Unsupervised fault detection method for transformer voiceprint signal based on feature clustering combined with generative learning

Abstract

An unsupervised detection method for power transformer voiceprint signal fault based on feature clustering combined with generative learning belongs to the technical field of power transformer fault diagnosis. A voiceprint signal sensor is placed on the transformer and the signal is extracted; a sliding window is used to slice the extracted voiceprint signal to obtain a sample set; a DAE is constructed to extract the trained bottleneck layer; a DBSCAN model is designed to cluster the bottleneck layer to obtain a sample X to be trained; the VAE is trained and the encoder part in the trained DAE is connected to the input end of the VAE to obtain a D-Vae model; the voiceprint signal to be tested is input into the trained D-Vae and a generated sequence sample is output; the generated sequence sample is compared with the real feature extraction sequence, and the 3-sigma rule is used to detect the fault. The present invention reduces the false alarm rate and the missed alarm rate of voiceprint signal fault detection, realizes unsupervised detection through feature clustering and generative learning, and improves the accuracy and robustness of fault detection.

Description

Technical Field

[0001] The present invention belongs to the technical field of power transformer fault detection, and in particular relates to an unsupervised detection method for transformer voiceprint signal faults based on feature clustering combined with generative learning. Background Art

[0002] In the power system, transformers, as one of the important equipment, play a key role in the transmission and distribution of electric energy. However, due to long-term use and environmental factors, power transformers have various potential fault risks, which may cause equipment damage, power outages and even accidents. Therefore, it is crucial to detect faults in power transformers and perform effective maintenance and repair. Traditional power transformer fault detection methods mainly rely on manual inspections or checking equipment operating status information. However, this method is usually time-consuming and cannot obtain the true status of the equipment in real time, making it difficult to detect potential fault risks in a timely manner.

[0003] In the field of power transformer fault detection, the application of voiceprint signals has attracted widespread attention from researchers in recent years. Mechanical vibration and electromagnetic induction caused by internal faults will produce specific sound characteristics, namely voiceprint signals. These voiceprint signals contain important information about the internal state of the transformer. By analyzing and processing the voiceprint signals of power transformers, the fault conditions of transformers can be effectively detected, the health status of the equipment can be accurately judged, and corresponding repair and maintenance measures can be taken to ensure the normal operation of the equipment.

[0004] The analysis process of voiceprint signals mainly includes the steps of signal acquisition, preprocessing, feature extraction and fault identification. First, the transformer voiceprint signal needs to be collected through the sensor. Then, the collected signal is preprocessed, including filtering, noise reduction, interference removal, etc., to improve the accuracy and reliability of subsequent analysis. Next, by applying signal processing and feature extraction algorithms, the fault-related features in the voiceprint signal are extracted, such as spectral features, energy features, time domain features, etc. Finally, under the guidance of the trained model, the extracted features are used for fault identification and classification to achieve automated fault detection and diagnosis.

[0005] In recent years, with the rapid development of technologies such as digital signal processing, machine learning and pattern recognition, machine learning and neural network detection methods based on supervised learning have gradually been applied to power transformer fault detection, providing new ideas for improving detection efficiency and accuracy. However, due to the complexity and variability of power transformer voiceprint signals, traditional fault detection methods often cannot accurately identify and classify various fault types. Among the current mainstream algorithms, clustering algorithms are widely used in fault identification, such as K-means algorithms, but the false alarm rate and missed alarm rate of fault signals obtained by clustering are very high, because many noise points in the signal samples cannot be well classified by clustering algorithms. Neural network methods are also widely used in voiceprint signal fault detection. The neural network is supervised by prior data and the input voiceprint signal is classified to obtain the fault signal. However, since the sample sets need to be collected in different scenarios, the training model is not suitable for fault prediction applications in a new scenario and has limitations.

[0006] Patent document CN115862678A discloses an unsupervised transformer voiceprint anomaly detection method, which uses deep learning to determine the abnormal operation state of the transformer, deploys the detection model into the transformer voiceprint processing system, extracts the transformer voiceprint data through 2.2CNN network model inference, obtains the transformer voiceprint features, and then inputs them into the trained voiceprint detection model to calculate the MSELoss loss function. According to the pre-set threshold and loss value, if the transformer voiceprint is greater than the threshold, it is an abnormal voiceprint, and if it is less than the threshold, it is normal, so as to achieve the purpose of transformer anomaly detection. However, the invention uses the threshold as a hyperparameter to determine abnormal behavior. In actual application scenarios, different scenarios need to set the optimal threshold based on experience to achieve accurate detection, and autonomous and adaptive anomaly detection in multiple scenarios cannot be achieved. The present invention uses the 3-sigma rule to effectively overcome this difficulty. There is no need to set a threshold. The abnormal points are excluded by calculating the error mean and standard deviation, which can automatically adapt to abnormal detection in different scenarios. Summary of the invention

