A microseismic signal feature dynamic change adaptive classification method

Abstract

The application relates to a self-adaptive classification method for dynamic changes of microseismic signal characteristics, and relates to the technical field of intelligent monitoring and information technology.Ada-VIT model integrating a feature extractor, a multi-label classifier and a domain discriminator is constructed; an adaptive mechanism is used to optimize the features so as to improve the recognition accuracy of the microseismic signals under complex change environment; clean microseismic signals and pure noise signals are used as source domain samples for model training, and the microseismic signals which are more complex and contain noise after feature change are used as target domain samples for recognition and classification; and the adaptive capacity of the model under the complex change environment is verified.The application not only improves the recognition accuracy of the microseismic signals and the self-adaptive capacity of the model, but also optimizes the multi-label classification performance, improves the calculation efficiency and resource utilization rate, and uses comprehensive evaluation indexes to verify the performance; and the application has important application value and popularization prospect in the field of geological disaster monitoring and early warning.

Description

Technical Field

[0001] This invention belongs to the field of intelligent monitoring and information technology, and relates to an adaptive classification method for the dynamic changes of microseismic signal characteristics. Background Technology

[0002] Microseismic monitoring technology, as an important tool for monitoring and early warning of geological disasters, has been widely applied in various fields. In underground engineering, microseismic detection technology is mainly used for early warning of rockburst and collapse risks; in mining, it is often used to predict disasters such as gas outbursts and mine water inrushes; and in reservoir and dam engineering, it is used to monitor leakage behavior and structural deformation. By monitoring and processing microseismic signals in real time, potential engineering risks can be identified in advance, allowing for preventative measures to be taken before disasters occur. However, signal identification, as the first step in microseismic monitoring data processing, is crucial for assessing key parameters of disasters. Changes in environmental conditions (such as temperature, humidity, construction disturbances, and interference from strong noise backgrounds) can significantly alter microseismic waveform characteristics, leading to insufficient adaptability of traditional microseismic signal identification methods in long-term monitoring and complex environments. This not only reduces the accuracy of signal identification but also directly affects the reliability and timeliness of disaster early warning systems. To address waveform characteristic changes caused by environmental variations and improve the accuracy and stability of microseismic signal identification, the research and application of adaptive microseismic signal identification technology is particularly important.

[0003] Significant progress has been made in previous research on microseismic signal identification using various methods. Early studies mainly fell into two categories: waveform spectrum analysis and statistical analysis methods. Arrowsmith et al. (2006), Derr (1970), and Hedlin et al. (1990) primarily used waveform spectrum analysis to identify microseismic signals; Booker and Mitronovas (1964), Taylor (1996), and Wüster (1993) tended to use statistical analysis methods. However, these methods typically rely on specific features extracted from the overall signal, leading to the loss of some signal information. Furthermore, they are highly dependent on the number of sensors and specialized knowledge, making them difficult for field personnel to operate efficiently. With the development of computer technology, artificial intelligence algorithms have attracted widespread attention from scholars due to their advantages in efficiency and automation. Vallejos and McKinnon (2013) used Logistic Regression and Artificial Neural Networks (ANNs) to distinguish between blasting and microseismic signals; Zhang et al. (2019a) decomposed microseismic events and blasting signals, obtained feature vectors, and then used an Extreme Learning Machine (ELM) model for signal recognition; Fan et al. (2022) developed an artificial intelligence recognition model using WSD transform and SVM, providing a fast and highly accurate method for classification. Although these methods have promoted the development of microseismic signal recognition, their performance heavily depends on the effectiveness of feature extraction, and the feature extraction process usually requires manual design. To overcome this limitation, Bi et al. (2021) proposed an understandable time-frequency convolutional neural network (XTF-CNN) for automatic identification of rock fracture and noise signals, achieving an accuracy of 95.17%. Li et al. (2022) used VGG16, ResNet18, AlexNet, and ensemble models to automatically identify and classify current, blasting, microseismic, noise, and siren noise signals, achieving an accuracy of up to 98%. Ma et al. (2023) proposed a microseismic signal identification model based on short-time Fourier transform (STFT) and deep learning techniques for classifying microseismic, blasting, and Gaussian environmental noise, achieving a microseismic event identification accuracy of 97%. These methods can automatically extract features from microseismic signals, thus significantly improving the accuracy and efficiency of identification.

[0004] Despite significant achievements in microseismic signal identification, complex and variable environmental factors (such as high temperature, high humidity, high ground stress, and construction methods) severely impact signal identification accuracy in deep engineering construction environments. Especially during long-term monitoring, changes in environmental conditions lead to significant alterations in signal waveform characteristics. Traditional classification models, lacking adaptability, struggle to adjust feature extraction methods promptly, resulting in a significant decrease in classification accuracy. For example, Wangand Tang (2022) found that different excavation methods in the same project severely affected the reliability of the ResSCA model's signal identification, reducing its accuracy from 97% to 80.6%. Therefore, there is an urgent need to explore an adaptive microseismic signal identification method capable of accurately extracting effective signal features and achieving high-precision classification under changing environments. Furthermore, processing large-scale sample data typically requires substantial time and human resources. Even with large datasets, traditional computing platforms still face excessive computational resource demands when processing deep learning models. For instance, Donget et al. (2023) pointed out that deep convolutional neural networks typically require more time and computational resources compared to traditional classification methods. This further highlights the importance of developing efficient and accurate adaptive signal recognition methods under limited sample conditions, in order to effectively alleviate the dual pressure of time cost and computing resources, and improve the applicability of the model in practical engineering.

[0005] Effective identification of microseismic signals is a fundamental and practical engineering problem, and many scholars have proposed different research solutions. With the rapid development of deep neural network architectures, ResNet, with its skip connection mechanism, significantly alleviates the gradient vanishing problem in deep networks, enabling the network to learn transformations superimposed on the input. It has achieved breakthrough results in image recognition and signal detection. Its deep feature extraction capability can identify subtle changes in signal patterns. Chen et al. (2021) and Souza et al. (2023) proposed its application for microseismic signal detection and classification in complex noisy environments. Recurrent Neural Networks (RNNs), due to their recurrent structure, can maintain information memory and capture the dynamic changes in sequence data, thus accurately extracting the temporal features of microseismic signals in signal detection tasks and enhancing the real-time detection capability of microseismic events. The Transformer model, proposed by Vaswani et al. (2017), effectively solves the long-distance dependency problem in sequence data through a self-attention mechanism, and significantly improves training efficiency because it can process the entire sequence in parallel. The Visual Transformer (VIT) is based on the Transformer architecture. It segments images into multiple patches and processes these patches as a sequence, effectively capturing the global features and temporal dependencies of seismic signals using a self-attention mechanism. Furthermore, for complex signals with extremely low signal-to-noise ratios, VIT, through its self-attention mechanism, can extract more refined features, achieving accurate identification of effective signals and noise in low-SNR microseismic signals, thereby improving detection accuracy.

