in sequence.

[0011] The initial lightweight microseismic signal classification model was trained iteratively based on the microseismic signal classification dataset until convergence.

[0012] The trained lightweight microseismic signal classification model was validated using a validation set. The model parameters were adjusted and retrained to determine the final lightweight microseismic signal classification model.

[0013] The final lightweight microseismic signal classification model was tested using a test set, and evaluation metrics were statistically analyzed.

[0014] The lightweight feature extraction backbone network for constructing the initial lightweight microseismic signal classification model includes the following steps:

[0015] Construct downsampling units: connect a 3×1 max pooling layer with a stride of 2 and a 3×1 convolutional layer with a stride of 2 in sequence.

[0016] Feature extraction units are constructed: a dual-attention adaptive residual shrinkage module and an enhanced shuffle network unit module are constructed. One dual-attention adaptive residual shrinkage module and n enhanced shuffle network unit modules are connected sequentially. The input and output of the dual-attention adaptive residual shrinkage module are added together as the input of the first enhanced shuffle network unit module. The dual-attention adaptive residual shrinkage module is an adaptive residual shrinkage mechanism based on deep learning, which is constructed by combining channel attention and spatial attention for noise suppression and feature enhancement. The enhanced shuffle network unit module is an improved ShuffleNet unit for lightweight feature extraction.

[0017] Construct a dilated convolutional pyramid pooling unit: Construct five branches. The first branch consists of an adaptive average pooling layer, a 1×1 convolutional layer, and an upsampling layer connected in sequence. The second branch is a 3×1 convolutional layer with a dilation rate of 18. The third branch is a 3×1 convolutional layer with a dilation rate of 12. The fourth branch is a 3×1 convolutional layer with a dilation rate of 6. The fifth branch is a 1×1 convolutional layer.

[0018] The outputs of the five branches of the dilated convolutional pyramid pooling unit are concatenated and then connected to a dual-attention adaptive residual shrinking module. The input and output of the dual-attention adaptive residual shrinking module are added together and then connected to an adaptive average pooling layer. The output of the adaptive average pooling layer is connected to the input of the classifier.

[0019] The initial lightweight microseismic signal classification model is trained iteratively based on the training set until convergence. The specific steps include the following:

[0020] The input layer of the lightweight microseismic signal classification model receives vibration signals from the microseismic signal classification dataset.

[0021] The input signal undergoes preliminary feature extraction and downsampling through a 3×1 max pooling layer with a stride of 2 and a 3×1 convolutional layer with a stride of 2.

[0022] The downsampled signal features are further abstracted by a feature extraction unit consisting of an enhanced shuffling network unit module and a dual-attention adaptive residual shrinking module, while simultaneously achieving adaptive noise reduction.

[0023] The dilated convolutional pyramid pooling unit further extracts multi-scale features.

[0024] After concatenating the multi-scale features, the input is fed into a dual-attention adaptive residual shrinking module. The input and output of the dual-attention adaptive residual shrinking module are added together and then fed into an adaptive average pooling layer to obtain high-dimensional features.

[0025] The classifier receives high-dimensional features and provides classification results for microseismic signals. It then compares the predicted values ​​of the classified microseismic signals with the actual values, calculates the error value, and iterates through the training process.

[0026] Optionally, a dual-attention adaptive residual shrinkage module is constructed, which specifically includes the following steps:

[0027] Constructing a channel attention mechanism: Performing global max pooling on the input features to obtain a vector. And global average pooling to obtain vectors ; to vector and Input a small multilayer perceptron and sum the output features to obtain the channel weight vector. In this small multilayer perceptron, the first layer is a fully connected layer, ReLU, and Dropout; the second layer is a fully connected layer and a Hard-Sigmoid activation function; the weights are normalized using the Sigmoid function and then compared with the global maximum value M. max Multiply to obtain the channel-level threshold. .

[0028] Constructing a spatial attention mechanism: Perform global max pooling on the input features and compress them along the channel dimension to obtain a vector. And global average pooling to obtain vectors ; to vector and After concatenation, input into a 3×1 convolutional layer to obtain the spatial weight vector. Normalize the weights using the Sigmoid function and compare them with the global average A. avg Multiplying yields the spatial threshold. .

