Self-adaptive classification method for dynamic change of micro-seismic signal characteristics

Abstract

The invention discloses a self-adaptive classification method for dynamic changes of micro-seismic signal features, and relates to the technical field of intelligent monitoring and information. According to the technical scheme, an Ada-VIT model integrating a feature extractor, a multi-label classifier and a domain discriminator is constructed; the features are optimized by using a self-adaptive mechanism so as to improve the recognition precision of the micro-seismic signals in a complex change environment; performing model training by using the clean micro-seismic signals and the pure noise signals as source domain samples, and performing identification and classification by using the micro-seismic signals which are more complex and contain noise after feature change as target domain samples; and verifying the adaptive ability of the model in a complex change environment. The method has the beneficial effects that the recognition precision of the microseismic signals and the model adaptive capacity are improved, the multi-label classification performance is optimized, the calculation efficiency and the resource utilization rate are improved, and comprehensive evaluation indexes are adopted for performance verification; therefore, the method has important application value and popularization prospect in the field of geological disaster monitoring and early warning.

Description

Technical Field

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

[0002] As an important tool for geological disaster monitoring and early warning, microseismic monitoring technology has been widely applied in many fields. In underground engineering, microseismic detection technology is mainly used for early warning of rockburst and collapse risks; in the field of mine exploitation, it is often used to predict disasters such as gas outburst and mine water inrush; while in the reservoir dam project, it is used to monitor leakage behavior and structural deformation. By real-time monitoring and processing of microseismic signals, potential engineering risks can be identified in advance, so as to take preventive measures before disasters occur. However, signal recognition, as the primary step in microseismic monitoring data processing, its accuracy is particularly important for evaluating key parameters of disasters. Changes in environmental conditions (such as temperature, humidity, construction disturbances, and interference from strong noise backgrounds) will significantly change the microseismic waveform characteristics, resulting in the problem of insufficient adaptability of traditional microseismic signal recognition methods in long-term monitoring and complex environments. This not only reduces the accuracy of signal recognition, but also directly affects the reliability of the disaster warning system and the timeliness of response. In order to cope with the waveform characteristic changes caused by environmental changes and improve the accuracy and stability of microseismic signal recognition, it is particularly important to study and apply adaptive microseismic signal recognition technology.

[0003] In previous research on microseismic signal recognition, significant progress has been made through various methods. Early research mainly fell into two categories: waveform spectrum analysis and statistical analysis methods. Arrowsmith et al. (2006), Derr (1970), and Hedlin et al. (1990) mainly used waveform spectrum analysis to identify microseismic signals; Booker and Mitronovas (1964), Taylor (1996), and Wüster (1993) preferred to use statistical analysis methods. However, these methods usually rely on specific features extracted from the overall signal, resulting in the loss of some signal information. In addition, they rely heavily on the number of sensors and expertise, making it difficult for on-site personnel to operate efficiently. With the development of computer technology, artificial intelligence algorithms have received extensive attention from scholars due to their advantages in efficiency and automation. Vallejos and McKinnon (2013) used Logistic regression and artificial neural network (ANN) to distinguish blasting and microseismic signals; Zhang et al. (2019a) decomposed microseismic events and blasting signals, and after extracting feature vectors, used the extreme learning machine model for signal recognition; Fan et al. (2022) developed an artificial intelligence recognition model through WSD transform and SVM, providing a fast and highly accurate classification method. Although these methods have promoted the development of microseismic signal recognition, their performance highly depends on the effectiveness of feature extraction and usually requires manual design of the feature extraction process. To overcome this limitation, Bi et al. (2021) proposed an interpretable time-frequency convolutional neural network (XTF-CNN) for automatic recognition of rock fracture and noise signals, with an accuracy rate of 95.17%; Li et al. (2022) used VGG16, ResNet18, AlexNet, and an integrated model for automatic recognition and classification of current, blasting, microseismic, noise, and siren noise signals, with the highest accuracy rate reaching 98%; Ma et al. (2023) proposed a microseismic signal recognition model based on short-time Fourier transform STFT and deep learning technology for classifying microseismic, blasting, and Gaussian environmental noise, achieving a microseismic event recognition accuracy rate of 97%. These methods can automatically extract the features of microseismic signals, thus significantly improving the accuracy and efficiency of recognition.

[0004] Although previous studies have achieved remarkable results in the field of microseismic signal recognition, in the deep engineering construction environment, complex and variable environmental factors (such as high temperature, high humidity, high in-situ stress, and construction methods) seriously affect the signal recognition accuracy. Especially during long-term monitoring, the signal waveform characteristics will change significantly with the variation of environmental conditions. Due to the lack of adaptability, traditional classification models are difficult to adjust the feature extraction method in a timely manner, resulting in a significant decline in classification accuracy. For example, Wang and Tang (2022) found that in the same project, different excavation methods had a serious impact on the signal recognition reliability of the ResSCA model, and its recognition accuracy decreased from 97% to 80.6%. Therefore, there is an urgent need to explore an adaptive microseismic signal recognition method that can accurately extract the effective features of signals and achieve high-precision classification under environmental changes. At the same time, the processing of large-scale sample data usually requires a large amount of time and human resources. And even if a large amount of data is obtained, traditional computing platforms still face the problem of excessive computational resource requirements when dealing with deep learning models. For example, Dong et al. (2023) pointed out that deep convolutional neural networks usually 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 to effectively alleviate the double pressure of time cost and computational resources and improve the applicability of the model in practical engineering.

[0005] The effective identification of microseismic signals is a basic and practical engineering problem, and many scholars have proposed different research solutions. With the rapid development of deep neural network architecture, ResNet has significantly alleviated the problem of gradient vanishing in deep networks with its jump connection mechanism, enabling the network to learn transformations superimposed on the input, and has achieved breakthrough results in the fields of 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 using it for microseismic signal detection and classification in complex noise environments. Recurrent neural networks (RNNs) can maintain information memory and capture the dynamic change characteristics of sequence data due to their recursive structure, so that the temporal characteristics of microseismic signals can be accurately extracted in signal detection tasks, enhancing the real-time detection capability of microseismic events. The Transformer model was proposed by Vaswani et al. (2017) in 2017. The model effectively solves the long-distance dependency problem in sequence data through the self-attention mechanism, and significantly improves the training efficiency because it can process the entire sequence in parallel. Visual Transformer (VIT) is based on the Transformer architecture. It divides the image into multiple small patches and processes these small patches as sequences, using the self-attention mechanism to effectively capture the global characteristics and time dependencies of seismic signals. In addition, for complex signals with extremely low signal-to-noise ratio, VIT can extract more refined features with the help of the self-attention mechanism, and achieve accurate identification of effective signals and noise in low signal-to-noise ratio microseismic signals, thereby improving detection accuracy.

