Deep Clustering Variational Autoencoder and Electrical Signal Clustering Analysis Method Based on It

Abstract

The present invention relates to a deep variational autoencoder algorithm for cluster analysis of aircraft electrical signals. The algorithm comprises a variational autoencoder module, a multi-scale residual convolution network module and a signal clustering module. When performing cluster analysis on aircraft electrical signals (101), signal acquisition and transmission (102) are first performed, and the signals are sent to a multi-scale residual convolution module (103) and a maximum pooling layer (104) to complete feature extraction. The feature maps are respectively passed through two parallel fully connected layers (105) and (106) to obtain feature mean (107) and feature variance (108). The feature maps are then passed through a decoder network with a structure completely symmetrical with that of the encoder, and a reconstructed signal (113) is finally output, which is intended to completely match the input signal. The network is trained by constructing a loss function and a two-stage training method, and finally cluster analysis (115) is completed by reconstructing a feature vector (109).

Description

Technical Field

[0001] The present invention relates to a deep clustering variational autoencoder and an electrical signal clustering analysis method based on the same. Background Art

[0002] During aircraft operational testing, continuous time-series signals collected by various sensors form the primary data set. Since these sensor-collected data sets are mostly unlabeled, further analysis, particularly for signal identification and monitoring through classification, requires data capture and analysis. Domain experts manually label various signals to construct a labeled data set and form an expert database. This allows for supervised training of classifier models and the development of signal recognition systems.

[0003] However, databases and datasets containing large amounts of data are both massive and complex, with signal entries reaching hundreds of thousands, each of which is a long time series with a length exceeding several thousand. Manual classification of such large amounts of complex data is extremely time-consuming and labor-intensive, making rapid labeling of large amounts of data a difficult, time-consuming, and labor-intensive task. In the field of machine learning, unlabeled data can be automatically categorized and organized through unsupervised learning. This type of unsupervised learning algorithm is known as a cluster analysis algorithm. Because signal data of the same category share certain similarities, cluster analysis algorithms are used to categorize and organize multiple categories of signals, and manual identification is used to assist in building an expert knowledge base. This is feasible for organizing large datasets.

[0004] Traditional clustering analysis algorithms include K-means, density clustering, and fuzzy clustering. However, these algorithms rely heavily on the representative attributes of the data's features. While they perform well on datasets with low-dimensional features, they perform poorly on high-dimensional data. For high-dimensional data, unsupervised feature extraction algorithms are often used to perform cluster analysis. These algorithms map the high-dimensional data into a low-dimensional feature space and then perform cluster analysis based on the feature vectors in this low-dimensional space. However, electrical signal data is noisy and extremely large, making the features extracted using these traditional methods unrepresentative. Furthermore, these methods are algorithmically disconnected, lacking data feedback between the clustering and feature extraction algorithms, resulting in poor clustering results. Summary of the Invention

[0005] In response to the above-mentioned problems in the prior art, the inventors utilized the feature extraction capabilities of deep neural networks to improve the design of an integrated deep clustering algorithm based on variational autoencoders, and constructed an electrical signal clustering analysis method based on deep clustering variational autoencoders, which can perform autonomous clustering analysis on complex electrical signal data.

[0006] Currently, building classification models for aircraft electrical signal testing requires supervised training with labeled electrical signal data to optimize classifier performance. However, most raw data is unlabeled and requires preprocessing, such as labeling, before it can be used in actual classification tasks. Traditional data labeling often relies on manual labor from experts, a time-consuming and labor-intensive process that significantly impacts model building efficiency.

[0007] To this end, the present invention has made important substantial improvements to the clustering method based on the deep convolutional variational autoencoder network, so that the electrical signal clustering analysis method based on the deep clustering variational autoencoder of the present invention can realize unsupervised preliminary clustering of the original electrical signal data without using prior data labels, and the clustering results can be used for auxiliary labeling of electrical signals, which can effectively improve the labeling efficiency of expert data.