[0029] The nonlinear soft threshold function is constructed as follows:

[0030] .

[0031] Where x is the feature before denoising, y is the feature after denoising, and τ is the channel-level threshold or spatial-level threshold.

[0032] Residual connections are introduced: Input features are preprocessed sequentially through a 3×1 depthwise separable convolutional layer and a 1×1 convolutional layer; the preprocessed features are then processed by a channel attention mechanism, and a nonlinear soft thresholding function is used to perform a first transformation between the channel-level threshold and the preprocessed features; the features after the first transformation are then processed by a spatial attention mechanism, and a nonlinear soft thresholding function is used to perform a second transformation between the spatial-level threshold and the features after the first transformation; the input features and the features after the second transformation are added element-wise to obtain the output of the dual-attention adaptive residual shrinkage module.

[0033] Optionally, an enhanced shuffling network unit module is constructed, specifically including the following steps:

[0034] Channel grouping: grouping input features The channels are divided into G groups, and each group contains C / G channels.

[0035] Intra-group convolution: Channel expansion is performed using 1×1 convolution, increasing the number of channels in each group from C / G to K*C / G, where K is the expansion factor. Feature extraction is then performed using 3×1 depthwise separable convolution, ultimately outputting the features. .

[0036] Squeeze operation: on features Perform global max pooling and global average pooling to obtain two vectors. .

[0037] Excitation operation: Excitation vector and After concatenation, the input is given to a small multilayer perceptron, and the output features are summed to obtain the weight vector. The small multilayer perceptron includes two linear layers: the first layer is a fully connected layer with ReLU and Dropout, and the second layer is a fully connected layer with a Hard-Sigmoid activation function.

[0038] Scale operation: Scales the weight vector With features Multiply to obtain the weighted features .

[0039] Introducing residual connections: features The result after 1×1 convolution and the input features splicing, or, features The result after 1×1 convolution and the input features The results are stitched together after passing through a 3×1 depth separable convolutional layer and a 1×1 convolutional layer in sequence.

[0040] Channel shuffling: feature map Rearrange the channels into H×1×G×(KC / G); transpose the channel dimensions to interleave channels from different groups; restore the format to H×1×KC to obtain the output of the enhanced shuffle network unit module.

[0041] By adopting the above technical solution, the present invention has at least the following beneficial effects:

[0042] 1) Lightweight design: The model significantly reduces the number of model parameters (only about 0.155M) and floating-point operation volume (about 38.237M) through the enhanced shuffling network unit module, making it suitable for deployment on resource-constrained edge devices.

[0043] 2) Strong noise robustness: The dual attention adaptive residual shrinkage module in the model combines adaptive threshold and dual attention mechanism, which can maintain a high recognition accuracy under low signal-to-noise ratio conditions.

[0044] 3) Adaptive denoising capability: The model can automatically adjust the denoising intensity according to the degree of contamination of the input signal without manual intervention.

[0045] 4) High classification accuracy: It achieves an accuracy rate of 97.7% in classification experiments, far exceeding the mainstream lightweight models. Attached Figure Description

[0046] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0047] Figure 1 This is a flowchart illustrating a lightweight model construction method for microseismic signal classification provided by the present invention.

[0048] Figure 2 This is a schematic diagram of the architecture of a lightweight model for microseismic signal classification provided by the present invention.

[0049] Figure 3 This is a schematic diagram of the architecture of the dual-attention adaptive residual shrinkage module provided by the present invention.

[0050] Figure 4 The following are schematic diagrams of the architecture of the enhanced shuffling network unit module provided by the present invention: (a) is a schematic diagram of the architecture of the enhanced shuffling network unit module with a step size of 1, and (b) is a schematic diagram of the architecture of the enhanced shuffling network unit module with a step size of 2.

[0051] Figure 5 This is the confusion matrix of the microseismic signal classification results.

[0052] Figure 6 This is a comparison chart of the model complexity and accuracy of eight different models. Detailed Implementation

[0053] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the 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.

[0054] like Figure 1 As shown, this embodiment of the invention provides a lightweight model construction method for microseismic signal classification, including the following steps:

[0055] The micro-fracture signals collected by the microseismic monitoring system are acquired and preprocessed to obtain a microseismic signal classification dataset. The microseismic signal classification dataset is then divided into a training set, a validation set, and a test set according to a preset ratio.