[0006] However, the noise in actual microseismic signal analysis usually contains multiple different categories of features. Although the above architecture has achieved remarkable success in single-label classification tasks, in multi-scenario microseismic signal processing, multiple noise types are superimposed on weak effective signals, which greatly increases the difficulty of noise and signal identification. In addition, the existence of multiple noise categories further increases the complexity of distinguishing different noise types. In order to solve the problems of semantic redundancy and co-dependency in multi-label classification, Wang et al. (2016) proposed an architecture combining convolutional neural network (CNN) and recurrent neural network (RNN) in 2016. The architecture uses an encoding attention mechanism to extract image features based on previous prediction results. In view of the weak co-dependency between noise types in microseismic signals, the CNN-RNN architecture is adjusted to achieve adaptive recognition and classification of noise types in microseismic signals. Specifically, CNN is first used to extract features from the input signal, and then the extracted features are processed by the fully connected layer (FC), followed by the RNN network for time series modeling, and finally multi-label classification is performed through the Softmax layer. The detailed description of the model architecture is as followsFigure 1 As shown in Figure 1 . The CNN-RNN hybrid model architecture shows significant performance advantages in multi-classification problems. However, when facing unknown noise categories and scenarios with complex noise combinations and superpositions, its adaptive recognition ability has limitations. Summary of the Invention

[0007] To solve the above technical problems existing in the prior art, the present invention provides an adaptive classification method for dynamically changing microseismic signal features, and proposes an Ada-VIT model suitable for adaptive classification of microseismic signals with low signal-to-noise ratio in deep and complex environments. This model consists of a feature extractor, a multi-label classifier, and a domain discriminator. By introducing an adaptive mechanism, the signal features are dynamically optimized to improve the recognition accuracy of microseismic signals in complex changing environments. Clean microseismic signals and pure noise signals collected from the Han River to Wei River Diversion Project are used as source domain samples for model training, and microseismic signals with changed features are used as target domain samples for detection and classification. By comparing with the ResNet model and the CNN-RNN model, the superiority of the Ada-VIT model is verified, and it is further applied to the microseismic data set of the Jinping II Hydropower Station.

[0008] To achieve the above object, the present invention provides the following technical solutions:

[0009] An adaptive classification method for dynamically changing microseismic signal features, the steps are as follows:

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

[0011] S2. Use an adaptive mechanism to optimize the features to improve the recognition accuracy of microseismic signals in complex changing environments;

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

[0013] S4. Verify the adaptive ability of the model in complex changing environments.

[0014] Further, 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 regarded as two different domains for processing. The source domain dataset covers the following types of signals: 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 regarded as an independent class label, and each sample is assigned a binary value according to whether it contains a specific class of signal: if the signal of this class exists in the sample, it is labeled as 1; if not, it is labeled as 0. The label i is represented as a one-hot vector e i =[0,...,1,0,...,0] with the i-th position being 1 and the others being 0. The production method of multi-labels is obtained by multiplying the one-hot vector e i and the label matrix U l to get the label vector y i which can be expressed as:

[0015] y i =U l ·e i (1)

[0016] The target domain dataset contains microseismic signal samples under noise interference. In the process of constructing the labels of the target domain dataset, the same annotation method as the source domain dataset is followed.

[0017] Further, in step S1, the feature extractor uses a pre-trained VIT model. This architecture relies on the global self-attention mechanism to accurately learn the mutual associations between small patches in the image, build long-range dependencies between any two small patches, 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 the features learned by the feature extractor and uses these features to effectively identify the source domain and target domain signals, and at the same time classify the noise carried by the samples. This process is realized through two fully connected layers, which map high-dimensional features to a low-dimensional space and accordingly identify the class to which the sample belongs.

[0019] The domain discriminator receives the feature vectors generated by the feature extractor and realizes binary classification decision through the sigmoid activation function to judge the source of the feature vectors. By introducing the adversarial training mechanism of the domain discriminator, the feature extractor is guided to learn feature representations that can make the domain discriminator unable to accurately distinguish the source domain from the target domain, so as to establish a higher degree of feature similarity between the source domain and the target domain.

[0020] Further, in step S1, the input of the model is a signal image. After cropping and transformation, its size is uniformly adjusted to 224×224 pixels. Both the source domain data and the target domain data are input into the feature extractor after passing through this preprocessing step. The VIT network divides the input image into small blocks of 16×16 pixels, totaling (224 / 16)*(224 / 16) = 196 small blocks. Each small block is mapped to a vector space of a fixed dimension through a linear layer. Subsequently, position encodings are appended to each image small block and the classification token through a position embedding layer to retain the position information in the sequence. The classification token is used to aggregate global image information, which is represented by an asterisk (*) in the figure. After these preprocessing steps, the small blocks and the CLS token are input into the Transformer encoder together to capture a more refined feature representation.

[0021] Further, for a single sample, when facing a multi-label problem, it is expanded into independent binary classification for the prediction of each label according to the above modeling method, and the final loss is the sum of the binary classification losses of each label. The calculation formula of the cross-entropy loss function of the multi-label classifier is:

[0022]

[0023] where y i is the true label vector, is the predicted probability of the model for the i-th class, that is, the probability generated by the Softmax function. Its calculation formula is:

[0024]

[0025] The calculation formula of the binary cross-entropy loss function used by the domain discriminator is:

[0026]

[0027] where y represents the true source domain and target domain labels, 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] Verified by experiments, when α takes the value of 1 and β takes the value of 0.5, the performance of this model can reach balance and stability.

