Deep learning-based few-sample micro-seismic signal identification method

Abstract

The invention discloses a small-sample micro-seismic signal identification method based on deep learning. The method comprises the step of classifying a double-branch prototype network and a composite distance metric function. The dual-branch network cooperatively extracts multi-scale features through DEA-CNN and CNN, and the feature expression ability is enhanced; the composite distance measurement is combined with the standard Euclidean distance and the normalized Euclidean distance, and the weight is adaptively adjusted to improve the classification precision. Experiments show that the method achieves 82.52% of classification precision and 93.05% of classification precision in 4-way-1-shot tasks and 4-way-5-shot tasks on a data set of a spathiphyllum beach hydropower station respectively, and the method is remarkably superior to an existing method. The problems of fuzzy feature difference, strong noise interference and dependence on large-scale labeled data in a complex construction environment can be effectively solved, and technical support is provided for tunnel microseism monitoring.

Description

Technical Field

[0001] This invention belongs to the field of microseismic monitoring technology in tunnel engineering, specifically, it relates to a method for identifying microseismic signals with few samples based on deep learning. Background Technology

[0002] Microseismic monitoring technology, as a passive monitoring method, has been widely used in tunnel construction. As excavation progresses, microseismic activity continuously occurs in the surrounding rock, and monitoring systems can record a large amount of signal data in real time. This data not only includes effective microseismic signals caused by rock fracturing but also interference signals from blasting operations, mechanical excavation, drilling, and environmental noise. Therefore, accurately identifying microseismic events from massive, complex, and heterogeneous monitoring data has become the core objective of microseismic monitoring in tunnel engineering. In recent years, researchers have proposed various automatic identification methods based on waveform characteristics, such as methods based on statistical analysis, spectral analysis, and energy ratios. These methods have achieved certain results in some tunnel monitoring applications, but they usually require experienced researchers to adjust parameters according to specific geological conditions and construction environments, thus exhibiting poor adaptability in complex and variable tunnel environments.

[0003] With the rapid development of deep learning technology, intelligent identification of microseismic signals has ushered in new opportunities. Deep learning methods have demonstrated higher recognition accuracy and robustness under complex excavation conditions. For example, researchers have developed end-to-end 3D convolutional networks, introducing residual structures and attention mechanisms, or combining spatial pyramid pooling layers to enhance adaptability to multi-channel monitoring data and complex waveform patterns. In terms of feature enhancement, models combining temporal attention mechanisms and gated recurrent units, as well as frameworks fusing VGG and channel-spatial attention mechanisms, have significantly improved the extraction capability of key features in noisy environments. In addition, optimization strategies such as transfer learning and genetic algorithms have been used to improve the generalization ability of models under different geological conditions and construction stages. However, the above methods generally rely on large-scale, high-quality labeled datasets, while in tunnel engineering, due to the complex and variable signal characteristics and the high cost of data acquisition and labeling, it is often difficult to obtain sufficient training data. In actual construction, the amount of manually labeled microseismic data available is limited, making it difficult for deep learning models to train high-precision signal recognizers, seriously affecting the accuracy and reliability of monitoring results.

[0004] The proposed few-shot learning method offers a new approach to solving the aforementioned problems. Few-shot learning can achieve near-conventional model recognition performance with a very small number of labeled samples, and mainly includes three categories: transfer learning, meta-learning, and metric learning. Among them, metric learning measures the similarity between the support set and the query samples by constructing a prototype representation space, offering advantages such as clear modeling and suitability for low-resource scenarios. Typical prototype networks extract embedded features through deep neural networks, calculate category prototypes, and perform classification based on Euclidean distance, combining mathematical simplicity with strong generalization ability, and have achieved success in few-shot tasks such as medical image analysis and natural image classification. However, in tunnel microseismic signal classification tasks, due to the complex and easily confused feature differences between signal categories, prototypes often fail to accurately represent the features of each category. Furthermore, traditional Euclidean distance metrics struggle to effectively distinguish similar signals under strong noise interference conditions, further limiting the application of this method in practical tunnel environments.