[0008] According to a more specific aspect, the present invention improves the variational autoencoder feature extraction model, integrates the convolutional neural network structure and the independently designed clustering structure, and proposes an electrical signal clustering analysis method based on deep clustering variational autoencoder according to the present invention. The method of the present invention has stronger feature extraction and generalization capabilities, better applicability to high-dimensional or low-dimensional signals, and demonstrates good algorithm performance.

[0009] According to one aspect of the present invention, a deep variational autoencoder for aircraft electrical signal clustering is provided, which is characterized by comprising:

[0010] Variational autoencoder module;

[0011] Multi-scale residual convolutional network module;

[0012] Signal clustering module.

[0013] The aircraft electrical signal clustering analysis based on deep variational autoencoder in the present invention overcomes the problem that high-dimensional signals are deeply troubled by noise and difficult to effectively extract features in existing aircraft signal clustering analysis. It effectively solves the problems of large data volume and strong noise in high-dimensional signals, and significantly improves the accuracy of aircraft electrical signal clustering analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 A flowchart of an electrical signal clustering analysis method based on a deep clustering variational autoencoder according to an embodiment of the present invention is shown;

[0015] Figure 2 A training flowchart of an electrical signal clustering analysis method based on a deep clustering variational autoencoder according to an embodiment of the present invention is shown. DETAILED DESCRIPTION

[0016] like Figure 1 As shown, the electrical signal clustering analysis method based on deep clustering variational autoencoder according to one embodiment of the present invention includes:

[0017] When the aircraft electrical signal is clustered and analyzed (101), the signal is first collected and transmitted (102);

[0018] The collected electrical signal is sent to a first multi-scale residual dilation convolution module (103);

[0019] Then, the maximum pooling layer (104) is combined to perform stacking. In the encoder network, the input of the network is the original signal, and after passing through the stacked feature extraction module, the feature map of the input signal is obtained;

[0020] Then, the feature map is passed through two parallel first fully connected layers (105) and second fully connected layers (106) to obtain the feature mean (107) and feature variance (108), respectively. The encoder structure is similar to that of the variational autoencoder.

[0021] In the decoder network, the reconstructed feature vector (109) is used as the input of the decoder network and passes through the third fully connected layer (110) to enter the feature reconstruction module; the feature reconstruction module includes an upsampling layer (111) and a second multi-scale residual dilation convolution module (112). The structure of the decoder is completely symmetrical with that of the encoder, including the stacking of layers and output channels; the upsampling layer (111) is the reverse operation of the maximum pooling layer (104); after passing through the feature reconstruction module, the output of the decoder is a reconstructed signal (113), which is intended to completely match the input signal of the encoder network; the decoder network also includes a clustering module after the feature reconstruction module, in which the reconstructed feature vector is subjected to feature transformation through the fourth fully connected layer (114), and the Softmax function is combined in the final layer to output the clustering result (115) in the mode of outputting the classification result.

[0022] Next, the first and second multi-scale residual dilated convolution modules (103, 112) and the variational autoencoder are described respectively.

[0023] The variational autoencoder is based on the autoencoder, and is used to constrain the obtained eigenvectors so that each value in the eigenvectors has a universal statistical meaning, thereby obtaining the algorithm part of the generative model. During the encoding process of the autoencoder, it can be assumed that each value in the eigenvector can represent a potential attribute of the original data. In the variational autoencoder algorithm, each potential attribute obeys a probability distribution. In order to enable the network to be trained smoothly in practice, each eigenvalue is assumed to obey a standard normal distribution. From a statistical point of view, it is believed that the original data X satisfies a certain probability distribution p(X) in its dimensional space, and its distribution law is difficult to calculate. The current original data sample is just a sample x obtained by sampling from this distribution law. In the variational autoencoder, it is set to express the distribution law of X by constructing an implicit variable Z, that is:

[0024]

[0025] Therefore, p(X|Z) here represents a model for generating sample X from Z. By adjusting the values ​​in the Z vector within the distribution, the generated sample X will also satisfy the distribution of the dataset X in its dimensional space. Therefore, the variational autoencoder is a generative algorithm, that is, its encoder part serves as the recognition model, and the decoder part serves as the generation model.