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

[0031] During the data division process, following the principle that the ratio of the training set, validation set, and test set is 7:2:1, the Adam optimizer is used in the training process of the Ada-VIT model to enable the model to converge quickly. The learning rates of the feature extractor and classifier are set to 0.00001, and the learning rate of the domain discriminator is set to 0.0001. Iterative training is performed 500 times; effective signal feature extraction is achieved through the convolution and attention mechanism modules in the transformer; t-SNE visualization is performed on the signal features of different categories in the source domain and target domain datasets. This method maps high-dimensional features to a two-dimensional space, thereby visually displaying the internal structure of high-dimensional data and revealing its potential distribution relationship;

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

[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, the entire network layers are unfrozen for training, and comparative experiments are conducted on different frozen layers. The last two layers of the VIT model are selected to be unfrozen to train the dataset for feature extraction.

[0034] Further, in step S4, the accuracy, precision, recall, F1 score, and AUC metrics are comprehensively considered. Among them, accuracy is used to measure the accuracy of the overall prediction results of the model, that is, the proportion of samples correctly predicted by the model in the total samples; precision evaluates the precision of the model when predicting positive samples, that is, the proportion of samples actually positive among the samples predicted as positive; recall reflects the recall ability of the model for positive samples, that is, the proportion of samples actually positive that are correctly identified by the model; the F1 score comprehensively considers precision and recall, and reflects the balance of the model in identifying positive and negative samples during the classification process through the value of it; while the AUC metric measures the robustness of the model to changes in different classification thresholds through the area under the ROC curve. The higher the AUC value, the stronger the ability of the model to distinguish positive and negative samples at different thresholds.

[0035] Further, the calculation formulas of the evaluation metrics are as shown in formulas (6) - (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] Further, the mean average precision (mAP) is calculated to comprehensively evaluate the overall performance of the Ada-VIT model in the multi-label classification task. Its calculation formula is as shown in (12) - (13):

[0039] For a single sample i, its true label set is Y i , and the prediction scores of the model for all labels are f i . Sort all labels according to the scores to obtain the sorted label sequence R i ; the calculation formula for the average accuracy is:

[0040]

[0041] Among them, n represents the total number of labels; P i (k) represents the proportion of the number of correctly predicted labels in the first k predictions to k, and R i (k) represents the k-th label after sorting, and δ(R i (k) ∈ Y i is an indicator function, which takes the value of 1 when R i (k) belongs to Y i , and 0 otherwise; the calculation formula for mAP is:

[0042]

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

[0044] Advantages of the present invention:

[0045] Compared with the prior art, the adaptive classification method for dynamic changes in microseismic signal features of the present invention has the following technical features or advantages:

[0046] (1) Improve the recognition accuracy of microseismic signals: By constructing the Ada-VIT model that integrates a feature extractor, a multi-label classifier, and a domain discriminator, the present invention can accurately capture the global features of microseismic signals, effectively distinguish microseismic signals from noise, and significantly improve the recognition accuracy of microseismic signals even in the complex and changeable deep engineering environment. This benefits from the adaptive mechanism in the Ada-VIT model, which can dynamically optimize the signal features to adapt to signal changes in different environments.

[0047] (2) Enhance the adaptive ability of the model: The present invention introduces a domain discriminator and an adversarial training mechanism to prompt the feature extractor to learn feature representations that can prevent the domain discriminator from accurately distinguishing the source domain from the target domain. This design enables the Ada-VIT model to establish a higher degree of feature similarity between the source domain and the target domain, thereby enhancing the adaptive ability of the model in different environments. This is of great significance for long-term monitoring and early warning of geological disasters because environmental changes will significantly affect the microseismic waveform features.

[0048] (3) Optimize the multi-label classification performance: Aiming at various types of noise that may exist in microseismic signals, the present invention uses a multi-label classifier for processing. By assigning an independent class label to each type of noise and using binary values for annotation, the Ada-VIT model can simultaneously identify and classify multiple types of noise in microseismic signals. This multi-label classification method improves the accuracy and robustness of the model when dealing with complex microseismic signals.

[0049] (4) Improve the computational efficiency and resource utilization rate: Although deep learning models usually consume a large amount of time and computational resources when processing large-scale sample data, the present invention effectively improves the computational efficiency by using the pre-trained VIT model as the feature extractor and unfreezing the last two layers for training. In addition, by optimizing parameter settings such as the learning rate, the convergence speed and resource utilization rate of the model are further improved.

[0050] (5) Comprehensive evaluation indicators are comprehensive: When verifying the performance of the model in the present invention, multiple indicators such as accuracy, precision, recall, F1 score, and AUC are used for comprehensive consideration. These indicators can comprehensively reflect the performance of the model in different aspects, including the overall prediction accuracy, the precision of positive sample prediction, the recall ability of positive samples, the balance of positive and negative sample recognition, and the robustness of the model to changes in different classification thresholds. In addition, the overall performance of the Ada-VIT model in multi-label classification tasks is comprehensively evaluated by calculating mAP, further demonstrating the superiority and practicality of the model.

[0051] (6) Broad engineering application prospects: The Ada-VIT model of the present invention has been applied and verified in actual projects such as the Han-Ji-Wei Water Diversion Project and the Jinping II Hydropower Station, and excellent results have been obtained. 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 microseismic signals, potential engineering risks can be identified in advance and preventive measures can be taken, which can significantly reduce the occurrence probability and harm degree of geological disasters.