[0056] In this embodiment, a complete microseismic monitoring system is deployed in the tunnel engineering project. It includes multiple high-sensitivity acceleration sensors. The arrangement of the sensors is determined according to the on-site construction conditions, but should meet the relevant deployment specifications in order to accurately capture microseismic events caused by rock mass fracture. The sampling period lasts for several months, and all events monitored during the monitoring period are recorded and saved.

[0057] The microseismic monitoring system continuously collects waveform data and stores the raw events on a local server. Subsequently, technicians perform preliminary screening of the collected data, eliminating obvious abnormal events (such as equipment failure, human interference, etc.) and classifying them into three categories based on signal characteristics: blasting events, microseismic events, and noise events.

[0058] All collected events were further classified channel by channel and saved in txt format in the corresponding folders (microseismic, blasting, and noise folders). Each channel signal was represented by one txt file to establish a microseismic signal classification dataset. Each channel signal had 4000 sampling points. Normalization and data augmentation were performed on each channel signal to ensure that the number of signals in each class remained consistent and to avoid class imbalance. All signals were divided into training set, validation set, and test set according to a preset ratio.

[0059] like Figure 2 As shown, an initial lightweight microseismic signal classification model is constructed, and the hyperparameters for training the lightweight microseismic signal classification model are set. The initial lightweight model includes an input layer, a lightweight feature extraction backbone network, and a classifier connected in sequence. The lightweight feature extraction backbone network includes a downsampling unit, three feature extraction units consisting of alternating stacks of an enhanced shuffling network unit module and a dual-attention adaptive residual shrinking module, and a dilated convolutional pyramid pooling unit connected in sequence. Specifically:

[0060] (1) Constructing downsampling units

[0061] Connect the 3×1 max pooling layer with a stride of 2 and the 3×1 convolutional layer with a stride of 2 in sequence.

[0062] (2) Constructing feature extraction units

[0063] like Figure 3 As shown, a dual-attention adaptive residual shrinkage module is constructed. This module is a deep learning-based adaptive residual shrinkage mechanism that combines channel attention and spatial attention. It is used for noise suppression and feature enhancement, and specifically includes the following steps:

[0064] Constructing a channel attention mechanism: Performing global max pooling on the input features to obtain a vector. And global average pooling to obtain vectors ; to vector and Input a small multilayer perceptron and sum the output features to obtain the channel weight vector. In this small multilayer perceptron, the first layer is a fully connected layer, ReLU, and Dropout; the second layer is a fully connected layer and a Hard-Sigmoid activation function; the weights are normalized using the Sigmoid function and then compared with the global maximum value M.max Multiply to obtain the channel-level threshold. .

[0065] Constructing a spatial attention mechanism: Perform global max pooling on the input features and compress them along the channel dimension to obtain a vector. And global average pooling to obtain vectors ;; will vector and After concatenation, input into a 3×1 convolutional layer to obtain the spatial weight vector. Normalize the weights using the Sigmoid function and compare them with the global average A. avg Multiplying yields the spatial threshold. .

[0066] The nonlinear soft threshold function is constructed as follows:

[0067] .

[0068] Where x is the feature before denoising, y is the feature after denoising, and τ is the channel-level threshold or spatial-level threshold.

[0069] Residual connections are introduced: Input features are preprocessed sequentially through a 3×1 depthwise separable convolutional layer and a 1×1 convolutional layer; the preprocessed features are then processed by a channel attention mechanism, and a nonlinear soft thresholding function is used to perform a first transformation between the channel-level threshold and the preprocessed features; the features after the first transformation are then processed by a spatial attention mechanism, and a nonlinear soft thresholding function is used to perform a second transformation between the spatial-level threshold and the features after the first transformation; the input features and the features after the second transformation are added element-wise to obtain the output of the dual-attention adaptive residual shrinkage module.

[0070] The dual-attention adaptive residual shrinkage module, as a key component of the classification model, achieves efficient noise suppression and feature enhancement through the following innovations:

[0071] 1) Depthwise separable convolution: significantly reduces computational cost and parameter count.

[0072] 2) Dual attention mechanism: Channel attention focuses on important feature channels, while spatial attention focuses on key time segments.