[0026] The variational autoencoder realizes its inherent statistical functions in network design and loss function design. During the encoding process, the network will fit two vectors of equal length, representing the mean and variance of the feature Z respectively. The encoder network structure here is the joint distribution p(X,Z) of X and Z. However, this probability distribution is difficult to solve, so the idea of ​​the network is to make p(X,Z) as close as possible to an approximate distribution Q(X,Z), where q is set to the standard normal distribution. In order to make the feature mean and feature variance satisfy the standard normal distribution, in the design of the loss function, in addition to the constraints on the reconstruction error, the loss function also considers the KL divergence between the distribution of Z and the standard normal distribution, that is:

[0027] Where μ is the characteristic mean, σ 2 is the characteristic variance, and N(0,I) is the standard normal distribution.

[0028] From a probabilistic perspective, we expand the latent variable Z of the original variational autoencoder to (Z, Y), where Y is the discrete variable output by the clustering layer, representing the output category. Therefore, its KL loss can be written as:

[0029]

[0030] (x represents "sample x obtained by sampling from the distribution law"; z represents one dimension in the latent variable Z) In order to make the loss function easier to calculate, the present invention makes a series of settings in the model:

[0031] Assume p(Z|X) is a normal distribution, and its output is the feature mean and feature variance;

[0032] Setting q(X|Z) to be a normal distribution with a constant variance is equivalent to using the mean square error as the reconstruction error;

[0033] Assume that q(Z|Y) is a normal distribution with a mean equal to the mean of the Y output and a variance of 1;

[0034] Assume q(Y) to be uniformly distributed so that the distribution of categories is balanced;

[0035] Then use the clustering module to fit p(Y|Z),

[0036] Finally, the loss function of the network is represented as:

[0037]

[0038] Among them, the meanings of the three losses are: ⑴Reconstruction loss It is used to constrain the similarity between the reconstructed signal and the input signal, and calculate the error between the input signal and the reconstructed signal through the mean square error and minimize it;

[0039] ⑵KL loss Used to constrain the distribution of features. Specifically, the KL divergence is calculated for the distribution of p(Y|Z) and q(Y) to constrain the distribution of latent vectors.

[0040] ⑶Category loss It is used to constrain the correspondence between the output clustering categories and the feature vectors, and to limit the output of the clustering results through the cross entropy error to achieve the alignment of the clustering categories and the feature vectors.

[0041] The clustered variational autoencoder of the present invention uses a deep neural network to extract deep features from high-dimensional data. The multi-scale residual dilated convolution module has the ability to extract features from complex electrical signals, and its feature extraction performance is better than that of fully connected neural networks (DNNs) and convolutional neural networks (CNNs). Therefore, the present invention introduces the multi-scale residual dilated convolution module into the clustered variational autoencoder as a major component of the encoder and decoder in the network, enabling the autoencoder network to obtain improved signal feature extraction and signal reconstruction capabilities.

[0042] The first and second multi-scale residual dilated convolution modules include the following features:

[0043] Multi-scale feature extraction branch: For high-dimensional signal features, the network is expected to capture large-scale features with a smaller number of layers while not neglecting small-scale features. To this end, drawing on the famous GoogleNet network, the present invention introduces a multi-scale feature extraction branch in the module design. The feature extraction module (multi-scale residual dilated convolution module (103) plus maximum pooling layer module (104)) includes two branches with convolution kernels of different scales:

[0044] -Small-scale feature extraction branch, whose convolution kernel size is 3, is used to perceive small-scale features;

[0045] - Large-scale feature extraction tributary, whose convolution kernel size is 5, is used to perceive large-scale features,

[0046] Through the separate actions of these two tributaries, the feature extraction module can capture signal features of different scales in parallel.