[0006] However, noise in actual microseismic signal analysis often contains multiple different categories of features. Although the above-mentioned architecture has achieved significant success in single-label classification tasks, in multi-scene microseismic signal processing, multiple noise types superimposed on the weak effective signal greatly increase the difficulty of noise and signal identification. Furthermore, the existence of multiple noise categories further increases the complexity of distinguishing different noise types. To address the semantic redundancy and co-occurrence dependency issues in multi-label classification, Wang et al. (2016) proposed an architecture combining convolutional neural networks (CNNs) and recurrent neural networks (RNNs) in 2016. This architecture utilizes an encoding attention mechanism to extract image features based on previous predictions. Given the weak co-occurrence dependency between noise types in microseismic signals, the CNN-RNN architecture was modified to achieve adaptive identification and classification of noise types in microseismic signals. Specifically, firstly, CNNs are used to extract features from the input signal, then fully connected layers (FCs) are used to process the extracted features, followed by temporal modeling using an RNN network, and finally, multi-label classification is performed using a Softmax layer. A detailed description of this model architecture is as follows:Figure 1 As shown, the CNN-RNN hybrid model architecture exhibits significant performance advantages in multi-class classification problems. However, its adaptive recognition capability is limited when facing scenarios with unknown noise categories or complex combinations of noise. Summary of the Invention

[0007] To address the aforementioned technical problems in existing technologies, this invention provides an adaptive classification method for dynamically changing microseismic signal features. It proposes an Ada-VIT model suitable for adaptive classification of low signal-to-noise ratio microseismic signals in deep, complex environments. This model consists of a feature extractor, a multi-label classifier, and a domain discriminator. By introducing an adaptive mechanism, it dynamically optimizes signal features, improving the recognition accuracy of microseismic signals in complex and changing environments. Clean microseismic signals and pure noise signals collected from the Hanjiang-to-Weihe River Water Diversion Project are used as source domain samples for model training, while microseismic signals with changed features are used as target domain samples for detection and classification. The superiority of the Ada-VIT model is verified through comparison with ResNet and CNN-RNN models, and further applied to the Jinping Phase II Hydropower Station microseismic dataset.

[0008] To achieve the above objectives, the present invention provides the following technical solution:

[0009] An adaptive classification method for dynamically changing microseismic signal characteristics, comprising the following steps:

[0010] S1. Construct the Ada-VIT model, which integrates a feature extractor, a multi-label classifier, and a domain discriminator;

[0011] S2. Utilize an adaptive mechanism to optimize features in order to improve the identification accuracy of microseismic signals under complex and changing environments;

[0012] S3. Use clean microseismic signals and pure noise signals as source domain samples for model training, and use microseismic signals with more complex and noisy features after feature changes as target domain samples for identification and classification.

[0013] S4. Verify the model's adaptive capability in complex and changing environments.

[0014] Furthermore, in step S1, a CNN-RNN model architecture is used to identify noise, treating the single-noise microseismic signal dataset and the multi-noise microseismic signal dataset as two different domains for processing. The source domain dataset covers the following signal types: clean microseismic signals, current noise signals, environmental noise signals, current and environmental noise signals, blasting noise signals, and siren noise signals. Each signal type is treated as an independent category label, and each sample is assigned a binary value based on whether it contains a signal of a specific category: if the sample contains a signal of that category, it is labeled as 1; otherwise, it is labeled as 0. The label i is represented by a one-hot vector e. i = [0,...,1,0,...,0] where the i-th position is 1 and the others are 0. The multi-label method involves creating a one-hot vector e. i and label matrix U l Multiplying them together yields the label vector y. i It can be represented as:

[0015] y i =U l ·e i (1)

[0016] The target domain dataset contains microseismic signal samples under noise interference. The labeling process for the target domain dataset follows the same annotation method as the source domain dataset.

[0017] Furthermore, in step S1, the feature extractor uses a pre-trained VIT model. This architecture uses a global self-attention mechanism to accurately learn the interrelationships between small blocks in the image and build long-range dependencies between any two small blocks, so as to have the ability to accurately capture the global features of microseismic signals and efficiently distinguish microseismic signals from noise.

[0018] The multi-label classifier receives features learned by the feature extractor and uses these features to effectively identify signals in the source and target domains, while also classifying the noise carried by the samples. This process is achieved through two fully connected layers, which map high-dimensional features to a low-dimensional space and identify the category to which the samples belong.

[0019] The domain discriminator receives the feature vectors generated by the feature extractor and performs binary classification decisions using the sigmoid activation function to determine the source of the feature vectors. By introducing an adversarial training mechanism for the domain discriminator, the feature extractor is guided to learn that the domain discriminator cannot accurately distinguish the feature representations of the source and target domains, thereby establishing a higher degree of feature similarity between the source and target domains.

[0020] Furthermore, in step S1, the input to the model is a signal image, which is cropped and transformed to a uniform size of 224×224 pixels. Both the source domain data and the target domain data are input to the feature extractor after this preprocessing step. The VIT network divides the input image into 16×16 pixel blocks, totaling (224 / 16)*(224 / 16) = 196 blocks. Each block is mapped to a fixed-dimensional vector space through a linear layer. Subsequently, a positional embedding layer is used to add positional encoding to each image block and the classification label to preserve the positional information in the sequence. The classification label is used to aggregate global image information and is represented by an asterisk (*) in the figure. After these preprocessing steps, the blocks and CLS labels are input together to the Transformer encoder to capture a more refined feature representation.

[0021] Furthermore, for a single sample, when facing a multi-label problem, the modeling method described above is extended to perform independent binary classification for the prediction of each label, and the final loss is the sum of the binary classification losses for each label; the cross-entropy loss function of the multi-label classifier is calculated as follows:

[0022]

[0023] Where y i For the actual label vector, Let be the predicted probability of the model for the i-th class, i.e., the probability generated by the Softmax function; its calculation formula is:

[0024]

[0025] The binary cross-entropy loss function used by the domain discriminator is calculated as follows:

[0026]

[0027] Where y represents the actual source and target domain labels. This represents the label predicted by the discriminator; the total loss of the Ada-VIT model is:

[0028] Loss all =αLoss c +βLoss d (5)

[0029] Experiments have shown that the model can achieve stable performance when α is 1 and β is 0.5.

[0030] Furthermore, in step S3, 6970 sets of engineering monitoring data are collected, specifically distributed as follows: 1001 sets of standard microseismic data, 1001 sets of current microseismic data, 1001 sets of noise microseismic data, 1001 sets of current-noise microseismic data, 501 sets of pure current data, 967 sets of pure noise data, 501 sets of noise-current data, 500 sets of blasting data, and 497 sets of car horn data. The collected source domain dataset will be used to build a training model, while the target domain dataset is further divided into a validation set and a test set.