[0052] By proposing an adaptive classification method for dynamically changing microseismic signal features and its Ada-VIT model, the present invention not only improves the recognition accuracy of microseismic signals and the model's adaptive ability, but also optimizes the multi-label classification performance, improves the computing efficiency and resource utilization rate, and uses comprehensive evaluation indicators for performance verification; these beneficial effects make the present invention have important application value and promotion prospects in the field of geological disaster monitoring and early warning. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the present invention will be described in detail below in conjunction with the drawings and detailed embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0054] Among them:

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

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

[0057] Figure 3 is the geographical location of the Han-Ji-Wei Water Diversion Project;

[0058] Figure 4It is the visualization diagram of the source domain, target domain dataset division and features: a Visualization image of the t-SNE feature distribution of the source domain and target domain datasets; b Visualization image of the signal features extracted during the training process of the Ada-VIT model; c Visualization image of the t-SNE features of various types of signals in the source domain dataset; d Visualization image of the t-SNE features of various types of signals in the target domain dataset;

[0059] Figure 5 It is the test effect diagram of the Ada-VIT model;

[0060] Figure 6 It is the t-SNE visualization diagram of the domain adaptation effect of the Ada-VIT model (best viewed in terms of color): (a) The effect without being activated by Ada-VIT; (b) The effect after the domain adversarial training of the Ada-VIT model; (c) The distribution effect of various types of signals after the domain adversarial training of the Ada-VIT model;

[0061] Figure 7 It is the comparison diagram of the test effects of different models;

[0062] Figure 8 It is the F1 score diagram of each model;

[0063] Figure 9 It is the geographical location of Jinping Mountain;

[0064] Figure 10 It is the t-SNE visualization diagram of the dataset of Jinping II Hydropower Station;

[0065] Figure 11 It is the application result diagram;

[0066] Figure 12 It is the model performance evaluation index diagram. Specific implementation manners

[0067] In order to make the objectives, technical solutions and advantages of the present invention clearer and more understandable, the present 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 only used to explain the present invention and are not used to limit the present invention. The following combines the attached Figure 1-12 A further description is made on the adaptive classification method for the dynamic changes of microseismic signal features, and the technical solutions in the embodiments of the present application are clearly and completely described; obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without making creative efforts belong to the scope of protection of the present application.

[0068] The construction environment of deep underground engineering is complex and changeable, and the resulting interference signals with different characteristics reduce the long-term recognition accuracy of microseismic signals. The present invention proposes a method for recognizing microseismic signals adaptable to environmental changes. This method constructs an Ada-VIT model integrating feature extraction, multi-label classification, and domain adaptation capabilities, and uses an adaptive mechanism to optimize the features, thereby improving the recognition accuracy of microseismic signals in a complex and changing environment. Clean microseismic signals and pure noise signals collected from the Han River to Wei River Diversion Project are used as source domain samples for model training, and microseismic signals that are more complex and contain noise after feature changes are used as target domain samples for recognition and classification to verify the adaptive ability of the model in a complex and changing environment. The results show that the Ada-VIT model significantly outperforms the ResNet50 and improved CNN-RNN models in the recognition and classification of microseismic signals, and its mean average precision (mAP) is as high as 0.98. In addition, when the adaptive classification model is applied to the Jinping II Hydropower Station, it is found that the model has excellent adaptive classification performance in actual engineering. The present invention provides important theoretical support and technical guarantee for the long-term recognition and accurate early warning of microseismic signals in deep engineering.

[0069] Example 1

[0070] Project modeling: In the case of low signal-to-noise ratio and limited sample size, it is a very challenging task to accurately extract the effective signal features. Traditional convolutional neural networks (CNNs) are no longer sufficient to address this challenge, and using a ResNet network with a large number of layers results in poor training effects and excessive time consumption due to insufficient sample size. Given the sequential feature characteristics of microseismic signals, the present invention considers using a CNN-RNN model architecture to identify noise. However, for microseismic signals with multi-class superimposed noise, their feature distributions have changed compared to microseismic signals with a single type of noise. It is no longer a simple linear superposition, but a fundamental transformation in the distribution domain. Therefore, the present invention proposes to treat 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 types of signals: 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 regarded as an independent class label, and each sample is assigned a binary value according to whether it contains a specific class of signal: if the signal of this class exists in the sample, it is marked as 1; if not, it is marked as 0. The label i is represented as a one-hot vector e i =[0,...,1,0,...,0] The i-th position is 1, and the others are 0. The production method of multi-labels can be obtained by multiplying the one-hot vector e i and the label matrix U l to get the label vector y i can be expressed as:

[0071] y i =U l ·e i (1)

[0072] The target domain dataset contains microseismic signal samples under noise interference. In the process of label construction of the target domain dataset, the same annotation method as the source domain dataset is followed.

[0073] Proposed method: In order to solve the problem of being unable to effectively identify various types of noise in complex noise superposition scenarios, this paper 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 types of superimposed noise. The structural diagram of the model is shown in the figure. Figure 2 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, wherein the feature extractor uses a pre-trained VIT model. This architecture uses a global self-attention mechanism to accurately learn the correlation between small blocks in the image and build a long-range dependency relationship between any two small blocks, thereby having the ability to accurately capture the global features of microseismic signals and efficiently distinguish microseismic signals from noise. For small sample microseismic datasets, since the VIT model has been trained and learned on massive image datasets and has mastered common visual features, fine-tuning can not only save training time, but also achieve good classification results.

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

[0076] The domain discriminator receives the feature vector generated by the feature extractor and implements a binary classification decision through the sigmoid activation function to determine the source of the feature vector (target domain or source domain). By introducing the adversarial training mechanism of the domain discriminator, the feature extractor can be guided to learn feature representations that make it impossible for the domain discriminator to accurately distinguish between the source domain and the target domain, thereby establishing a higher degree of feature similarity between the source domain and the target domain. This process significantly reduces the distribution difference between the source domain and the target domain, and effectively enhances the generalization ability of the model in the target domain. In this way, the model can better adapt to the data distribution of the target domain and improve the overall performance of cross-domain transfer learning.

[0077] The input of the model is a signal image. After cropping and transformation, its size is uniformly adjusted to 224×224 pixels. Both the source domain data and the target domain data are input into the feature extractor after passing through this preprocessing step. The VIT network divides the input image into small patches of 16×16 pixels, totaling (224 / 16)*(224 / 16) = 196 patches. Each patch is mapped to a vector space of a fixed dimension through a linear layer. Subsequently, position encoding is appended to each image patch and a specially designed classification token (CLS token) through the Position Embedding layer to preserve the position information in the sequence. The classification token (CLS token) is used to aggregate global image information, which is represented by an asterisk (*) in the figure. After these preprocessing steps, the patches and the CLS token are input into the Transformer encoder together to capture a more refined feature representation.