[0047] Dilated convolution: The so-called dilated convolution (or hole convolution) refers to adding holes separated by several elements (dilation rate) between each element of the convolution kernel, thereby expanding the receptive field of the convolution layer without increasing the number of convolution kernel parameters. In the present invention, the second convolution layer of the small-scale feature extraction tributary adopts convolution with a dilation rate of 2, and the first and second layers of the large-scale feature extraction tributary adopt convolution with dilation rates of 2 and 5, respectively. By adopting this technology, the receptive field of the module can be effectively expanded. When the dilated convolution technology is not adopted, in a two-layer convolutional neural network, the convolution kernel sizes of 3 and 5 can bring receptive fields of lengths of 5 and 9, respectively; when dilated convolution is adopted, taking the dilated convolution module proposed in the present invention as an example, the receptive fields can be expanded to 9 and 25, respectively.

[0048] By integrating multiple advanced neural network techniques, this feature extraction module effectively solves the problem of extracting multi-scale features from high-dimensional signals, while overcoming the vanishing gradient problem that occurs in deep networks. Furthermore, thanks to the use of dilated convolutions, the number of network parameters and complexity are effectively controlled, further improving operational efficiency. In summary, the proposed first and second multi-scale residual dilated convolution modules can effectively meet the requirements for feature extraction from high-dimensional aircraft electrical signals.

[0049] Compared with traditional variational autoencoders, the clustered variational autoencoder of the present invention has better adaptability in clustering tasks. First, the clustered variational autoencoder of the present invention transforms the dimensionality reduction algorithm into a clustering algorithm, solving the algorithmic fragmentation problem when clustering high-dimensional data in the past. That is, the features extracted by the deep autoencoder are directly related to the final output cluster category and are interpretable from a probabilistic perspective. At the same time, in view of the data structure characteristics of electrical signals, the deep clustered variational autoencoder of the present invention introduces a multi-scale residual dilation convolution module to extract features and reconstruct signals, improving the adaptability of its neural network to electrical signal data.

[0050] The algorithm flow chart of the present invention using two-stage training to train the deep clustering variational autoencoder network is as follows: Figure 2 As shown. The first stage of training (201) is performed first. First, the iteration step is initialized (202) to start a round of training (203). By inputting unlabeled training samples (204), the network reconstructs the signal (205) and outputs the clustering results. The error and loss function (206) are calculated for these results. Then, the network parameters are updated by backpropagation using the first stage learning rate α1 (207). It is determined whether the training set has been traversed (208). If not, the method returns to step (204). If it has been completed, the number of iteration steps is increased by 1 (209). Then, it is determined whether the number of rounds has reached the number of training rounds in the first stage (210). If not, the method returns to step (203). If yes, the method ends the first stage training and starts the second stage training.

[0051] In the second phase of training, the network parameters are inherited and multiple rounds of training are started. The difference is that the network parameters are updated by backpropagation using the second phase learning rate α2 (217). When the number of rounds reaches the set target (220), training ends (221). The loss function used in training is shown in Equation (4).

[0052] The advantages and benefits of the deep variational autoencoder algorithm for aircraft electrical signal clustering analysis according to the present invention include:

[0053] ⑴ Compared with traditional variational autoencoders, the deep clustering variational autoencoder of the present invention has better adaptability when completing clustering tasks. First, the clustering variational autoencoder of the present invention transforms the dimensionality reduction algorithm into a clustering algorithm, solving the problem of algorithm splitting when clustering high-dimensional data in the past, that is, the features extracted by the deep autoencoder are directly related to the clustering category of the final output, and are interpretable from a probabilistic perspective; at the same time, in view of the data structure characteristics of electrical signals, the deep clustering variational autoencoder of the present invention extracts features and reconstructs signals by introducing a multi-scale residual dilation convolution module, which has stronger adaptability to electrical signal data;

[0054] (2) The deep clustered variational autoencoder of the present invention introduces a multi-scale residual dilated convolution module to extract features from the signal. By integrating multiple advanced neural network technologies, it effectively solves the problem of extracting multi-scale features from high-dimensional signals, while overcoming the vanishing gradient problem in deep networks. At the same time, due to the use of dilated convolution, the number and complexity of the network parameters are also effectively controlled, the operating efficiency is further improved, and the feature extraction requirements of high-dimensional electrical signals of aircraft are effectively met.