[0031] During data partitioning, the ratio of training set, validation set, and test set was 7:2:1. The Ada-VIT model training process used Adam as the optimizer to achieve rapid convergence. The learning rates for the feature extractor and classifier were set to 0.00001, and the learning rate for the domain discriminator was also set to 0.0001. Training was iterated 500 times. Signal features were effectively extracted using the convolution and attention mechanisms in the transformer. t-SNE visualization was performed on the signal features of different categories in the source and target domain datasets. This method maps high-dimensional features to a two-dimensional space, thus visually displaying the internal structure of high-dimensional data and revealing its potential distribution relationships.

[0032] After processing by the feature extractor in the Ada-VIT model, different types of noise features of the signal are differentially identified. At the same time, the feature distribution patterns of different distribution domains are learned. Through adversarial training of the domain discriminator, the distribution difference between the source domain and the target domain becomes smaller and smaller, thereby increasing the model's domain generalization ability.

[0033] The VIT network architecture has a total of 12 layers. When calling the pre-trained model, its pre-trained weight parameters are retained. First, all network layers are unfrozen for training. Comparative experiments were conducted on different numbers of frozen layers. The last two layers of the unfrozen VIT model were selected to train the dataset for feature extraction.

[0034] Furthermore, in step S4, accuracy, precision, recall, F1 score, and AUC are used for comprehensive evaluation. Accuracy measures the overall accuracy of the model's prediction results, i.e., the proportion of samples correctly predicted by the model out of the total samples. Precision evaluates the model's accuracy in predicting positive samples, i.e., the proportion of samples predicted as positive that are actually positive. Recall reflects the model's recall ability for positive samples, i.e., the proportion of samples that are actually positive that are correctly identified by the model. The F1 score takes into account both precision and recall, and its value reflects the balance between the model's identification of positive and negative samples during the classification process. The AUC metric measures the robustness of the model to changes in different classification thresholds by the area under the ROC curve. The higher the AUC value, the stronger the model's ability to distinguish between positive and negative samples under different thresholds.

[0035] Furthermore, the calculation formulas for the evaluation indicators are as follows: (6) to (11):

[0036]

[0037] Among them, TP represents true positive; TN represents true negative; FP represents false positive; FN represents false negative; and FPR is the false positive rate.

[0038] Furthermore, the mean average precision (mAP) is calculated to comprehensively evaluate the overall performance of the Ada-VIT model in multi-label classification tasks. The calculation formulas are as follows (12) to (13):

[0039] For a single sample i, its true label set is Y. i The model's prediction score for all labels is f. i Sort all labels according to their scores to obtain the sorted label sequence R. i The formula for calculating average accuracy is:

[0040]

[0041] Where n represents the total number of tags; P i (k) represents the proportion of correctly predicted labels out of the first k predictions, Rk i (k) represents the k-th label after sorting, δ(R) i (k)∈Y i For indicator functions, when R i (k) belongs to Y i The value is 1 if the condition is met, and 0 otherwise; the formula for calculating mAP is:

[0042]

[0043] Where N represents the total number of samples, AP i This represents the average accuracy of the i-th sample.

[0044] The beneficial effects of this invention are:

[0045] Compared with existing technologies, the adaptive classification method for dynamic changes in microseismic signal characteristics described in this invention has the following technical features or beneficial effects:

[0046] (1) Improved accuracy of microseismic signal identification: By constructing an Ada-VIT model that integrates a feature extractor, a multi-label classifier, and a domain discriminator, this invention can accurately capture the global features of microseismic signals and effectively distinguish microseismic signals from noise. Even in complex and variable deep engineering environments, it can significantly improve the identification accuracy of microseismic signals. This is due to the adaptive mechanism in the Ada-VIT model, which can dynamically optimize signal features to adapt to signal changes in different environments.

[0047] (2) Enhancing Model Adaptability: This invention introduces a domain discriminator and an adversarial training mechanism, prompting the feature extractor to learn feature representations that the domain discriminator cannot accurately distinguish between the source and target domains. This design enables the Ada-VIT model to establish a higher degree of feature similarity between the source and target domains, thereby enhancing the model's adaptability under different environments. This is of great significance for long-term monitoring and early warning of geological disasters, as changes in environmental conditions can significantly affect the characteristics of microseismic waveforms.

[0048] (3) Optimizing Multi-Label Classification Performance: To address the various noise types that may exist in microseismic signals, this invention employs a multi-label classifier. By assigning an independent category label to each noise type and annotating it with binary values, the Ada-VIT model can simultaneously identify and classify multiple noises in microseismic signals. This multi-label classification method improves the accuracy and robustness of the model when processing complex microseismic signals.

[0049] (4) Improved computational efficiency and resource utilization: Although deep learning models typically require significant time and computational resources when processing large-scale sample data, this invention effectively improves computational efficiency by using a pre-trained VIT model as a feature extractor and unfreezing the last two layers for training. Furthermore, by optimizing parameter settings such as the learning rate, the convergence speed and resource utilization of the model are further enhanced.

[0050] (5) Comprehensive Evaluation Metrics: This invention employs multiple metrics, including accuracy, precision, recall, F1 score, and AUC, to comprehensively evaluate model performance. These metrics comprehensively reflect the model's performance in various aspects, including overall prediction accuracy, positive sample prediction precision, positive sample recall, positive and negative sample recognition balance, and robustness to different classification thresholds. Furthermore, mAP is calculated to comprehensively evaluate the overall performance of the Ada-VIT model in multi-label classification tasks, further demonstrating the model's superiority and practicality.

[0051] (6) Broad Engineering Application Prospects: The Ada-VIT model of this invention has been applied and verified in actual projects such as the Hanjiang-to-Weihe River Water Diversion Project and the Jinping Phase II Hydropower Station, achieving excellent results. This indicates that the model has broad engineering application prospects, especially in the field of geological disaster monitoring and early warning. By real-time monitoring and processing of microseismic signals, potential engineering risks can be identified in advance, and preventive measures can be taken, which can significantly reduce the probability and severity of geological disasters.

[0052] This invention proposes an adaptive classification method based on the dynamic changes of microseismic signal characteristics and its Ada-VIT model. This method not only improves the identification accuracy and adaptive capability of microseismic signals, but also optimizes multi-label classification performance, improves computational efficiency and resource utilization, and uses comprehensive evaluation indicators for performance verification. These beneficial effects make this invention of significant application value and promising prospects in the field of geological disaster monitoring and early warning. Attached Figure Description

[0053] To more clearly illustrate the technical solutions of the embodiments of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and detailed embodiments. 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.