[0005] Currently, traditional methods for identifying microseismic signals in tunnels largely rely on statistical features, spectral features, and energy ratios. Their core drawback lies in the need for extensive manual parameter adjustments, making them ill-suited for complex construction environments. With the development of deep learning, researchers have proposed end-to-end convolutional networks, attention mechanisms, and transfer learning, which have improved recognition accuracy and robustness to some extent. However, deep learning methods generally depend on large-scale, high-quality labeled data. In practical tunnel engineering applications, labeled samples are scarce, and signal features are complex, making it difficult for models to obtain sufficient training support. The emergence of few-shot learning methods provides a new approach for microseismic signal identification under limited sample conditions. Prototype networks, with their simple and effective metric learning mechanisms, have achieved good results in few-shot tasks. However, in complex tunnel environments, the differences between microseismic signal categories are blurred, and noise interference is strong. Existing prototype networks, based on Euclidean distance, are prone to inaccurate differentiation of similar signals and feature overlap between different signal categories, thus affecting recognition accuracy. These problems indicate that existing few-shot learning methods still have shortcomings in the application of tunnel microseismic signal identification, requiring further improvements to the model's feature extraction and metric mechanisms to enhance its applicability and reliability in complex environments.

[0006] In summary, in practical tunnel engineering, due to the complex and variable characteristics of signals, coupled with the high cost of data acquisition and annotation, it is often difficult to obtain sufficient training data. Furthermore, the limited amount of manually annotated microseismic data available makes it difficult for deep learning models to train high-precision signal recognizers, severely impacting the accuracy and reliability of monitoring results. Therefore, developing a microseismic signal recognition method capable of efficiently extracting features and accurately classifying them under conditions of limited labeled samples has significant practical importance and application value. Summary of the Invention

[0007] The purpose of this invention is to provide a few-sample microseismic signal identification method based on deep learning, which mainly solves the problems of blurred feature differences, strong noise interference, and reliance on large-scale high-quality labeled datasets in existing tunnel microseismic signal identification methods under complex construction environments.

[0008] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0009] A method for identifying few-sample microseismic signals based on deep learning includes the following steps:

[0010] S1, establish a classification dual-branch prototype network using a weighted combination of DEA-CNN and CNN networks;

[0011] S2 extracts signal features from the input sample signal through a classification dual-branch prototype network and performs feature fusion based on weights to obtain feature representations of the support set and query set of the sample signal;

[0012] S3, calculate the class prototype of the sample signal and its L2 normalized class prototype;

[0013] S4, calculates the standardized Euclidean distance between the feature vectors of the class prototype and the feature vectors of the query set, and the L2 normalized Euclidean distance between the feature vectors of the class prototype and the feature vectors of the query set based on the composite distance metric function.

[0014] S5 weights and fuses the standardized Euclidean distance and the L2 normalized Euclidean distance to obtain the final distance metric, which serves as the basis for classification.

[0015] S6 uses the SoftMax probability calculation module to output the classification probability based on the distance metric results.

[0016] Furthermore, in this invention, the DEA-CNN network uses depthwise separable convolutional modules to construct core units, extracts spatial features through 3×3 depthwise convolutional kernels, and combines 1×1 pointwise convolution to complete channel-dimensional feature reorganization, effectively reducing computational load while ensuring feature representation capabilities. After batch normalization layers and ReLU activation function processing, a channel-space dual-path attention mechanism is introduced. Through channel-dimensional feature recalibration and spatial domain focus on regions of interest, dynamic enhancement of key features and simultaneous suppression of unnecessary information are achieved. A 2×2 max pooling layer is used to perform feature dimensionality reduction, and a hierarchical feature extraction process is constructed through cascaded network modules, realizing spatial feature deepening and channel information fusion during the layer-by-layer transmission process.