[0055] ⑶ The deep clustering variational autoencoder of the present invention significantly improves the clustering accuracy of aircraft electrical signals, has stronger adaptability and robustness to electrical signal data, and makes outstanding contributions to the core fault identification of aircraft health management strategies.

Claims

1. A deep variational autoencoder for aircraft electrical signal clustering tasks, characterized by include: Variational autoencoder module; Signal clustering module, in: The variational autoencoder module includes: an encoder, a reconstruction layer and a decoder. The encoder includes: A signal acquisition and transmission module (102) is used to acquire and transmit signals (101) when cluster analysis is performed on aircraft electrical signals; A first multi-scale residual dilation convolution module (103) is used to receive the electrical signal collected by the signal collection and transmission module (102); a maximum pooling layer (104) for performing stacking in combination with the first multi-scale residual dilated convolution module (103) to obtain a feature map from the electrical signal; A first fully connected layer (105) and a second fully connected layer (106) are arranged in parallel, wherein the feature map passes through the first fully connected layer (105) and the second fully connected layer (106) in parallel, respectively obtaining a feature mean (107) and a feature variance (108); The reconstruction layer (109) is used to reconstruct the feature mean and feature variance into a feature vector as the input of the decoder network module. The decoder includes: The third fully connected layer (110), Upsampling layer (111), The second multi-scale residual dilation convolution module (112) and the upsampling layer (111) are combined with the second multi-scale residual dilation convolution module (112) to perform the reverse operation of the maximum pooling layer (104), and the feature vector enters the upsampling layer (111) through the third fully connected layer (110). A reconstruction module (113) for reconstructing the output of the decoder to fully match the input of the encoder; The signal clustering module includes: The fourth fully connected layer (114) is used to perform feature transformation on the reconstructed feature vector. The final layer (115) is used to combine the reconstructed feature vector after feature transformation of the fourth fully connected layer (114) with the Softmax function to output the clustering result. Wherein, in the first and second multi-scale residual dilated convolution modules: Assume that the original data X satisfies the probability distribution p(X) in its dimensional space, Assume that the current original data sample is the sample x obtained by sampling from the distribution p(X), and construct the implicit variable Z to represent the distribution law of X, that is: p(X|Z) represents a model that generates sample X from Z. By adjusting the value in vector Z within the distribution, the generated sample X will also satisfy the distribution of dataset X in its dimensional space. During the encoding process, the variational autoencoder module fits two vectors of equal length, representing the mean and variance of Z respectively. The network structure of the variational autoencoder module is the joint distribution p(X,Z) of X and Z. Make p(X,Z) as close as possible to an approximate distribution q(X,Z), where q is set to the standard normal distribution, In order to make the mean and feature variance of Z satisfy the standard normal distribution, the loss function not only constrains the reconstruction error but also considers the KL divergence between the distribution of Z and the standard normal distribution, that is: Where μ is the characteristic mean, σ 2 is the characteristic variance, N(0,I) is the standard normal distribution, Thus, Z is expanded to (Z, Y), where Y is the discrete variable output by the clustering layer, representing the output category, so the KL loss of Z is represented as: Where x represents the sample obtained by sampling from the distribution law of X, z represents one dimension in the latent variable Z, To make the loss function easier to calculate: Assume p(Z|X) is a normal distribution, and its output is the mean and characteristic variance of Z; Assume that q(X|Z) represents a normal distribution with constant variance; Set q(Z|Y) to a normal distribution with a mean equal to the mean of the Y output and a variance of 1; Set q(Y) to be uniformly distributed so that the distribution of categories is balanced; Use the clustering module to fit p(Y|Z) and express the loss function as: The three losses are: Reconstruction loss It is used to constrain the similarity between the reconstructed signal and the input signal, and calculate the error between the input signal and the reconstructed signal through the mean square error and minimize it; KL Loss Used to constrain the distribution of features; Class loss It is used to constrain the correspondence between the output clustering categories and the feature vectors, and to limit the output of the clustering results through the cross entropy error to achieve the alignment of the clustering categories and the feature vectors.