[0054] in:

[0055] Figure 1 This is a diagram of the CNN-RNN network model architecture;

[0056] Figure 2 This is the Ada-VIT model architecture diagram;

[0057] Figure 3 The geographical location of the Hanjiang-to-Weihe River Water Diversion Project;

[0058] Figure 4The datasets are divided into source and target domains, and their features are visualized as follows: a) Visualization of the t-SNE feature distribution of the source and target domain datasets; b) Visualization of signal features extracted during the Ada-VIT model training process; c) Visualization of the t-SNE features of each category of signals in the source domain dataset; d) Visualization of the t-SNE features of each category of signals in the target domain dataset.

[0059] Figure 5 This is a test result diagram of the Ada-VIT model;

[0060] Figure 6 The following is a visualization of the domain adaptation effect of the Ada-VIT model using t-SNE (preferably viewed in color): (a) the effect without Ada-VIT activation; (b) the effect after domain adversarial training of the Ada-VIT model; (c) the distribution effect of each category of signal after domain adversarial training of the Ada-VIT model.

[0061] Figure 7 This is a comparison chart of the test results of different models;

[0062] Figure 8 These are the F1 scores for each model;

[0063] Figure 9 The geographical location of Jinping Mountain;

[0064] Figure 10 This is a t-SNE visualization of the Jinping Phase II Hydropower Station dataset;

[0065] Figure 11 This is a diagram showing the application results;

[0066] Figure 12 This is a graph showing the model performance evaluation metrics. Detailed Implementation

[0067] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention. The following description, in conjunction with the accompanying drawings... Figure 1-12 The adaptive classification method for the dynamic changes of microseismic signal characteristics is further explained, and the technical solutions in the embodiments of this application are clearly and completely described. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0068] The construction environment of deep underground engineering is complex and variable, and the resulting interference signals with different characteristics reduce the long-term accuracy of microseismic signal identification. This invention proposes an adaptive microseismic signal identification method for environmental changes. This method constructs an Ada-VIT model that integrates feature extraction, multi-label classification, and domain adaptation capabilities. The adaptive mechanism optimizes features to improve the identification accuracy of microseismic signals under complex and changing environments. Clean microseismic signals and pure noise signals collected from the Hanjiang-to-Weihe River Water Diversion Project are used as source domain samples for model training. More complex and noisy microseismic signals with changed features are used as target domain samples for identification and classification, verifying the model's adaptive capability under complex and changing environments. The results show that the Ada-VIT model significantly outperforms ResNet50 and the improved CNN-RNN model in microseismic signal identification and classification, with a mean average accuracy (mAP) as high as 0.98. Furthermore, the adaptive classification model is applied to the Jinping Phase II Hydropower Station, and the model demonstrates excellent adaptive classification performance in actual engineering. This invention provides important theoretical support and technical assurance for long-term identification and accurate early warning of microseismic signals in deep engineering.

[0069] Example 1

[0070] Project Modeling: Accurate extraction of effective signal features is a highly challenging task when signal-to-noise ratios are low and sample sizes are limited. Traditional convolutional neural networks (CNNs) are insufficient to meet this challenge, while ResNet networks with multiple layers suffer from poor training performance and excessive time consumption due to insufficient sample sizes. Given the sequential characteristics of microseismic signals, this invention considers using a CNN-RNN model architecture to identify noise. However, for microseismic signals with multiple superimposed noise types, the feature distribution changes compared to microseismic signals with a single noise type; it is no longer a simple linear superposition but a fundamental shift in the distribution domain. Therefore, this invention proposes treating single-noise microseismic signal datasets and multi-noise microseismic signal datasets as two different domains. The source domain datasets cover the following signal types: clean microseismic signals, current noise signals, environmental noise signals, current and environmental noise signals, blasting noise signals, and siren noise signals. Each signal type is treated as an independent category label, and each sample is assigned a binary value based on whether it contains a signal of a specific category: if the sample contains a signal of that category, it is marked as 1; otherwise, it is marked as 0. Label i is represented as a one-hot vector e i = [0,...,1,0,...,0] where the i-th position is 1 and the others are 0. Multi-labeling can be achieved by converting the one-hot vector e... i and label matrix U l Multiplying them together yields the label vector y. i It can be represented as:

[0071] y i =U l ·e i (1)

[0072] The target domain dataset contains microseismic signal samples under noise interference. The labeling process for the target domain dataset follows the same annotation method as the source domain dataset.

[0073] Proposed Method: To address the problem of effectively identifying various noise types in complex noise superposition scenarios, this invention proposes an adaptive microseismic signal noise classification Ada-VIT network model. This model can not only adaptively identify unknown noise types, but also effectively detect multiple superimposed noise types. A schematic diagram of the model is shown below. Figure 2 As shown.

[0074] The proposed adaptive microseismic signal noise classification Ada-VIT network model consists of a feature extractor, a multi-label classifier, and a domain discriminator. The feature extractor employs a pre-trained VIT model. This architecture leverages a global self-attention mechanism to accurately learn the interrelationships between small patches in the image, establishing long-range dependencies between any two patches. This enables it to accurately capture the global features of microseismic signals and efficiently distinguish them from noise. For small-sample microseismic datasets, given that the VIT model has already been trained on massive image datasets and mastered general visual features, fine-tuning not only saves training time but also achieves good classification results.

[0075] The multi-label classifier receives features learned by the feature extractor and uses these features to effectively identify signals in the source and target domains, while also classifying the noise carried by the samples. This process is mainly achieved through two fully connected (FC) layers, which map high-dimensional features to a low-dimensional space and identify the category to which the sample belongs accordingly.

[0076] The domain discriminator receives feature vectors generated by the feature extractor and performs binary classification using the sigmoid activation function to determine the source (target domain or source domain) of the feature vectors. By introducing an adversarial training mechanism for the domain discriminator, the feature extractor is guided to learn to prevent the domain discriminator from accurately distinguishing the feature representations of the source and target domains, thereby establishing a higher degree of feature similarity between the two domains. This process significantly reduces the distributional differences between the source and target domains, effectively enhancing the model's generalization ability in the target domain. In this way, the model can better adapt to the data distribution of the target domain, improving the overall performance of cross-domain transfer learning.

[0077] The model's input is a signal image, which, after cropping and transformation, is uniformly resized to 224×224 pixels. Both source and target domain data undergo this preprocessing step before being input to the feature extractor. The VIT network divides the input image into 16×16 pixel blocks, totaling (224 / 16)*(224 / 16) = 196 blocks. Each block is mapped to a fixed-dimensional vector space through a linear layer. Subsequently, a positional encoding is added to each image block and a specially designed classification label (CLS label) through a position embedding layer to preserve positional information in the sequence. The classification label (CLS label) is used to aggregate global image information and is represented by an asterisk (*) in the diagram. After these preprocessing steps, the blocks and CLS labels are input together to the Transformer encoder to capture a more refined feature representation.