[0017] Furthermore, in this invention, the CNN network adopts a standard CNN network architecture, and multi-scale feature parsing is achieved through five cascaded convolutional modules; each convolutional module performs spatial feature extraction by 3×3 convolutional kernels, and after batch normalization and ReLU activation function processing, feature downsampling is implemented through 2×2 max pooling layers.

[0018] Furthermore, in step S4, the standardized Euclidean distance between the feature vector of the class prototype and the feature vector of the query set is... The calculation formula is as follows:

[0019] ,

[0020] in, Feature vectors representing class prototypes =(𝑥1,𝑥2,…,𝑥 𝐷 ); Represents the feature vector of the query set. ; This represents the scaling factor in the 𝑖-th dimension. Let represent the component of the eigenvector in the i-th dimension, and D represent the number of dimensions of the eigenvector.

[0021] Further, in step S4, the calculation process of the L2-normalized Euclidean distance between the feature vector of the L2-normalized class prototype and the feature vector of the query set is as follows:

[0022] Calculate the normalized feature vector of the class prototype and the feature vector of the normalized query set :

[0023] ;

[0024] Calculate the L2 normalized Euclidean distance :

[0025] .

[0026] Furthermore, in step S5, the final distance metric is calculated using the following formula:

[0027]

[0028] In the formula, These are weighting coefficients used to adjust the relative contributions of the standard Euclidean distance and the L2-normalized Euclidean distance to the overall metric.

[0029] Further, in step S6, the formula for calculating the classification probability based on the distance metric result is as follows:

[0030]

[0031] In the formula, y represents the feature vector Z of the query set. qj The category to which it belongs, m represents the category loop variable, and C represents the total number of categories.

[0032] Compared with the prior art, the present invention has the following beneficial effects:

[0033] (1) This invention constructs a classification dual-branch prototype network, which combines the DEA-CNN network with the standard CNN network in a weighted manner: DEA-CNN adopts a depthwise separable convolution and a channel-space dual-path attention mechanism to dynamically enhance key features and suppress noise; CNN achieves multi-scale feature parsing through five layers of cascaded convolution. Dual-branch feature fusion can fully capture the spatial details and channel semantic information of microseismic signals, effectively solve the problems of ambiguous signal feature differences and key features being easily masked by noise in complex construction environments, and improve the robustness of feature representation.

[0034] (2) This invention proposes a composite distance metric function: combining standard Euclidean distance and L2-normalized Euclidean distance, and balancing the contributions of the two through an adaptive weighting coefficient β. This mechanism can reduce the impact of feature scale differences on similarity measurement and enhance the model's ability to resist interference from complex backgrounds such as blasting interference and mechanical noise. In the experiment, the classification accuracy reached 82.52% on the 4-way-1-shot task of the Baihetan Hydropower Station dataset, which is significantly better than existing few-shot methods.

[0035] (3) Based on the prototype network framework, this invention can efficiently learn prototype features even in scenarios with scarce labeled samples through the collaborative design of dual-branch feature extraction and composite distance metric. The deep separable convolution and attention mechanism of DEA-CNN improves feature discriminativeness while reducing computational load. Combined with weighted fusion distance metric, the model achieves an accuracy of 93.05% in the 4-way-5-shot task. This method effectively solves the pain points of scarce labeled data and high acquisition costs of microseismic signals in tunnel engineering. It can achieve high-precision identification without large-scale labeled data, providing a low-cost and highly reliable technical solution for practical engineering monitoring. Attached Figure Description

[0036] Figure 1 This is a schematic diagram of the overall network flow of the present invention.

[0037] Figure 2 This is a schematic diagram of the Deep Enhanced Attention Convolutional Neural Network (DEA-CNN) structure of the present invention.

[0038] Figure 3 This is a schematic diagram of the standard convolutional neural network (CNN) structure of this invention.

[0039] Figure 4 This is a comparison chart of the normalized confusion matrix of the present invention on the Baihetan Hydropower Station dataset.

[0040] Figure 5 This is a visualization of the T-SNE effect of the present invention on the Baihetan Hydropower Station dataset. Detailed Implementation

[0041] The present invention will be further described below with reference to the accompanying drawings and embodiments. The embodiments of the present invention include, but are not limited to, the following embodiments.