2. The deep variational autoencoder according to claim 1, wherein: Loss in KL During the determination, the KL divergence between the distribution p(Y|Z) and the distribution q(Y) is calculated to constrain the distribution of the latent vector.

3. The deep variational autoencoder according to claim 1, characterized in that The first and second multi-scale residual dilated convolution modules each include: Two branches with convolution kernels of different scales: -Small-scale feature extraction branch, whose convolution kernel size is 3, is used to perceive small-scale features; - Large-scale feature extraction tributary, whose convolution kernel size is 5, is used to perceive large-scale features, Through the separate effects of these two branches, the first and second multi-scale residual dilated convolution modules can capture signal features of different scales in parallel. in: The second convolutional layer of the small-scale feature extraction tributary uses a convolution with a dilation rate of 2. The first and second layers of the large-scale feature extraction branch use convolution with dilation rates of 2 and 5 respectively.

4. The deep variational autoencoder according to any one of claims 1 to 3, characterized in that The deep variational autoencoder is trained using a two-stage training process, including: First, the first stage of training (201) is carried out. First, the iteration step is initialized (202), Start a round of training (203), Input unlabeled training samples (204), Reconstruct the sample signal (205) and output the clustering result, The error and loss function (206) are calculated for the clustering result. Then the first stage learning rate α1 is used to back propagate and update the network parameters (207). Determine whether to traverse the training set (208), If “No”, return to step (204) of inputting unlabeled training samples. If "yes", the number of iteration steps increases by 1 (209). Then determine whether the number of rounds reaches the number of training rounds in the first stage (210). If "no", return to the step of starting a round of training (203). If "yes", end the first stage training and start the second stage training. In the second stage of training, the network parameters are inherited and multiple rounds of training are started. The difference is that the network parameters are updated by back propagation with the second stage learning rate α2 (217). When the number of rounds reaches the set target (220), the training ends (221). The loss function used in training is shown in formula (4).