[0078] For a single sample, when facing a multi-label problem, the modeling method described above is extended to perform independent binary classification for each label prediction, and the final loss is the sum of the binary classification losses for each label. The formula for calculating the cross-entropy loss function of the multi-label classifier is:

[0079]

[0080] Where y i For the actual label vector, Let be the predicted probability of the model for the i-th class, i.e., the probability generated by the Softmax function. Its calculation formula is:

[0081]

[0082] The binary cross-entropy loss function used by the domain discriminator is calculated as follows:

[0083]

[0084] Where y represents the actual source and target domain labels. This represents the label predicted by the discriminator. The total loss of the Ada-VIT model is:

[0085] Loss all =αLoss c +βLoss d (5)

[0086] Experiments have shown that the model can achieve stable performance when α is 1 and β is 0.5.

[0087] All computation and neural network training were performed on a Windows PC equipped with the following hardware: CPU: Intel Core i5-14600KF; RAM: 32GB; GPU: NVIDIA RTX 4070Ti. All deep learning models and training were developed and executed on the open-source deep learning framework PyTorch.

[0088] Example 2

[0089] This embodiment provides an introduction to the engineering background and dataset of the Hanjiang-to-Weihe River Water Diversion Project in Shaanxi, China.

[0090] The dataset for this invention originates from the Hanjiang-to-Weihe River Water Diversion Project in Shaanxi, China, such as... Figure 3 This project spans the Yangtze and Yellow River basins and traverses the Qinling Mountains. It mainly consists of three parts: the Qinling Water Diversion Tunnel, the Huangjinxia Hydropower Project, and the Sanhekou Hydropower Project. The Qinling Water Diversion Tunnel, the core and key control project, is 98.3 kilometers long with a maximum burial depth exceeding 2000 meters. Due to the complex geological conditions in the construction area, including high ground stress, high seepage, and high temperatures, geological disasters such as rock bursts and water inrushes are highly likely to occur, posing a significant threat to construction safety and equipment operation. To reduce safety risks during construction, a microseismic monitoring system was introduced. This system includes a mainframe analysis module, a data acquisition instrument, and a single-axis accelerometer. The accelerometer's design parameters include a response frequency range of 50Hz to 5kHz, a sensitivity of 30V / g, and dimensions of 25.4mm in diameter and 122mm in height. The system sampling frequency is set to 10kHz, with each acquisition lasting 400ms.

[0091] Because microseismic signals with multiple types of noise superimposed exhibit significant changes in their characteristic distribution compared to microseismic signals with a single type of noise, this change is not a simple accumulation of features, but rather causes a transformation in the characteristic distribution domain, such as... Figure 4 As shown in Figure a, the blue dots represent source domain data features, and the red dots represent target domain data features. As can be seen from the figure, there are large inter-class differences between the data features of the source domain and the target domain.

[0092] This study collected 6970 sets of monitoring data from the Hanjiang-to-Weihe River Water Diversion Project, distributed as follows: 1001 sets of standard microseismic data, 1001 sets of current-based microseismic data, 1001 sets of noise-based microseismic data, 1001 sets of current-noise-based microseismic data, 501 sets of pure current data, 967 sets of pure noise data, 501 sets of noise-current data, 500 sets of blasting data, and 497 sets of car horn data. The collected source domain dataset was used to build the training model, while the target domain dataset was further divided into validation and test sets. The data partitioning followed a 7:2:1 ratio for the training, validation, and test sets. The Ada-VIT model was trained using Adam as the optimizer, which enabled rapid convergence. Experiments showed that setting the learning rate of the feature extractor and classifier to 0.00001, and the learning rate of the domain discriminator to 0.0001, resulted in a more stable model with the highest classification accuracy, after 500 iterations of training. The signal features extracted during the training process using the Ada-VIT model are visualized as follows: Figure 4 As shown in b, signal features can be effectively extracted through the convolution and attention mechanisms in the transformer. Figure 4 As shown in Figure cd, this invention visualizes the signal features of different categories in the source and target domain datasets using t-SNE. This method maps high-dimensional features to a two-dimensional space, thereby intuitively displaying the internal structure of high-dimensional data and revealing its potential distribution relationships. Different colors are used in the figure to distinguish features of different label categories. It can be observed that the data distribution in the source domain has significant separation, but some categories exhibit some overlap and dispersion.

[0093] In the high-dimensional feature space, the sample feature points of microseismic signals, blasting signals, and car horn signals exhibit high clustering, reflecting significant consistency in features among samples within these categories. Conversely, the sample feature points of noise, current-noise, and current signals are more spatially dispersed, with some even deviating from the main cluster centers. This indicates significant differences in features among samples within these categories, suggesting that some samples are affected by complex external noise. Furthermore, microseismic signals and blasting signals show clear spatial intervals with other signal categories, indicating high efficiency of the model in distinguishing these signal categories. However, current signals and noise current signals almost completely overlap in the feature space, and noise signals and car horn signals also have partial overlap, revealing a certain similarity in features among these signal categories in the high-dimensional feature space, thus increasing the complexity of model recognition. In the target domain, the sample feature points of current microseismic signals, noise microseismic signals, and current-noise microseismic signals also exhibit high discreteness and overlap with each other. In particular, noise microseismic signals and noise-current microseismic signals almost completely overlap, further indicating that these signal categories have high similarity in feature distribution, making accurate differentiation between them extremely difficult. After processing by the feature extractor in the Ada-VIT model designed in this paper, different types of noise features of the signal can be differentially identified. Simultaneously, it learns the feature distribution patterns of different distribution domains. Through adversarial training of the domain discriminator, the distribution differences between the source and target domains gradually decrease, thereby increasing the model's domain generalization ability.

[0094] The VIT network architecture has 12 layers. When calling the pre-trained model, its pre-trained weight parameters are retained. First, all network layers are unfrozen for training. To balance model performance, computational cost, and training time, comparative experiments were conducted with different numbers of layers frozen. The experimental results are shown in Table 1 below. Finally, the last two layers of the VIT model were unfrozen to train the dataset for feature extraction. This reduced training time and saved memory while achieving good training results.

[0095] Table 1. Relationship between the number of unfrozen layers in the VIT pre-trained model and the performance of the Ada-VIT model.

[0096]

[0097]

[0098] Example 3

[0099] This embodiment provides model evaluation metrics.