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

[0079]

[0080] where y i is the true label vector, is the predicted probability of the model for the i-th class, that is, the probability generated by the Softmax function. Its calculation formula is:

[0081]

[0082] The calculation formula of the binary cross-entropy loss function used by the domain discriminator is:

[0083]

[0084] where y represents the true source domain and target domain labels, 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] Verified by experiments, when α is taken as 1 and β is taken as 0.5, the performance of the model can reach balance and stability.

[0087] All calculations and neural network training were carried out 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 example provides an introduction to the project background and dataset of the Hanjiang-to-Weihe River Diversion Project in Shaanxi, China.

[0090] The dataset of the present invention is derived from the Hanjiang-to-Weihe River Diversion Project in Shaanxi, China. As Figure 3 , this project straddles the Yangtze River and Yellow River basins and crosses the Qinling Mountains region. The project mainly consists of three major parts: the Qinling Water Conveyance Tunnel, the Huangjinxia Water Control Project, and the Sanhekou Water Control Project. Among them, the Qinling Water Conveyance Tunnel, as the core control project of the entire project, has a total length of 98.3 kilometers and a maximum burial depth of more than 2,000 meters. Due to the complex geological conditions in the construction area, including factors such as high ground stress, high seepage flow, and high temperature, it is extremely easy to induce geological disasters such as rock bursts and water inrushes, posing a great threat to construction safety and equipment operation. To reduce the safety risks during construction, a microseismic monitoring system was introduced. The system includes a host analysis module, a data collector, and a uniaxial accelerometer. Among them, the design parameters of the accelerometer include a response frequency range of 50Hz to 5kHz, a sensitivity of 30V / g, and a size of 25.4mm in diameter and 122mm in height. The system sampling frequency is set at 10kHz, and the duration of each acquisition is 400ms.

[0091] Since the microseismic signals with multiple types of noise superimposed have significantly changed in feature distribution compared to the microseismic signals with a single type of noise, this change is not a simple feature accumulation but rather causes a transformation of the feature distribution domain. As Figure 4 shown in Figure a, the blue dots in the figure represent the source domain data features, and the red dots represent the target domain data features. It can be seen from the figure that there are large inter-class differences between the data features of the source domain and the target domain.

[0092] A total of 6,970 sets of monitoring data for the Han River to Wei River Diversion Project were collected in this study, with the specific distribution as follows: 1,001 sets of standard microseismic data, 1,001 sets of current microseismic data, 1,001 sets of noise microseismic data, 1,001 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 construct a training model, while the target domain dataset is further divided into a validation set and a test set. During the data division process, the principle of a 7:2:1 ratio for the training set, validation set, and test set was followed. The Adam optimizer was used in the training process of the Ada-VIT model, which can enable the model to converge quickly. Through experimental verification, the learning rates of the feature extractor and classifier were set to 0.00001, and the learning rate of the domain discriminator was set to 0.0001. The model is relatively stable and has the highest classification accuracy, with 500 iterations of training. The visualization of the signal features extracted during the training process using the Ada-VIT model is shown in Figure 4 as shown in b. Effective extraction of signal features can be achieved through the convolution and attention mechanism modules in the transformer. As shown in Figure 4 c-d, the present invention performed t-SNE visualization on the signal features of different categories in the source domain and target domain datasets. 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 relationship. Different colors were used in the figure to distinguish the features of different label categories. It can be found that the data distribution in the source domain has relatively significant separability, but there is a certain degree of overlap and dispersion in some categories.

[0093] In the high-dimensional feature space, the sample feature points of microseismic signals, blasting signals, and car horn signals show a high degree of clustering, which reflects the significant consistency of the samples within these categories in terms of features. On the contrary, the sample feature points of noise, current-noise, and current signals are more discretely distributed in space, and some sample feature points even deviate from the main clustering center, indicating that there are significant differences in the features of the samples within these categories, and some samples are affected by external complex noises. In addition, there is an obvious gap in space between microseismic signals and blasting signals and other signal categories, indicating that the model has high efficiency in distinguishing these signal categories. However, the current signal and the noise current signal almost completely overlap in the feature space, and there is also a partial overlap area between the noise signal and the car horn signal, which reveals that the features of these signal categories have a certain degree of similarity 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 show a high degree of discreteness and there are overlapping areas between them. In particular, the noise microseismic signal and the noise-current microseismic signal almost completely overlap, further indicating that these signal categories have a high degree of similarity in feature distribution, and it is extremely difficult to accurately distinguish these signal categories. After being processed by the feature extractor in the Ada-VIT model designed in this paper, it can differentially identify the noise features of different types of signals, and at the same time learn the feature distribution laws of different distribution domains. Through the adversarial training of the domain discriminator, the distribution differences between the source domain and the target domain become smaller and smaller, thereby increasing the domain generalization ability of the model.

[0094] The VIT network architecture has a total of 12 layers. When calling the pre-trained model, its pre-trained weight parameters are retained. First, the entire network layers are unfrozen for training. In order to balance the performance, computational cost, and training duration of the model, comparative experiments are carried out on freezing different numbers of layers. The experimental results are shown in Table 1 below. Finally, the last two layers of the VIT model are selected to unfreeze and train the dataset for feature extraction. While reducing the training duration and saving memory overhead, the model can also achieve good training results.

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

[0096]

[0097]

[0098] Example 3

[0099] This example provides model evaluation metrics.

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

[0101]

[0102] Among them, TP (True Positive) represents true positive; TN (True Negative) represents true negative; FP (False Positive) represents false positive; FN (False Negative) represents 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 , and the prediction scores of the model for all labels are f i . Sort all labels according to the scores to obtain the sorted label sequence R i . The calculation formula for the average accuracy is as follows:

[0104]

[0105] Among them, n represents the total number of labels. P i (k) represents the proportion of the number of correctly predicted labels in the first k predictions to k, and R i (k) represents the kth label after sorting, and δ(R i (k) ∈ Y i is an indicator function, which takes the value of 1 when R i (k) belongs to Y i , and 0 otherwise. The calculation formula for mAP is as follows:

[0106]

[0107] where \(N\) represents the total number of samples, and \(AP\) i represents the average accuracy of the \(i\)-th sample.