[0007] In view of the technical problems existing in the above-mentioned prior art, the present invention proposes an unsupervised detection method for power transformer voiceprint signal faults based on feature clustering combined with generative learning, which realizes unsupervised detection of the entire process. Through unsupervised learning of power transformer voiceprint signals, the accuracy and robustness of fault detection are improved. Traditional supervised learning methods require a large amount of labeled data, but in practical applications, it is often difficult to obtain sufficient labeled samples. Unsupervised learning methods do not require labeled data, and model learning can be performed by mining the inherent distribution laws of the data. They are suitable for scenarios such as fault detection where it is difficult to obtain labeled data. At the same time, the advantages of generative learning are used to model the distribution of power transformer voiceprint signals, so as to expand the training set by generating samples and improve the generalization ability of the model. This training process can be directly used in various new scenarios with better prediction results.

[0008] The technical solution adopted by the present invention is as follows:

[0009] An unsupervised detection method for power transformer voiceprint signal fault based on feature clustering combined with generative learning, characterized in that it includes the following steps:

[0010] Step 1: Place a voiceprint signal sensor on the transformer. The sensor records the signal of each transformer. Use a readout circuit to collect and extract the signal to obtain a sample set M.

[0011] Step 2: Preprocess the sample set by using a sliding window to slice the voiceprint signal extracted by each sensor with a certain step length to obtain a sample set M';

[0012] Step 3: Construct a noisy autoencoder neural network (DAE), where DAE consists of an encoder and a decoder. The sample set M' is used as the input and output of DAE to train DAE, and the trained DAE model is obtained. The encoder part is extracted, and the extracted feature sequence sample set X' is output, that is, the bottleneck layer vector of the trained DAE is extracted.

[0013] Step 4: Design a DBSCAN model to cluster the extracted feature sequence sample set X' to obtain the sample X to be trained;

[0014] Step 5: Build a VAE model, use the training sample X as the input and output of the VAE model, train the VAE, and obtain the trained VAE model;

[0015] Step 6: Connect the encoder part of the trained DAE model and the VAE model together, and use the output of the encoder of the DAE model as the input of the VAE to form the trained D-Vae model; input the voiceprint signal to be tested into the trained D-Vae model, and output the generated sequence sample; Step 7: Compare the generated sequence sample with the real feature sample output by the encoder of the trained DAE model, calculate the anomaly score, and obtain the anomaly score set R;

[0016] Step 8. Use the 3-sigma rule to eliminate abnormal points in the abnormal score set R, and extract new voiceprint signals to be tested after 2 to 4 cycles, and input them into the trained D-Vae model to obtain a new abnormal score set. Use the 3-sigma rule to eliminate abnormal points so that all fault voiceprint signals can be detected as much as possible. When new faults exist during the operation of the transformer, they can also be detected in time, resulting in lower false alarm and missed alarm rates.

[0017] Furthermore, the sample set M obtained in step 1 = {m1,m2,..,m num}, where num is the number of samples, ranging from 100 to 3000, and m i =(m i 1 ,m i 2 ,..,m i L ), i = 1, 2, ..., num, is the time series of the i-th sample, where L is the length of the sequence, ranging from 50 to 500;

[0018] Furthermore, in step 2, a sliding window is used to slice the voiceprint signal extracted by each sensor with a certain step size. The specific steps are as follows:

[0019] Step 2.1, define the length range of the window as win, with a value range of 10 to 300;

[0020] Step 2.2, set the starting position of the sliding window to the starting position of the sequence;

[0021] Step 2.3: Take the first sample time series of M samples, and start from the beginning of the sequence by sliding the window size, and gradually move the window, and cut each window out. The moving step of the window is 2, which means that 2 data points are moved each time, and the sliced ​​sample M_win of the first sample time series is obtained. 1 ={M_win1 1 ,M_win2 1 …M_win nw 1}, where nw is the number of sequences obtained after sliding window slicing, M_wini 1 ={ss i 1 (1),ss i 1 (2),…,ss i 1 (win)}, i = 1, 2, ..., nw, is the i-th sequence obtained by slicing the first sample, where the length of the sequence is win, i.e., the size of the window;

[0022] Step 2.4, repeat step 2.3, take the 2nd to num sample time series, and finally slice the sample composition to obtain the training sample set M' = {M_win 1 ,M_win 2 …M_win num}.