[0073] 3) Nonlinear soft threshold function: The improved soft threshold function avoids the defects of hard / soft thresholds.

[0074] 4) Residual structure: alleviates gradient vanishing and improves training stability.

[0075] 5) Adaptive threshold learning: Dynamically adjusts the denoising intensity to adapt to input signals with different levels of contamination.

[0076] like Figure 4 As shown, an enhanced shuffle network unit module is constructed. This enhanced shuffle network unit module is an improved ShuffleNet unit used for lightweight feature extraction. The specific steps include the following:

[0077] Channel grouping: grouping input features The channels are divided into G groups, and each group contains C / G channels.

[0078] Intra-group convolution: Channel expansion is performed using 1×1 convolution, increasing the number of channels in each group from C / G to K*C / G, where K is the expansion factor. Feature extraction is then performed using 3×1 depthwise separable convolution, ultimately outputting the features. .

[0079] Squeeze operation: on features Perform global max pooling and global average pooling to obtain two vectors. .

[0080] Excitation operation: Excitation vector and After concatenation, the input is given to a small multilayer perceptron, and the output features are summed to obtain the weight vector. The small multilayer perceptron includes two linear layers: the first layer is a fully connected layer with ReLU and Dropout, and the second layer is a fully connected layer with a Hard-Sigmoid activation function.

[0081] Scale operation: Scales the weight vector With features Multiply to obtain the weighted features .

[0082] Introducing residual connections: features The result after 1×1 convolution and the input features splicing, or, features The result after 1×1 convolution and the input features The results are stitched together after passing through a 3×1 depth separable convolutional layer and a 1×1 convolutional layer in sequence.

[0083] Channel shuffling: feature map Rearrange the channels into H×1×G×(KC / G); transpose the channel dimensions to interleave channels from different groups; restore the format to H×1×KC to obtain the output of the enhanced shuffle network unit module.

[0084] The enhanced shuffling network unit module, as a key component in the classification model, achieves efficient and lightweight feature extraction through the following innovations:

[0085] 1) Intragroup convolution and depthwise separable convolution: significantly reduce computation and parameter count.

[0086] 2) Channel rearrangement mechanism: Breaks the channel isolation of group convolution and improves feature fusion capability.

[0087] 3) SE attention mechanism: dynamically adjust channel weights to enhance attention to key features.

[0088] 4) Residual connections: alleviate gradient vanishing and improve training stability.

[0089] 5) Multi-stage stacking structure: Gradually extract abstract features to adapt to the information abstraction needs at different levels.

[0090] A dual-attention adaptive residual shrinking module and n enhanced shuffling network unit modules are connected sequentially, and the input and output of the dual-attention adaptive residual shrinking module are added together as the input of the first enhanced shuffling network unit module.

[0091] (3) Constructing dilated convolution pyramid pooling units

[0092] Five branches are constructed. The first branch consists of an adaptive average pooling layer, a 1×1 convolutional layer, and an upsampling layer connected in sequence. The second branch is a 3×1 convolutional layer with a dilation rate of 18. The third branch is a 3×1 convolutional layer with a dilation rate of 12. The fourth branch is a 3×1 convolutional layer with a dilation rate of 6. The fifth branch is a 1×1 convolutional layer.

[0093] The outputs of the five branches of the dilated convolutional pyramid pooling unit are concatenated and then connected to a dual-attention adaptive residual shrinking module. The input and output of the dual-attention adaptive residual shrinking module are added together and then connected to an adaptive average pooling layer. The output of the adaptive average pooling layer is connected to the input of the classifier.

[0094] After completing the initial lightweight microseismic signal classification model, the model is trained iteratively using the training set until convergence. The specific steps include:

[0095] The input layer of the lightweight microseismic signal classification model receives vibration signals from the microseismic signal classification dataset.

[0096] The input signal undergoes preliminary feature extraction and downsampling through a 3×1 max pooling layer with a stride of 2 and a 3×1 convolutional layer with a stride of 2.

[0097] The downsampled signal features are further abstracted by a feature extraction unit consisting of an enhanced shuffling network unit module and a dual-attention adaptive residual shrinking module, while simultaneously achieving adaptive noise reduction.