[0042] This invention discloses a few-sample microseismic signal recognition method based on deep learning, the overall process of which is as follows: Figure 1 As shown, the process includes the implementation of a dual-branch prototype network for classification, a feature fusion mechanism, and a composite distance metric function. Specifically, after preprocessing, the input signal is divided into a support set and a query set, which are used for class prototype calculation and classification decision, respectively. Each sample in the support set generates a feature representation through a dual-branch feature extraction network, and then generates the final feature representation through the feature fusion mechanism. Samples in the query set also generate feature representations through a dual-branch feature extraction network, and their similarity is measured with the class prototypes in the support set using a composite distance metric function, ultimately outputting the classification result.

[0043] The classification dual-branch prototype network uses a weighted combination of DEA-CNN and CNN networks; the DEA-CNN network is as follows: Figure 2 As shown, the network takes an 84×84 pixel preprocessed image as input and uses a five-layer depthwise separable convolutional module to construct the core unit. Spatial feature extraction is achieved through a 3×3 depthwise convolutional kernel, and channel dimension feature reorganization is completed by combining 1×1 pointwise convolution, which effectively reduces the computational load while ensuring feature representation capability. After batch normalization layer and ReLU activation function processing, a channel-space dual attention mechanism is introduced. Through channel dimension feature recalibration and spatial domain focus, the dynamic enhancement of key features and the synchronous suppression of unnecessary information are achieved. A 2×2 max pooling layer is used to perform feature dimensionality reduction operation. A hierarchical feature extraction process is constructed through cascaded network modules, and spatial feature deepening and channel information fusion are achieved in the process of layer-by-layer transmission. This effectively alleviates the information decay problem, enhances the network's ability to capture discriminative features, and thus improves the overall performance of the model. As shown in formula (1), the network performs feature transformation on the input preprocessed image x, and finally outputs a feature map Z1 with a dimension of 2×2×64.

[0044] (1)

[0045] The CNN network adopts the standard CNN network architecture (structure as follows) Figure 3As shown in the figure, an 84×84 pixel preprocessed image is input, and multi-scale feature parsing is achieved through a five-layer cascaded convolutional module. Each module performs spatial feature extraction with a 3×3 convolutional kernel (Padding=1), and after batch normalization and ReLU activation function processing, feature downsampling is performed through a 2×2 max pooling layer. This hierarchical design captures local details of the image to global semantic features through progressive convolutional layers. Pooling operation effectively suppresses feature redundancy and enhances spatial translation invariance, thereby improving the generalization ability and context information representation efficiency of the model. Each convolutional module extracts and fuses discriminative features under different receptive fields through a hierarchical feature abstraction mechanism, and finally outputs a feature map Z2 with a dimension of 2×2×64, the mathematical expression of which is shown in formula (2).

[0046] (2)

[0047] To achieve collaborative optimization of feature representations in a dual-branch network, this embodiment balances the contributions of bidirectional features through a weight-based feature fusion mechanism, maximizing the complementary advantages of heterogeneous features. Specifically, Feature Extractor 1 (composed of DEA-CNN) and Feature Extractor 2 (composed of CNN) fuse their output features with a fixed weight of 0.5. During real-time model training, this mechanism avoids additional learnable hyperparameters and directly achieves feature combination through this weighted fusion strategy. This method improves the overall classification efficiency of the model, as shown in Equation (3):

[0048] (3)

[0049] In the ProtoNet network, standard Euclidean distance is relied upon to measure the similarity between class prototypes and the query set. However, due to the limited number of support set samples and significant differences in feature distribution between classes in the microseismic event recognition task, standard Euclidean distance cannot fully account for the differences in local structure and feature vector orientation information in the feature space, thus affecting recognition accuracy. Therefore, a hybrid metric function is adopted to improve the stability of the distance metric and recognition accuracy. Standard Euclidean distance and normalized Euclidean distance are dynamically fused through learnable parameter delays, adaptively adjusting the weights of the two metrics. First, the feature vectors of the class prototypes are calculated using standard Euclidean distance. and the feature vector of the query set Geometric distance between them, distance Calculated by equation (4). In this context, the feature vector of the class prototype is represented as: =(𝑥1,𝑥2,…,𝑥 𝐷 ), and the feature vector of the query set ; This represents the scaling factor in the 𝑖-th dimension. Let represent the component of the eigenvector in the i-th dimension, and D represent the number of dimensions of the eigenvector.