[0108] To better evaluate the performance of the Ada-VIT model during the training process, the performance of the test set during the model training process was visualized (as Figure 5 shown). It can be seen from Figure 5 a that as the number of iterations increases, the validation loss of the model continuously decreases until it approaches 0.05, and the accuracy gradually increases and stabilizes at 0.91. Since the multi-label classification problem is transformed into multiple binary classification problems for processing in this paper, precision and recall are also used to comprehensively evaluate the performance of the model. From Figure 5 b, it can be seen that at the initial stage of training, both precision and recall increase significantly, indicating that the model can quickly learn the patterns in the data. As the training progresses, precision and recall tend to be stable, and when the model training reaches a stable state, the values of the two reach 91.45% and 92.17% respectively.

[0109] Another focus of this work is to train using known signals and different types of noise, and to achieve the identification and classification of different noise types in the microseismic signals with multiple noise superpositions through feature learning and domain adversarial training. For this purpose, the domain adaptation performance of the Ada-VIT model was analyzed, and the results are as Figure 6 shown: Figure 6 a shows the initial distribution of the source domain and target domain features only after the feature extractor, Figure 6 and b shows the comparison of the 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 the feature distributions between the source domain and the target domain is significantly reduced, and the features of the two domains gradually form significant overlaps in multiple regions, indicating that the similarity of the inter-domain features has been effectively improved, thus achieving the domain adaptation goal. Although some clusters are still mainly composed of data points in a single domain, after further visualizing the specific categories, it is found that the inter-domain differences between the source domain data (such as current signals, current-noise signals) and the target domain data (such as current-noise-microseismic signals) are significantly reduced and tend to the same clustering cluster. In addition, the inter-domain differences between the current signals in the source domain and the current-noise-microseismic signals and current-microseismic signals in the target domain are also reduced synchronously, and similar domain alignment laws are also shown among the three types of features of noise signals, microseismic signals, and noise-microseismic signals.

[0110] Interestingly, the noise-microseismic signal, with the dual characteristics of both noise and microseismic signals, forms a single cluster as a bridge connecting the two after domain adversarial training. It is also worth noting that even though the inter-domain distance is significantly reduced and similar features gradually converge into the same cluster after domain adversarial training, the distribution differences between noise and effective microseismic signals are still very obvious, and there is still a clear distinction between different types of signals.

[0111] In this paper, the classification performances of three models, namely the classic classification network ResNet50, the improved multi-label classification network CNN-RNN, and the proposed multi-label adaptive classification network Ada-VIT, are compared. The results are as Figure 7 shown. Through comparative analysis, it can be seen that ResNet50 performs poorly in multi-label classification. CNN-RNN has a significant improvement compared to ResNet50, and the average precision mAP for classifying the microseismic dataset used in this paper can reach about 0.8. However, the effect still needs to be improved. This is because although the CNN-RNN model increases the richness of feature extraction and can further capture the sequence features of seismic signals, for the changes in the distribution domain caused by the superposition of different types of noise and the low signal-to-noise ratio microseismic signals with small samples, CNN-RNN can no longer accurately distinguish the noise mixed in the microseismic signals. The Ada-VIT model just solves this problem through a VIT-based feature extractor and a domain discriminator, achieving effective recognition of low signal-to-noise ratio microseismic signals and accurate classification of multiple types of superimposed noise, with an mAP as high as 0.98.

[0112] The number of parameters, validation loss, and average precision of validation for the three models are shown in Table 2. From the parameters in the table, it can be seen that Ada-VIT has the largest number of parameters, which is 3.6 times that of ResNet50 and 134.8 times that of CNN-RNN. This shows that the Ada-VIT model has a high complexity and learns richer features, but it also comes with an increase in computational complexity and the demand for graphics card resources, resulting in an increase in the training duration of the model. To balance the computational amount and training duration of the model, the parameter weights of the first 10 layers are frozen and do not participate in network training when calling the VIT model, and the last two convolutional and Transformer layers are unfrozen for training to extract features. Through experimental verification, compared with other models, the Ada-VIT model has the shortest average single training duration, and the validation loss is as low as 0.24, indicating that the model can quickly learn to converge and has stable and accurate classification performance.

[0113] Table 2 Comparison of the number of parameters and training performances of different models

[0114]

[0115] As shown in Table 3, the present invention conducted performance tests on three models, including mean average precision (mAP) and accuracy (Acc), as well as the test time consumption for each test sample. The test results show that although the time required for the Ada-VIT model to process each sample increased by 0.0222 seconds and 0.0484 seconds compared to the ResNet50 and CNN-RNN models respectively, its mAP value reached 0.99, which was increased by 83.33% and 90.38% compared to the ResNet50 and CNN-RNN models respectively. In addition, the test accuracy of the Ada-VIT model reached 278.78% and 184% of the ResNet50 and CNN-RNN models respectively. These results together verify the significant advantage of the Ada-VIT model in multi-label classification performance.

[0116] Further analysis Figure 8 The F1 score in [[]] is used to show the performance of the model on different signal types. The higher the value, the better the model performs in terms of both precision and comprehensiveness. It should be noted that due to the lack of samples of blasting and siren noise signals in the test set, the F1 scores of all models for these two types of signals are zero. However, there are significant differences in the F1 scores of each model for the classification of microseismic, current, noise, and noise-current signals. Specifically, the F1 scores of the Ada-VIT model for these signal types are as high as 1, 0.97, 0.95, and 0.88 respectively, indicating its excellent classification performance for various signals. In contrast, the ResNet50 model only achieved an F1 score of over 0.5 for microseismic signals, while the CNN-RNN model only performed on microseismic and noise-current signals, and the highest F1 score was only 0.56. In summary, the performance of the F1 scores of the Ada-VIT model for different signal types not only highlights its robustness in signal classification tasks, but also further verifies the adaptability of the Ada-VIT model to different signal characteristics and its superiority in the field of signal multi-classification.

[0117] Table. 3 Comparison of test performances of different models

[0118] Model Test times (s / samples) Test-mAP Test-Acc 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 example provides an application example of the Jinping II Hydropower Station.