[0098] The dilated convolutional pyramid pooling unit further extracts multi-scale features.

[0099] After concatenating the multi-scale features, the input is fed into a dual-attention adaptive residual shrinking module. The input and output of the dual-attention adaptive residual shrinking module are added together and then fed into an adaptive average pooling layer to obtain high-dimensional features.

[0100] The classifier receives high-dimensional features and provides classification results for microseismic signals. It then compares the predicted values ​​of the classified microseismic signals with the actual values, calculates the error value, and iterates through the training process.

[0101] The trained lightweight microseismic signal classification model was validated using a validation set. The model parameters were adjusted and retrained to determine the final lightweight microseismic signal classification model.

[0102] The final lightweight microseismic signal classification model was tested using a test set, and evaluation metrics were statistically analyzed.

[0103] To quantitatively analyze the accuracy of the model in classifying microseismic signals, experiments were conducted to analyze its classification performance, lightweight performance, and noise robustness, as detailed below:

[0104] (1) Classification performance evaluation

[0105] To test the classification performance of the proposed lightweight microseismic signal classification model, the model was trained and tested on 30,000 single-channel acceleration waveforms from tunnel projects in southwestern China. The dataset was divided into three categories: blasting signals, microseismic signals, and noise signals, with 10,000 samples in each category and 4,000 sampling points in each sample.

[0106] 1) Data preprocessing and partitioning

[0107] The data was normalized to the maximum absolute value; the training set, validation set, and test set were divided in an 8:1:1 ratio; to address the issue of a small number of blasting signals, data augmentation methods (such as left-right shifting and up-down flipping) were used to balance the data distribution; the labels were corrected multiple times by multiple experts to reduce noise interference.

[0108] 2) Model performance metrics

[0109] The evaluation is conducted using three core metrics: accuracy, precision, and recall.

[0110] .

[0111] .

[0112] .

[0113] Where TP is the number of correctly identified positive samples, FP is the number of negative samples that were misclassified as positive samples, TN is the number of correctly identified negative samples, and FN is the number of positive samples that were misclassified as negative samples.

[0114] 3) Experimental Results

[0115] The confusion matrix of the microseismic signal classification results is as follows Figure 5 As shown, in the test set, the lightweight microseismic signal classification model constructed in this application achieved an overall classification accuracy of 97.7%. The precision rates for each category were: 100% for blasting signals, 96.9% for microseismic signals, and 96.2% for noise signals. The recall rates were: 100% for blasting signals, 96.2% for microseismic signals, and 96.9% for noise signals. The confusion matrix showed that all blasting signals were correctly identified, while there was a small amount of confusion between microseismic signals and noise signals, mainly due to the high similarity of their waveforms. Further analysis revealed that the misclassified microseismic signals and noise signals had extremely low amplitudes and energy, making it difficult even for experienced engineers to accurately distinguish them based on a single-channel waveform. Therefore, these misclassifications had little impact on the actual warning level judgment, and the overall classification accuracy far exceeded the industrial application standard.

[0116] (2) Comparison of lightweight performance

[0117] To verify the applicability of the lightweight microseismic signal classification model in resource-constrained environments, we selected mainstream lightweight network architectures for comparison, including GoogLeNet, ResNet, ShuffleNet, EfficientNet, MobileNet, DenseNet, and Conformer-NSE. Model complexity was measured by the number of parameters and floating-point operations (FLOPs), and the classification accuracy of each model was compared. The comparison results are as follows: Figure 5 As shown.

[0118] 1) Comparison of model complexity

[0119] The lightweight microseismic signal classification model has only 0.155M parameters, which is much lower than other models (such as ShuffleNet-V2 at 0.338M and MobileNet-V2 at 2.184M). The lightweight microseismic signal classification model has 38.237M floating-point operations, which is slightly higher than Convformer-NSE (24.059M), but significantly better than other models (such as EfficientNet at 529.144M).

[0120] 2) Performance Comparison

[0121] The lightweight microseismic signal classification model achieved a classification accuracy of 97.7%, which is the best among all lightweight models. Although EfficientNet's accuracy is close (97.4%), its number of parameters and computational cost are tens of times higher. The lightweight microseismic signal classification model showed excellent performance in the early stages of training, indicating that its small number of parameters can achieve fast convergence.