[0050] (4)

[0051] Because the feature dimensions of the class prototype and the query sample differ in scale, directly calculating the Euclidean distance will be affected by the scale factor λ. 𝑖 Interference. When 𝑘 𝑖 When the differences are significant, the distance will be biased towards the larger dimension, leading to the failure of feature similarity judgment. For this reason, this embodiment introduces standardized Euclidean distance. The feature vectors are normalized so that the modulus of each feature vector is 1, thereby shifting the focus from the size of the feature to its orientation information, effectively eliminating the influence of feature scale, and making the distance metric more stable and discriminative. We calculate the normalized class prototype feature vectors. and query set feature vector It can be obtained from the calculation formula (5):

[0052] (5)

[0053] The formula for calculating the normalized Euclidean distance (6) is as follows:

[0054] (6)

[0055] Furthermore, this embodiment employs a weighted fusion of the standard Euclidean distance and the normalized Euclidean distance to fully leverage their complementary advantages in measuring spatial structures. The weighted Euclidean distance formula is shown in equation (7):

[0056] (7)

[0057] in, The weighting coefficients are used to adjust the relative contributions of the standard Euclidean distance and the normalized Euclidean distance to the overall metric. Weighted Euclidean distance mitigates the sensitivity of traditional Euclidean distance to feature scale differences to some extent, while enhancing the model's focus on feature orientation information. By increasing the weight of the normalized Euclidean distance in the total distance, the model can more accurately capture semantic similarity between samples, thereby improving robustness and discriminative ability during feature learning.

[0058] Finally, based on the weighted fusion distance metric results, the SoftMax probability calculation module uses formula (8) to make classification decisions:

[0059] (8)

[0060] In the formula, y represents the feature vector Z of the query set. qjThe category to which the model belongs is defined, where m represents the category loop variable and C represents the total number of categories. By normalizing the feature vector, the model's adaptability to changes in feature scale is enhanced, and the risk of overfitting is reduced.

[0061] To verify the effectiveness of the proposed model in microseismic event identification, this embodiment compares the proposed model with other few-shot learning methods using the best classification results of each model. For each model, we save the model state for each time period. Then, we select four classes, with one sample per class as the support set and 15 samples per class as the query set. The effectiveness of the model is verified on the Baihetan Hydropower Station dataset. As shown in Table 1, on the Baihetan Hydropower Station dataset, DB-ProtoNet significantly outperforms the baseline model ProtoNet and other few-shot methods in both the 1-sample and 5-sample settings, achieving the highest classification accuracy.

[0062] Table 1 compares the performance of this model with other few-shot classification models on the Baihetan Hydropower Station dataset.

[0063]

[0064] Considering the potential for imbalanced predictions when handling different categories, this embodiment uses both the ProtoNet and DB-ProtoNet models to perform a 4-class 1-sample test on the Baihetan Hydropower Station dataset. Figure 4 As shown, the model's classification results for four events in the normalized confusion matrix were observed. The results indicate that the model's ability to distinguish between different categories is enhanced, and the classification accuracy and misclassification rate do not fluctuate significantly. This demonstrates that the model has good discriminative ability and stability, illustrating the effectiveness of the proposed method.

[0065] To further demonstrate the superiority of the features extracted by the model in category recognition, T-SNE dimensionality reduction technology was used to visualize the intra-spatial distribution of features for samples of different categories. Figure 5 As can be seen from the Baihetan Hydropower Station dataset, when comparing the ProtoNet model with the DB-ProtoNet model, the distribution of the four classes of samples in the DB-ProtoNet model shows obvious clustering, and the boundaries between different classes are clear.