[0100] In model evaluation, this invention uses indicators such as accuracy, precision, recall, F1 score, and AUC (Area Under the ROC Curve) for comprehensive evaluation. Among them, accuracy is used to measure the overall accuracy of the model's prediction results, that is, the proportion of samples correctly predicted by the model to the total number of samples; precision focuses on evaluating the model's accuracy in predicting positive samples, that is, the proportion of samples predicted as positive that are actually positive; recall reflects the model's recall ability for positive samples, that is, the proportion of samples that are actually positive that are correctly identified by the model. F1 score comprehensively considers precision and recall, and its value reflects the balance of the model in identifying positive and negative samples during the classification process. The AUC indicator measures the robustness of the model to changes in different classification thresholds by the area under the ROC curve. The higher the AUC value, the stronger the model's ability to distinguish between positive and negative samples under different thresholds. Its calculation formula is shown in equations (6) to (11). In addition, this invention also calculates mAP to comprehensively evaluate the overall performance of the Ada-VIT model in multi-label classification tasks. Its calculation formula is shown in equations (12) to (13).

[0101]

[0102] Among them, TP (True Positive) represents a true positive; TN (True Negative) represents a true negative; FP (False Positive) represents a false positive; FN (Falase Negative) represents a false negative; and FPR (False Positive Rate) is the false positive rate.

[0103] For a single sample i, its true label set is Y. i The model's prediction score for all labels is f. i Sort all labels according to their scores to obtain the sorted label sequence R. i The formula for calculating average accuracy is:

[0104]

[0105] Where n represents the total number of tags. P i (k) represents the proportion of correctly predicted labels out of the first k predictions, Rk i (k) represents the k-th label after sorting, δ(R) i (k)∈Y i As an indicator function, when R i (k) belongs to Y i The value is 1 if the condition is met, and 0 otherwise. The formula for calculating mAP is:

[0106]

[0107] Where N represents the total number of samples, AP i This represents the average accuracy of the i-th sample.

[0108] To better evaluate the performance of the Ada-VIT model during training, the performance on the test set during training was visualized (e.g., ...). Figure 5 As shown), from Figure 5 As shown in Figure a, with the increase of iterations, the model's validation loss continuously decreases until it approaches 0.05, while the accuracy gradually increases and stabilizes at 0.91. Because this paper transforms the multi-label classification problem into multiple binary classification problems, precision and recall are also used to comprehensively evaluate the model's performance. Figure 5 As shown in b, in the initial stage of training, precision and recall both increase significantly, indicating that the model can quickly learn the patterns in the data. As training progresses, precision and recall tend to stabilize, and when the model training reaches a stable state, their values ​​reach 91.45% and 92.17%, respectively.

[0109] Another key aspect of this work is the use of known signals and different types of noise for training. Through feature learning and domain adversarial training, the model aims to identify and classify different noise types in multi-noise superimposed microseismic signals. To this end, the domain adaptation performance of the Ada-VIT model is analyzed, and the results are as follows: Figure 6 As shown: Figure 6 a shows the initial distribution of source and target domain features after passing through the feature extractor only. Figure 6 Figure b shows a comparison of feature distributions after introducing the Ada-VIT model and performing domain adversarial training. It can be observed that through the optimization of the domain adversarial mechanism, the difference in feature distributions between the source and target domains is significantly reduced, and the features of the two domains gradually form significant overlap in multiple regions, indicating that the similarity of features between the domains has been effectively improved, thus achieving the domain adaptation goal. Although some clusters still mainly consist of single-domain data points, further visualization of specific categories reveals that the inter-domain differences between source domain data (such as current signals and current-noise signals) and target domain data (such as current-noise-microseismic signals) are significantly reduced and tend to converge into the same cluster. Furthermore, the inter-domain differences between the source domain current signals and the target domain current-noise-microseismic signals and current-microseismic signals are also reduced synchronously, while similar domain alignment patterns are also observed among the three types of features: noise signals, microseismic signals, and noise-microseismic signals.

[0110] Interestingly, the noise-microseismic signal, possessing characteristics of both noise and microseismic signals, acts as a bridge connecting the two, aggregating them into the same cluster after domain adversarial training. It is also noteworthy that even after domain adversarial training, although the inter-domain distance is significantly reduced and similar features gradually cluster together, the distribution differences between noise and effective microseismic signals remain quite significant, and clear distinctions still exist between different types of signals.

[0111] This paper compares the classification performance of three models: the classic ResNet50 classification network, the improved multi-label classification network CNN-RNN, and the proposed multi-label adaptive classification network Ada-VIT. The results are as follows: Figure 7 As shown in the comparative analysis, ResNet50 performs poorly in multi-label classification, while CNN-RNN shows a significant improvement over ResNet50, achieving an average accuracy (mAP) of around 0.8 for the microseismic dataset used in this paper. However, the performance still needs further improvement. This is because although the CNN-RNN model increases the richness of extracted features and can further capture the sequence features of seismic signals, it cannot accurately distinguish the noise mixed in with microseismic signals due to the superposition of different types of noise causing changes in their distribution domain, or for small samples of low signal-to-noise ratio microseismic signals. The Ada-VIT model, through its VIT-based feature extractor and domain discriminator, effectively solves this problem, achieving effective identification of low signal-to-noise ratio microseismic signals and accurate classification of multiple superimposed noise types, with an mAP as high as 0.98.

[0112] Table 2 shows the parameter count, validation loss, and average validation accuracy of the three models. The table reveals that Ada-VIT has the most parameters, 3.6 times that of ResNet50 and 134.8 times that of CNN-RNN. This indicates that Ada-VIT has higher model complexity and learns richer features, but this also increases computational complexity and GPU resource requirements, leading to longer training times. To balance computational cost and training time, the parameters and weights of the first 10 layers were frozen during VIT model training, while the last two convolutional and Transformer layers were unfrozen for feature extraction. Experiments show that Ada-VIT has the shortest average training time per run and the lowest validation loss (0.24), indicating that it can learn quickly, converge, and achieve stable and accurate classification performance.

[0113] Table 2 Comparison of parameter count and training performance of different models

[0114]

[0115] As shown in Table 3, the performance of the three models was tested, including mean average precision (mAP), accuracy (Acc), and testing time for each test sample. The test results show that although the Ada-VIT model takes 0.0222 seconds and 0.0484 seconds longer to process each sample compared to ResNet50 and CNN-RNN models, respectively, its mAP value reached 0.99, representing improvements of 83.33% and 90.38% compared to ResNet50 and CNN-RNN models, respectively. Furthermore, the test accuracy of the Ada-VIT model reached 278.78% and 184% of that of ResNet50 and CNN-RNN models, respectively. These results collectively validate the significant performance advantage of the Ada-VIT model in multi-label classification.