[0122] (3) Noise robustness analysis

[0123] Considering that microseismic signals collected at engineering sites are often subject to various noise interferences, this application evaluates the classification stability of a lightweight microseismic signal classification model under different noise intensities by artificially synthesizing noise signals with different signal-to-noise ratios.

[0124] 1) Noise Types and Synthesis Methods

[0125] Using Gaussian white noise and real noise collected on-site, six sets of test data were generated with signal-to-noise ratio ranges of [10,15), [15,20), [20,25), [25,30), [30,35), and [35,+∞) by adjusting the noise amplitude.

[0126] 2) Classification accuracy performance

[0127] When the signal-to-noise ratio is ≥20, all models can achieve a classification accuracy of approximately 100%.

[0128] In the low signal-to-noise ratio ranges [10, 15) and [15, 20), the lightweight microseismic signal classification model significantly outperforms other models:

[0129] In the [10,15) interval, the lightweight microseismic signal classification model achieved an accuracy of 85%, while ShuffleNet achieved only 3%;

[0130] In the [15,20) range, the lightweight microseismic signal classification model still maintains a performance of 94%, which is significantly better than other models.

[0131] The results show that the dual-attention adaptive residual shrinkage module introduced into the classification model has good adaptive denoising ability, can effectively retain key feature information, and still maintains a high recognition accuracy under low signal-to-noise ratio conditions.

[0132] The present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. A lightweight model construction method for microseismic signal classification, characterized in that, Specifically, the steps include the following: The micro-fracture signals collected by the microseismic monitoring system are acquired and preprocessed to obtain a microseismic signal classification dataset. The microseismic signal classification dataset is then divided into a training set, a validation set, and a test set according to a preset ratio. An initial lightweight microseismic signal classification model was constructed, and the hyperparameters for training the lightweight microseismic signal classification model were set. The initial lightweight model includes an input layer, a lightweight feature extraction backbone network, and a classifier connected in sequence. The lightweight feature extraction backbone network includes a downsampling unit, three feature extraction units consisting of an enhanced shuffling network unit module and a dual attention adaptive residual shrinking module stacked alternately, and a dilated convolution pyramid pooling unit connected in sequence. The initial lightweight microseismic signal classification model was trained iteratively using the training set until convergence. The trained lightweight microseismic signal classification model was validated using a validation set. The model parameters were adjusted and retrained to determine the final lightweight microseismic signal classification model. The final lightweight microseismic signal classification model was tested using a test set, and evaluation metrics were statistically analyzed. The lightweight feature extraction backbone network for constructing the initial lightweight microseismic signal classification model includes the following steps: Construct downsampling units, feature extraction units, and dilated convolutional pyramid pooling units; Feature extraction unit construction: Construct a dual-attention adaptive residual shrinkage module and an enhanced shuffle network unit module. Connect one dual-attention adaptive residual shrinkage module and n enhanced shuffle network unit modules in sequence. Add the input and output of the dual-attention adaptive residual shrinkage module as the input of the first enhanced shuffle network unit module. The dual-attention adaptive residual shrinkage module is a deep learning-based adaptive residual shrinkage mechanism, combined with channel attention and spatial attention, for noise suppression and feature enhancement. The enhanced shuffle network unit module is an improved ShuffleNet unit for lightweight feature extraction. Constructing a dual-attention adaptive residual shrinkage module includes the following steps: Constructing a channel attention mechanism: Performing global max pooling on the input features to obtain a vector. And global average pooling to obtain vectors ; to vector and Input a small multilayer perceptron and sum the output features to obtain the channel weight vector. In this small multilayer perceptron, the first layer is a fully connected layer, ReLU, and Dropout; the second layer is a fully connected layer and a Hard-Sigmoid activation function; the weights are normalized using the Sigmoid function and then compared with the global maximum value M. max Multiply to obtain the channel-level threshold. ; Constructing a spatial attention mechanism: Perform global max pooling on the input features and compress them along the channel dimension to obtain a vector. And global average pooling to obtain vectors ; to vector and After concatenation, input into a 3×1 convolutional layer to obtain the spatial weight vector. Normalize the weights using the Sigmoid function and compare them with the global average A. avg Multiply to obtain the spatial threshold. ; The nonlinear soft threshold function is constructed as follows: ； Where x is the feature before denoising, y is the feature after denoising, and τ is the channel-level threshold or spatial-level threshold; Residual connections are introduced: Input features are preprocessed sequentially through a 3×1 depthwise separable convolutional layer and a 1×1 convolutional layer; the preprocessed features are then processed by a channel attention mechanism, and a nonlinear soft thresholding function is used to perform a first transformation between the channel-level threshold and the preprocessed features; the features after the first transformation are then processed by a spatial attention mechanism, and a nonlinear soft thresholding function is used to perform a second transformation between the spatial-level threshold and the features after the first transformation; the input features and the features after the second transformation are added element-wise to obtain the output of the dual-attention adaptive residual shrinkage module. The construction of the enhanced shuffling network unit module includes the following steps: Channel grouping: grouping input features The channels are divided into G groups, and each group contains C / G channels; Intra-group convolution: Channel expansion is performed using 1×1 convolution, increasing the number of channels in each group from C / G to K*C / G, where K is the expansion factor. Feature extraction is then performed using 3×1 depthwise separable convolution, ultimately outputting the features. ; Squeeze operation: on features Perform global max pooling and global average pooling to obtain two vectors. ; Excitation operation: Excitation vector and After concatenation, the input is given to a small multilayer perceptron, and the output features are summed to obtain the weight vector. The small multilayer perceptron includes two linear layers: the first layer is a fully connected layer, ReLU, and Dropout; the second layer is a fully connected layer and a Hard-Sigmoid activation function. Scale operation: Scales the weight vector With features Multiply to obtain the weighted features ; Introducing residual connections: features The result after 1×1 convolution and the input features splicing, or, features The result after 1×1 convolution and the input features The results after sequentially passing through a 3×1 depth separable convolutional layer and a 1×1 convolutional layer are stitched together; Channel shuffling: feature map Rearrange the channels into H×1×G×(KC / G); transpose the channel dimensions to interleave channels from different groups; restore the format to H×1×KC to obtain the output of the enhanced shuffle network unit module.