5. An electrical signal clustering analysis method for aircraft electrical signal clustering tasks, characterized by include: Using a first multi-scale residual dilation convolution module (103), receiving the electrical signal collected by the signal collection and transmission module (102); Using a maximum pooling layer (104) in combination with a first multi-scale residual dilated convolution module (103) to perform stacking, a feature map is obtained from the electrical signal; Passing the feature map through two parallel first fully connected layers (105) and second fully connected layers (106) to obtain a feature mean (107) and a feature variance (108); Using the reconstruction layer (109), the feature mean and feature variance are reconstructed into feature vectors. Input the feature vector into the third fully connected layer (110), The feature vector is passed through the third fully connected layer (110) and enters the upsampling layer (111), The upsampling layer (111) is used in combination with the second multi-scale residual dilation convolution module (112) to perform an operation inverse to the maximum pooling layer (104). Reconstructing the output of the decoder network module to fully match the input of the encoder using a reconstruction module (113); The reconstructed feature vector is transformed using the fourth fully connected layer (114). Using the final layer (115), the reconstructed feature vector after feature transformation of the fourth fully connected layer (114) is combined with the Softmax function to output the clustering result. The first and second multi-scale residual dilated convolution modules are used to perform the following operations: Assume that the original data X satisfies the probability distribution p(X) in its dimensional space, Assume that the current original data sample is the sample x obtained by sampling from the distribution p(X), The distribution law of X is expressed by constructing the implicit variable Z, namely: Let p(X|Z) represent a model that generates sample X from Z. By adjusting the value in vector Z within the distribution, the generated sample X will also satisfy the distribution of dataset X in its dimensional space. During the encoding process, a variational autoencoder module including a signal acquisition and transmission module (102), a first multi-scale residual expansion convolution module (103), a maximum pooling layer (104), two parallel first fully connected layers (105) and a second fully connected layer (106) is used to fit two vectors of equal length, respectively representing the mean (107) and variance (108) of Z. The network structure of the variational autoencoder module is the joint distribution p(X, Z) of X and Z. Make p(X,Z) as close as possible to an approximate distribution q(X,Z), where q is set to the standard normal distribution, In order to make the mean and feature variance of Z satisfy the standard normal distribution, the loss function not only constrains the reconstruction error but also considers the KL divergence between the distribution of Z and the standard normal distribution, that is: Where μ is the characteristic mean, σ 2 is the characteristic variance, N(0,I) is the standard normal distribution, Thus, Z is expanded to (Z, Y), where Y is the discrete variable output by the clustering layer, representing the output category, so the KL loss of Z is represented as: Where x represents the sample obtained by sampling from the distribution law of X, z represents one dimension in the latent variable Z, To make the loss function easier to calculate: Assume p(Z|X) is a normal distribution, and its output is the mean and characteristic variance of Z; Assume that q(X|Z) represents a normal distribution with constant variance; Set q(Z|Y) to a normal distribution with a mean equal to the mean of the Y output and a variance of 1; Set q(Y) to be uniformly distributed so that the distribution of categories is balanced; Use the clustering module to fit p(Y|Z) and express the loss function as: The three losses are: Reconstruction loss It is used to constrain the similarity between the reconstructed signal and the input signal, and calculate the error between the input signal and the reconstructed signal through the mean square error and minimize it; KL Loss Used to constrain the distribution of features; Class loss It is used to constrain the correspondence between the output clustering categories and the feature vectors, and to limit the output of the clustering results through the cross entropy error to achieve the alignment of the clustering categories and the feature vectors.

6. The method for clustering analysis of electrical signals according to claim 5, wherein: Loss in KL During the determination, the KL divergence between the distribution p(Y|Z) and the distribution q(Y) is calculated to constrain the distribution of the latent vector.

7. The method for clustering analysis of electrical signals according to claim 5, characterized in that The first and second multi-scale residual dilated convolution modules each include: Two branches with convolution kernels of different scales: -Small-scale feature extraction branch, whose convolution kernel size is 3, is used to perceive small-scale features; - Large-scale feature extraction tributary, whose convolution kernel size is 5, is used to perceive large-scale features, Through the separate effects of these two branches, the first and second multi-scale residual dilated convolution modules can capture signal features of different scales in parallel. in: The second convolutional layer of the small-scale feature extraction tributary uses a convolution with a dilation rate of 2. The first and second layers of the large-scale feature extraction branch use convolution with dilation rates of 2 and 5 respectively.

8. The method for clustering analysis of electrical signals according to any one of claims 5 to 7, characterized in that Further comprising training using a two-stage training comprising: First, the first stage of training (201) is carried out. First, the iteration step is initialized (202), Start a round of training (203), Input unlabeled training samples (204), Reconstruct the sample signal (205) and output the clustering result, The error and loss function (206) are calculated for the clustering result. Then the first stage learning rate α1 is used to back propagate and update the network parameters (207). Determine whether to traverse the training set (208), If “No”, return to step (204) of inputting unlabeled training samples. If "yes", the number of iteration steps increases by 1 (209). Then determine whether the number of rounds reaches the number of training rounds in the first stage (210). If "no", return to the step of starting a round of training (203). If "yes", end the first stage training and start the second stage training. In the second stage of training, the network parameters are inherited and multiple rounds of training are started. The difference is that the network parameters are updated by back propagation with the second stage learning rate α2 (217). When the number of rounds reaches the set target (220), the training ends (221). The loss function used in training is shown in formula (4).

Images