[0116] Further analysis Figure 8 The F1 score is used to demonstrate the model's performance on different signal types; a higher value indicates better accuracy and comprehensiveness. It's worth noting that due to the lack of samples for blasting and siren noise signals in the test set, all models achieved an F1 score of zero for these two signal types. However, significant differences emerged in the F1 scores for classifying microseismic, current, noise, and noise-current signals. Specifically, the Ada-VIT model achieved F1 scores as high as 1, 0.97, 0.95, and 0.88 for these signal types, demonstrating excellent classification performance across all signal types. In contrast, the ResNet50 model only achieved an F1 score exceeding 0.5 for microseismic signals, while the CNN-RNN model only performed well on microseismic and noise-current signals, with a highest F1 score of only 0.56. In summary, the F1 scores of the Ada-VIT model across different signal types not only highlight its robustness in signal classification tasks, but also further verify its adaptability to different signal characteristics and its superiority in the field of multi-class signal classification.

[0117] Table 3 Comparison of test performance of different models

[0118] ResNet50 0.0627 0.54 0.33 CNN-RNN 0.0365 0.52 0.50 Ada-VIT 0.0849 0.99 0.92

[0119] Example 4

[0120] This embodiment provides an application example of the Jinping Phase II Hydropower Station.

[0121] Jinping Phase II Hydropower Station is located in the middle reaches of the Yalong River in Sichuan Province. Figure 9The project is a key component of China's clean energy development. It primarily comprises an underground powerhouse, water diversion tunnels, and a flood discharge system, with a total installed capacity of 4.8 million kilowatts and an average annual power generation of approximately 16 billion kilowatt-hours. As the world's deepest water diversion tunnel complex, with a maximum tunnel depth of 2,525 meters, the project faced complex geological conditions including high ground stress, high seepage, and high temperatures, making rockburst risk particularly prominent. To ensure construction safety, the project employed an advanced microseismic monitoring system to identify rockburst risks. This system consists of a Hyperion digital signal processing system, a Paladin digital signal acquisition system, and six single-axis accelerometers, whose performance parameters are similar to those of the Hanjiang-to-Weihe River Water Diversion Project.

[0122] To systematically evaluate the performance of the Ada-VIT model in different real-world engineering scenarios, the Jinping Phase II Hydropower Station was selected as a case study, and 1000 sets of data were collected for empirical analysis of the Ada-VIT model. 850 sets of data were randomly selected to form the training set, and the remaining 150 sets were used as the test set. Detailed visualization analysis of the training and test sets was conducted (see...). Figure 10 The results revealed a significant gap between the training and test sets in the feature space. This phenomenon indicates that the feature distributions of the two sets of data differ and have low correlation, revealing that they originate from different distribution domains. Therefore, if the Ada-VIT model can accurately identify the categories of the two types of data, this will strongly demonstrate that the Ada-VIT model has excellent generalization ability and can effectively classify signals across different data distribution domains.

[0123] Figure 11 This paper presents a performance evaluation of the Ada-VIT model for identifying microseismic signals on the Jinping II Hydropower Station dataset. The analysis reveals a high degree of consistency between the number of samples predicted as microseismic signals (502 cases) and the actual number of microseismic signal samples (496 cases). Simultaneously, the difference between the number of samples predicted as noise (505 cases) and the actual number of noise samples (504 cases) is negligible. These small deviations between the predictions and the actual data further confirm its high accuracy and reliability in adaptive multi-classification tasks, highlighting the application potential of the Ada-VIT model in the field of adaptive signal multi-classification.

[0124] To further quantify the application potential of the Ada-VIT model in adaptive signal multi-class classification, a detailed evaluation was conducted on the prediction results of the Jinping Phase II Hydropower Station dataset. Key performance indicators, including Accuracy, Precision, Recall, F1 score, and AUC, were calculated. Figure 12As shown in the figure, the closer the value is to 1, the better its performance. It can be observed from the figure that all performance indicators consistently reach 0.99. These data fully demonstrate that the Ada-VIT model can not only accurately classify microseismic signals and noise under extremely limited sample conditions, but also exhibit excellent signal separation and signal recognition capabilities in engineering environments with strong noise and multiple interferences.

[0125] In long-term monitoring of deep underground engineering projects, although existing models demonstrate high accuracy in identifying microseismic signals, their performance often declines rapidly under dynamic changes in complex environments, making it difficult to maintain stable accuracy. Furthermore, the superimposed interference of multiple noise sources on weak, effective signals further increases the complexity of identification. To address this issue, this invention proposes an adaptive classification Ada-VIT model for low signal-to-noise ratio microseismic signals in deep, changing environments. This model consists of a feature extractor, a multi-label classifier, and a domain discriminator. Through an adaptive mechanism, it optimizes features to accurately identify microseismic signals while simultaneously detecting the types of noise embedded in the measured microseismic signals. This provides a novel solution for identifying low signal-to-noise ratio microseismic signals in complex deep environments.

[0126] Clean microseismic signals and pure noise signals collected from the Hanjiang-to-Weihe River Water Diversion Project were used as source domain signal samples for training, while microseismic signals after environmental changes were used as target domain signal samples for detection and identification. Results show that the Ada-VIT model performs excellently in microseismic signal identification, with an average precision (mAP) of 0.98, significantly outperforming the traditional ResNet50 and improved CNN-RNN models. Furthermore, the Ada-VIT model exhibits fast convergence speed and low loss during training and testing, validating its stability and effectiveness. By freezing and fine-tuning some layers of the VIT model, Ada-VIT achieves accurate classification of small-sample, low-signal-to-noise-ratio datasets while maintaining low computational and memory overhead.

[0127] To evaluate the performance of the Ada-VIT model in identifying and classifying microseismic signals in various engineering projects, it was applied to the signal identification and classification dataset of the Jinping Phase II Hydropower Station. The results show a high degree of consistency between the model's predictions and the actual results. In the classification task, key metrics such as Accuracy, Precision, Recall, F1 score, and AUC all consistently reached 0.99, fully demonstrating the superior classification performance of the Ada-VIT model under limited sample conditions. Furthermore, the model exhibits extremely high signal identification and classification capabilities in complex engineering environments with strong noise and multiple interferences, providing important theoretical support and technical assurance for signal monitoring and accurate early warning in deep engineering projects.

[0128] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An adaptive classification method for dynamically changing microseismic signal characteristics, characterized in that, The steps are as follows: S1. Construct an Ada-VIT model that integrates a feature extractor, a multi-label classifier, and a domain discriminator. The feature extractor uses a pre-trained VIT model. This architecture uses a global self-attention mechanism to accurately learn the interrelationships between small patches in an image and build long-range dependencies between any two patches. This enables it to accurately capture the global features of microseismic signals and efficiently distinguish microseismic signals from noise. The multi-label classifier receives features learned by the feature extractor and uses these features to effectively identify source and target domain signals, while classifying the noise carried by the samples. This process is achieved through two fully connected layers, which map high-dimensional features to a low-dimensional space and identify the category to which the samples belong. The domain discriminator receives the feature vector generated by the feature extractor and performs a binary classification decision through the sigmoid activation function to determine the source of the feature vector. By introducing an adversarial training mechanism for the domain discriminator, the feature extractor is guided to learn that the domain discriminator cannot accurately distinguish the feature representations of the source domain and the target domain, thereby establishing a higher degree of feature similarity between the source domain and the target domain. S2. Utilize an adaptive mechanism to optimize features in order to improve the identification accuracy of microseismic signals under complex and changing environments; S3. Use clean microseismic signals and pure noise signals as source domain samples for model training, and use microseismic signals with more complex and noisy features after feature changes as target domain samples for identification and classification. S4. Verify the model's adaptive capability in complex and changing environments.