2. The lightweight model construction method for microseismic signal classification according to claim 1, characterized in that, Construct downsampling units: connect a 3×1 max pooling layer with a stride of 2 and a 3×1 convolutional layer with a stride of 2 in sequence; Construct a dilated convolutional pyramid pooling unit: Construct five branches. The first branch consists of an adaptive average pooling layer, a 1×1 convolutional layer, and an upsampling layer connected in sequence. The second branch is a 3×1 convolutional layer with a dilation rate of 18. The third branch is a 3×1 convolutional layer with a dilation rate of 12. The fourth branch is a 3×1 convolutional layer with a dilation rate of 6. The fifth branch is a 1×1 convolutional layer. The outputs of the five branches of the dilated convolutional pyramid pooling unit are concatenated and then connected to a dual-attention adaptive residual shrinking module. The input and output of the dual-attention adaptive residual shrinking module are added together and then connected to an adaptive average pooling layer. The output of the adaptive average pooling layer is connected to the input of the classifier.

3. The lightweight model construction method for microseismic signal classification according to claim 1, characterized in that, It also includes iterative training of the initial lightweight microseismic signal classification model based on the microseismic signal classification dataset until convergence, specifically including the following steps: The input layer of the lightweight microseismic signal classification model receives vibration signals from the microseismic signal classification dataset; The input signal undergoes preliminary feature extraction and downsampling through a 3×1 max pooling layer with a stride of 2 and a 3×1 convolutional layer with a stride of 2; The downsampled signal feature input is further abstracted by a feature extraction unit composed of an enhanced shuffling network unit module and a dual-attention adaptive residual shrinking module, which simultaneously achieves adaptive noise reduction. Diffuse convolutional pyramid pooling units further extract multi-scale features; After concatenating the multi-scale features, the input is fed into a dual-attention adaptive residual shrinking module. The input and output of the dual-attention adaptive residual shrinking module are added together and then fed into an adaptive average pooling layer to obtain high-dimensional features. The classifier receives high-dimensional features and provides classification results for microseismic signals. It then compares the predicted values ​​of the classified microseismic signals with the actual values, calculates the error value, and iterates through the training process.

Images