[0121] The Jinping II Hydropower Station is located in the middle reaches of the Yalong River Basin in Sichuan Province ( Figure 9) is one of the important projects for clean energy development in China. The project mainly includes an underground power house, a water diversion tunnel, and a flood discharge system, with a total installed capacity of 4.8 million kilowatts and an annual average power generation of about 16 billion kWh. As the deepest water diversion tunnel group in the world, the maximum burial depth of its tunnels reaches 2,525 meters. During the construction process, it faces complex geological conditions such as high ground stress, high seepage flow, and high temperature, and the risk of rockburst is particularly prominent. To ensure construction safety, the project adopts an advanced microseismic monitoring system to identify the rockburst risk. The system consists of a Hyperion digital signal processing system, a Paladin digital signal acquisition system, and six single-axis accelerometers, and the performance parameters of the accelerometers are similar to those of the Hanjiang-to-Weihe River Diversion Project.

[0122] To systematically evaluate the performance of the Ada-VIT model in different actual engineering scenarios, the Jinping II Hydropower Station was selected as the case study object, and 1,000 groups of data were collected for empirical analysis of the Ada-VIT model. 850 groups of data were randomly selected to form the training set, and the remaining 150 groups of data were used as the test set. Through detailed visual analysis of the training set and the test set (see Figure 10 ), the results revealed a significant gap between the training set and the test set in the feature space. This phenomenon indicates that the feature distributions of the two groups of data are different and the correlation is low, revealing that they are from different distribution domains. Therefore, if the Ada-VIT model can accurately identify the categories of the two types of data, it will strongly prove that the Ada-VIT model has excellent generalization ability and can perform effective signal classification across different data distribution domains.

[0123] Figure 11 Shows the performance evaluation of the Ada-VIT model in identifying microseismic signals on the Jinping II Hydropower Station dataset. The analysis results reveal a high consistency between the number of samples predicted by the model as microseismic signals (502 cases) and the number of actual microseismic signal samples (496 cases). At the same time, the difference between the number of samples predicted by the model as noise (505 cases) and the number of actual noise samples (504 cases) is also minimal. The small deviation between the prediction results and the actual data further confirms 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 deeply quantify the application potential of the Ada-VIT model in the field of adaptive signal multi-classification, a detailed evaluation of the prediction results of the Jinping II Hydropower Station dataset was carried out, and key performance indicators including Accuracy, Precision, Recall, F1score (and AUC were calculated (such as Figure 12As shown in the figure, the closer its value is to 1, the better its performance. It can be observed from the figure that all performance indicators have stably reached 0.99. These data fully demonstrate that the Ada-VIT model can not only achieve accurate classification of microseismic signals and noise under extremely limited sample conditions, but also demonstrate excellent signal separation and signal recognition capabilities in an engineering environment with strong noise and multiple interferences.

[0125] In the long-term monitoring of deep underground engineering, although existing models show high accuracy in microseismic signal recognition, their recognition performance often drops rapidly and it is difficult to maintain stable accuracy when facing the dynamic changes of complex environments. In addition, the superimposed interference of multiple noise sources on weak effective signals further increases the complexity of recognition. To address this problem, the present invention proposes an Ada-VIT model for adaptive classification of microseismic signals with low signal-to-noise ratio in a deep changing environment. This model consists of a feature extractor, a multi-label classifier, and a domain discriminator, and optimizes the features through an adaptive mechanism to achieve accurate identification of microseismic signals while detecting the types of noise mixed in the measured microseismic signals. It provides a new solution for the recognition of microseismic signals with low signal-to-noise ratio in a deep complex environment.

[0126] Clean microseismic signals and pure noise signals collected from the Hanjiwei Project are used as source domain signal samples for training, and microseismic signals after environmental changes are used as target domain signal samples for detection and recognition. The results show that the Ada-VIT model performs excellently in the recognition of microseismic signals, with its average precision mAP reaching as high as 0.98, significantly superior to the traditional ResNet50 and the improved CNN-RNN models. In addition, the Ada-VIT model shows a fast convergence speed and low loss during the training and testing processes, verifying the stability and effectiveness of the model. By freezing and fine-tuning some layers of the VIT model, Ada-VIT achieves accurate classification of small sample low signal-to-noise ratio data sets while maintaining low computational cost and video memory overhead.

[0127] To evaluate the performance of the Ada-VIT model in the recognition and classification of different engineering microseismic signals, it is applied to the signal recognition and classification of the Jinping II Hydropower Station data set. The results show that there is a high degree of consistency between the model prediction and the actual results, and key indicators such as Accuracy, Precision, Recall, F1 score, and AUC in the classification task have all stably reached 0.99, fully demonstrating the excellent classification performance of the Ada-VIT model under limited sample conditions. In addition, the model shows extremely high signal recognition and classification capabilities in a complex engineering environment with strong noise and multiple interferences, providing important theoretical support and technical guarantee for deep engineering signal monitoring and precise early warning.

[0128] The above are only the preferred specific embodiments of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, making equivalent substitutions or changes should be covered within the protection scope of the present invention.

Claims

1. An adaptive classification method for the dynamic change of microseismic signal characteristics, characterized in that Here are the steps: S1. Construct the Ada-VIT model integrating feature extractor, multi-label classifier and domain discriminator; S2, using adaptive mechanisms to optimize features to improve the recognition accuracy of microseismic signals in complex and changing environments; S3, using clean microseismic signals and pure noise signals as source domain samples for model training, and using microseismic signals that are more complex and contain noise after feature changes as target domain samples for identification and classification; S4. Verify the model’s adaptability in a complex and changing environment.

2. The adaptive classification method for the dynamic change of microseismic signal characteristics according to claim 1, characterized in that In step S1, the CNN-RNN model architecture is adopted to identify noise. The single-noise microseismic signal dataset and the multi-noise microseismic signal dataset are regarded as two different domains for processing. The source domain dataset covers the following types of signals: 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 regarded as an independent class label, and each sample is assigned a binary value according to whether it contains a specific class of signal: if the signal of this class exists in the sample, it is labeled as 1; if not, it is labeled as 0. The label i is represented as a one-hot vector e i =[0,...,1,0,...,0], with the i-th position being 1 and the others being 0. The production method of multi-labels is obtained by multiplying the one-hot vector e i and the label matrix U l to get the label vector y i which can be expressed as: y i = U l ·e i (1) The target domain dataset contains microseismic signal samples under noise interference. During the label construction process of the target domain dataset, the same annotation method as the source domain dataset is followed.