2. The adaptive classification method for dynamic changes in microseismic signal characteristics as described in claim 1, characterized in that, In step S1, a CNN-RNN model architecture is used to identify noise. The single-noise microseismic signal dataset and the multi-noise microseismic signal dataset are treated as two different domains for processing. The source domain dataset covers the following signal types: clean microseismic signals, current noise signals, environmental noise signals, current and environmental noise signals, blasting noise signals, and siren noise signals. Each signal type is treated as an independent category label. Each sample is assigned a binary value based on whether it contains a signal of a specific category: if the sample contains a signal of that category, it is labeled 1; otherwise, it is labeled 0. i Represented as a one-hot vector No. i One position is 1, and the rest are 0. Multi-labeling is achieved by using one-hot vectors. and label matrix Multiplying them together yields the label vector. It can be represented as: The target domain dataset contains microseismic signal samples under noise interference. The labeling process for the target domain dataset follows the same annotation method as the source domain dataset.

3. The adaptive classification method for dynamic changes in microseismic signal characteristics as described in claim 1, characterized in that, In step S1, the input to the model is a signal image, which is cropped and transformed to a uniform size of 224×224 pixels. Both the source domain data and the target domain data are input to the feature extractor after this preprocessing step. The VIT network divides the input image into 16×16 pixel blocks, totaling (224 / 16). (224 / 16) = 196 small blocks; each small block is mapped to a fixed-dimensional vector space through a linear layer; subsequently, a location embedding layer is used to attach location encoding to each image small block and the classification label to preserve the location information in the sequence; Classification tags are used to aggregate global image information; After these preprocessing steps, the small blocks and CLS tags are fed together into the Transformer encoder to capture a more refined feature representation.

4. The adaptive classification method for dynamic changes in microseismic signal characteristics as described in claim 3, characterized in that, For a single sample, when facing a multi-label problem, it is expanded to perform independent binary classification for the prediction of each label, and the final loss is the sum of the binary classification losses of each label; the formula for calculating the cross-entropy loss function of the multi-label classifier is: in For the actual label vector, For the model to the first i The predicted probability of a class, i.e., the probability generated by the Softmax function; its calculation formula is: The binary cross-entropy loss function used by the domain discriminator is calculated as follows: in y Represents the actual source and target domain tags. This represents the label predicted by the discriminator; The total loss of the Ada-VIT model is: Experiments have verified that The value is 1. When the value is 0.5, the performance of the model can reach a balanced and stable state.

5. The adaptive classification method for dynamic changes in microseismic signal characteristics as described in claim 1, characterized in that, In step S3, 6970 sets of engineering monitoring data were collected, specifically distributed as follows: 1001 sets of standard microseismic data, 1001 sets of current microseismic data, 1001 sets of noise microseismic data, 1001 sets of current-noise microseismic data, 501 sets of pure current data, 967 sets of pure noise data, 501 sets of noise-current data, 500 sets of blasting data, and 497 sets of car horn data. The collected source domain dataset will be used to build a training model, while the target domain dataset will be further divided into a validation set and a test set. During data partitioning, the ratio of training set, validation set, and test set was 7:2:

1. The Ada-VIT model training process used Adam as the optimizer to enable rapid model convergence. The learning rate of the feature extractor and classifier was set to 0.00001, and the learning rate of the domain discriminator was set to 0.0001. The training was iterated 500 times. The convolution and attention mechanisms in the transformer were used to effectively extract signal features. t-SNE visualization is performed on signal features of different categories in source and target domain datasets. This method visualizes the internal structure of high-dimensional data and reveals its potential distribution relationship by mapping high-dimensional features to a two-dimensional space. After processing by the feature extractor in the Ada-VIT model, different types of noise features of the signal are differentially identified. At the same time, the feature distribution patterns of different distribution domains are learned. Through adversarial training of the domain discriminator, the distribution difference between the source domain and the target domain becomes smaller and smaller, thereby increasing the model's domain generalization ability. The VIT network architecture has a total of 12 layers. When calling the pre-trained model, its pre-trained weight parameters are retained. First, all network layers are unfrozen for training. Comparative experiments were conducted on different numbers of frozen layers. The last two layers of the unfrozen VIT model were selected to train the dataset for feature extraction.

6. The adaptive classification method for dynamic changes in microseismic signal characteristics as described in any one of claims 1-5, characterized in that, In step S4, accuracy, precision, recall, F1 score, and AUC are used for comprehensive evaluation. Accuracy measures the overall accuracy of the model's prediction results, i.e., the proportion of samples correctly predicted by the model out of the total samples. Precision evaluates the model's accuracy in predicting positive samples, i.e., the proportion of samples predicted as positive that are actually positive. Recall reflects the model's recall ability for positive samples, i.e., the proportion of samples that are actually positive that are correctly identified by the model. The F1 score takes into account both precision and recall, and its value reflects the balance between the model's identification of positive and negative samples during the classification process. The AUC metric measures the robustness of the model to changes in different classification thresholds by the area under the ROC curve. The higher the AUC value, the stronger the model's ability to distinguish between positive and negative samples under different thresholds.

7. The adaptive classification method for dynamic changes in microseismic signal characteristics as described in claim 6, characterized in that, The formulas for calculating the evaluation indicators are as follows: (6) to (11): (6) (7) (8) (9) (10) (11) Among them, TP represents true positive; TN represents true negative; FP represents false positive; FN represents false negative; and FPR is the false positive rate.

8. The adaptive classification method for dynamic changes in microseismic signal characteristics as described in claim 6, characterized in that, The mean accuracy is calculated to comprehensively evaluate the overall performance of the Ada-VIT model in multi-label classification tasks. The calculation formulas are as follows (12)~(13): For a single sample i, its true label set is The model's prediction score for all labels is Sort all tags according to their scores to obtain the sorted tag sequence. The formula for calculating average accuracy is: Where n represents the total number of tags; Indicates preceding k Of the predictions, the number of correctly predicted labels accounted for... k proportion, Indicates the sorted order of the first... k A tag, For indicator functions, when belong The value is 1 if the condition is met, and 0 otherwise; the formula for calculating mAP is: Where N represents the total number of samples. This represents the average accuracy of the i-th sample.

Images