[0066] The above embodiments are merely one of the preferred embodiments of the present invention and should not be used to limit the scope of protection of the present invention. Any modifications or refinements made to the main design concept and spirit of the present invention that are not of substantial significance, but solve the same technical problem as the present invention, should be included within the scope of protection of the present invention.

Claims

1. A deep learning-based few-shot microseismic signal recognition method, characterized in that, The method comprises the following steps: S1, a classification double-branch prototype network using DEA-CNN network and CNN network weighted combination is established; S2, signal feature extraction is performed on the input sample signal through the classification double-branch prototype network, and feature fusion is performed based on the weight, to obtain the feature representation of the support set and the query set of the sample signal; S3, the class prototype and the L2 normalized class prototype of the sample signal are calculated; S4, the standardized Euclidean distance between the feature vector of the class prototype and the feature vector of the query set and the L2 normalized Euclidean distance between the feature vector of the L2 normalized class prototype and the feature vector of the query set are calculated based on the composite distance measurement function; S5, the standardized Euclidean distance and the L2 normalized Euclidean distance are weighted and fused to obtain the final distance measurement as the basis for classification; S6, the SoftMax probability calculation module is used to output the classification probability according to the distance measurement result.

2. The few-shot microseismic signal recognition method based on deep learning according to claim 1, characterized in that, The DEA-CNN network adopts a deep separable convolution module to construct a core unit, realizes spatial feature extraction through a 3*3 deep convolution kernel, completes channel dimension feature reorganization through a 1*1 point-by-point convolution, effectively reduces the calculation load on the premise of ensuring the feature representation capability, is processed through a batch normalization layer and a ReLU activation function, introduces a channel-spatial dual-path attention mechanism, realizes dynamic enhancement of key features and synchronous suppression of unnecessary information through feature re-labeling in the channel dimension and focusing on the attention area in the spatial domain, and realizes spatial feature deepening and channel information fusion in the process of layer-by-layer transmission through a 2*2 maximum pooling layer to implement feature dimension reduction operation and a cascaded network module to construct a hierarchical feature extraction process.

3. The method according to claim 2, wherein, The CNN network adopts a standard CNN network architecture, realizes multi-scale feature analysis through five cascaded convolution modules, each convolution module performs spatial feature extraction through a 3*3 convolution kernel, is processed through a batch normalization and a ReLU activation function, and implements feature down-sampling through a 2*2 maximum pooling layer.

4. The few-shot microseismic signal recognition method based on deep learning according to claim 3, characterized in that, In step S4, the normalized Euclidean distance between the feature vector of the class prototype and the feature vector of the query set is... The calculation formula is as follows: ， wherein, represents a feature vector of a class prototype, = (x1, x2, …, x 𝐷 ); represents a feature vector of a query set, ; represents a scaling factor in the i-th dimension, represents a component of the feature vector in the i-th dimension, and D represents the number of dimensions of the feature vector.

5. The few-shot microseismic signal recognition method based on deep learning according to claim 4, characterized in that, In the step S4, the calculation process of the L2 normalized Euclidean distance between the feature vector of the L2 normalized class prototype and the feature vector of the query set is as follows: computing the feature vector of the normalized class prototype and the feature vector of the normalized query set : ； calculating the L2 normalized Euclidean distance : 。 6. The few-shot microseismic signal recognition method based on deep learning according to claim 5, characterized in that, In the step S5, the calculation formula of the final distance measurement is as follows: wherein are weighting factors used to adjust the relative contribution of the standard Euclidean distance and the L2 normalized Euclidean distance in the overall metric.

7. The few-shot microseismic signal recognition method based on deep learning according to claim 6, characterized in that, In the step S6, the calculation formula of the classification probability according to the distance measurement result is as follows: where y represents the feature vector Z of the query set qj where m represents the class loop variable and C represents the total number of classes.

Images