3. The adaptive classification method for the dynamic change of microseismic signal characteristics according to claim 1, characterized in that In step S1, the feature extractor uses a pre-trained VIT model. This architecture uses a global self-attention mechanism to accurately learn the correlation between small blocks in the image and build a long-range dependency relationship between any two small blocks, so as to accurately capture the global features of microseismic signals and efficiently distinguish microseismic signals from noise. The multi-label classifier receives the features learned by the feature extractor and uses these features to effectively identify the source domain and target domain signals, while classifying the noise carried by the samples; this process is implemented through two fully connected layers, which map high-dimensional features to low-dimensional space and identify the category to which the sample belongs based on this; The domain discriminator receives the feature vector generated by the feature extractor and implements a binary classification decision through a 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 feature representations that enable the domain discriminator to be unable to accurately distinguish between the source domain and the target domain, thereby establishing a higher degree of feature similarity between the source domain and the target domain.

4. The adaptive classification method for dynamically changing microseismic signal features according to claim 3, characterized in that In step S1, the input of the model is the signal image, which is uniformly resized to 224×224 pixels after cropping and transformation. Both the source domain data and the target domain data are input into the feature extractor after this preprocessing step. The VIT network divides the input image into small blocks of 16×16 pixels, totaling (224 / 16)*(224 / 16)=196 small blocks. Each small block is mapped to a vector space of fixed dimension through a linear layer. Subsequently, a positional encoding is added to each image small block and classification mark through a position embedding layer to retain the position information in the sequence. The classification mark is used to aggregate global image information, which is represented by an asterisk (*) in the figure. After these preprocessing steps, the patches are fed into the Transformer encoder along with the CLS tokens to capture finer feature representations.

5. The adaptive classification method for the dynamic change of microseismic signal characteristics according to claim 4, characterized in that, For a single sample, when facing a multi-label problem, the modeling method described in claim 4 is expanded to perform independent binary classification on the prediction of each label, and the final loss is the sum of the binary classification losses of each label; the calculation formula of the cross entropy loss function of the multi-label classifier is: where y i is the true label vector, is the predicted probability of the model for the i-th class, that is, 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: where y represents the true source and target domain labels, represents the label predicted by the discriminator; The total loss of the Ada-VIT model is: Loss all = αLoss c + βLoss d (5) Experimental verification shows that the performance of the model can achieve balanced stability when α is 1 and β is 0.

5.

6. The adaptive classification method for the dynamic change of microseismic signal characteristics according to claim 1, wherein In step S3, 6970 groups of engineering monitoring data are collected, and the specific distribution is as follows: 1001 groups of standard microseismic data, 1001 groups of current microseismic data, 1001 groups of noise microseismic data, 1001 groups of current-noise microseismic data, 501 groups of pure current data, 967 groups of pure noise data, 501 groups of noise-current data, 500 groups of blasting data, and 497 groups of car horn data; the collected source domain dataset will be used to construct a training model, while the target domain dataset is further divided into a validation set and a test set; During the data division process, following the principle that the ratio of the training set, validation set, and test set is 7:2:1, the Adam optimizer is used in the training process of the Ada-VIT model to enable the model to converge quickly. The learning rates of the feature extractor and classifier are set to 0.00001, and the learning rate of the domain discriminator is set to 0.0001, and iterative training is carried out 500 times; effective signal feature extraction is achieved through the convolutional and attention mechanism modules in the transformer; Perform t-SNE visualization on the signal features of different categories in the source domain and target domain datasets. 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 relationship; After being processed by the feature extractor in the Ada-VIT model, different types of noise features of the signal are differentially identified, and at the same time, the feature distribution laws of different distribution domains are learned. Through the adversarial training of the domain discriminator, the distribution difference between the source domain and the target domain becomes smaller and smaller, thereby increasing the domain generalization ability of the model; The VIT network architecture has a total of 12 layers. When calling the pre-trained model, its pre-trained weight parameters are retained. First, the entire network layers are unfrozen for training, and comparative experiments are carried out on freezing different layers. The last two layers of the VIT model are selected to unfreeze and train the dataset for feature extraction.

7. The adaptive classification method for the dynamic change of microseismic signal characteristics according to any one of claims 1-6, characterized in that In step S4, the accuracy, precision, recall, F1 score, and AUC metrics are comprehensively considered; among them, accuracy is used to measure the accuracy of the overall prediction result of the model, that is, the proportion of samples correctly predicted by the model in the total samples; precision evaluates the accuracy of the model when predicting positive samples, that is, the proportion of actually positive samples among the samples predicted as positive; recall reflects the recall ability of the model for positive samples, that is, the proportion of actually positive samples correctly identified by the model; the F1 score comprehensively considers precision and recall, and reflects the balance of the model's recognition of positive and negative samples in the classification process through the size of its value; and the AUC metric measures the robustness of the model to changes in different classification thresholds through the area under the ROC curve. The higher the AUC value, the stronger the model's ability to distinguish positive and negative samples at different thresholds.

8. The adaptive classification method for the dynamic change of microseismic signal characteristics according to claim 7, characterized in that The calculation formulas of the evaluation metrics are as shown in formulas (6) to (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.

9. The adaptive classification method for dynamically changing microseismic signal characteristics according to claim 7, characterized in that, Calculate the mean average precision to comprehensively evaluate the overall performance of the Ada-VIT model in the multi-label classification task. Its calculation formula is as shown in (12) - (13): For a single sample i, its true label set is Y i , the prediction scores of the model for all labels are f i , sort all labels according to the scores to obtain the sorted label sequence R i ; the calculation formula for the average precision rate is: where n represents the total number of labels; P i (k) represents the proportion of the number of correctly predicted labels in the first k predictions to k, R i (k) represents the k-th label after sorting, δ(R i (k) ∈ Y i is an indicator function, when R i (k) belongs to Y i the value is 1, otherwise it is 0; the calculation formula of mAP is: where N represents the total number of samples, and AP i represents the average accuracy of the i-th sample.

Images