[0023] Furthermore, in step 3, a noisy autoencoder neural network (DAE) is constructed, the middle layer is the bottleneck layer, from the input layer to the bottleneck layer is the DAE encoder, and from the bottleneck layer to the output layer is the DAE decoder. The specific structure is:

[0024] 1 input layer, the number of neurons is the length of the sliced ​​sequence win;

[0025] The number of hidden layers ranges from 3 to 11;

[0026] The number of neurons ranges from 5 to 200;

[0027] 1 output layer, the number of neurons is the length of the sliced ​​sequence win;

[0028] Furthermore, in step 4, a DBSCAN model is designed to cluster the extracted feature sequence sample set X'. The specific steps are:

[0029] Step 4.1, initialize parameters, define the density threshold ε, the value range is 0.1-0.8 and the minimum number of samples MinPts, the value range is 50-500, and form a DBSCAN clustering model;

[0030] Step 4.2: Input the sample set X', and after DBSCAN clustering, obtain the core points, cluster points and noise points;

[0031] Step 4.3, remove the noise points, extract the cluster points and use them as the training samples X;

[0032] Furthermore, in step 5, a VAE model is constructed, and the training sample X is used as the input and output of the VAE model. The VAE is trained to obtain a trained VAE model. The specific structure is:

[0033] 1 input layer, the number of neurons ranges from 5 to 200;

[0034] The number of hidden layers ranges from 3 to 11;

[0035] The number of neurons ranges from 5 to 200;

[0036] 1 output layer, the number of neurons ranges from 5 to 200;

[0037] Furthermore, the sequence samples generated in step 7 are compared with the real feature samples output by the encoder of the trained DAE model, and the anomaly score is calculated to obtain the anomaly score set R;

[0038] The calculation formula of the anomaly score is:

[0039]

[0040] Among them, y gen is the generated vector output by the D-Vae model, y true is the true vector of the sample, n_er is the length of the sequence sample, that is, the number of features extracted by the bottleneck layer in DAE;

[0041] y_gen i ,i=1,2…,n_er is the i-th value in the generated vector output by the D-Vae model, y_true i ,i=1,2…,n_er is the i-th value in the sample true vector. Finally, all anomaly scores ab form the anomaly score set R.

[0042] Furthermore, the specific steps of the 3-sigma rule in step 8 are:

[0043] Step 8.1, calculate the mean and standard deviation of the anomaly score set R;

[0044] Step 8.2, traverse all ab values ​​of the abnormal score set A. When the ab value is greater than the average plus 3 times the standard deviation or less than the average minus 3 times the standard deviation, determine that the sequence corresponding to the ab value is an abnormal sequence, that is, an abnormal voiceprint signal. Use the 3-sigma rule to eliminate abnormal points so that all fault voiceprint signals can be detected as much as possible;

[0045] Step 8.3: When a new fault occurs during the operation of the transformer, it can be detected in time, resulting in a lower false alarm rate and missed alarm rate.

[0046] The beneficial effects of the present invention are:

[0047] The present invention proposes an unsupervised detection method for power transformer voiceprint signal faults based on feature clustering combined with generative learning. The method uses a feature clustering algorithm to divide the power transformer voiceprint signal and clusters signals with similar features together to obtain different features of fault signals and normal signals. By comparing and analyzing signals of different categories, key feature vectors representing fault features can be effectively extracted.

[0048] Then, the generative learning method is used to build a fault detection model by learning the distribution law of fault features. As an innovative fault detection method, this method realizes unsupervised fault detection through feature clustering and generative learning of power transformer voiceprint signals, thus improving the accuracy and robustness of fault detection.

[0049] In practical applications, the unsupervised learning characteristics of this method are more practical and are expected to have important application value in the field of power transformer fault detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 The training process of the unsupervised detection model of power transformer voiceprint signal fault based on feature clustering combined with generative learning in Example 1;

[0051] Figure 2 This is the detection process of the faulty transformer voiceprint signal using the D-Vae model proposed in Example 1. DETAILED DESCRIPTION

[0052] The implementation process of the present invention is described in detail below in conjunction with the accompanying drawings, but this should not limit the protection scope of the present invention.

[0053] See also Figure 1 , Figure 1 The training process of the unsupervised detection model of power transformer voiceprint signal fault based on feature clustering combined with generative learning is demonstrated. First, the voiceprint signal is sent to the DAE model for training. After the trained model is obtained, the feature vector is extracted and clustered. The clustering result is used as a training set to send it to the VAE model for training. As shown in the figure, an unsupervised detection method of power transformer voiceprint signal fault based on feature clustering combined with generative learning includes the following steps:

[0054] Step 1: Place a voiceprint signal sensor on the transformer. The sensor records the signal of each transformer. Use the readout circuit to collect and extract the signal to obtain a sample set M; where the sample set M = {m1,m2,..,m num}, where num is the number of samples, the value is 100, m i =(m i 1 ,mi 2 ,..,m i L ), i = 1, 2, ..., num, is the time series of the i-th sample, where L is the length of the sequence, and its value range is 159;

[0055] Step 2: Preprocess the sample set, use a sliding window, and slice the voiceprint signal extracted by each sensor with a certain step size. The steps are as follows:

[0056] Step 2.1, define the length range of the window as win, with a value of 100;

[0057] Step 2.2, set the starting position of the sliding window to the starting position of the sequence;

[0058] Step 2.3: Take the first sample time series of M samples, and start from the beginning of the sequence by sliding the window size, and gradually move the window, and cut each window out. The moving step of the window is 2, which means that 2 data points are moved each time, and the sliced ​​sample M_win of the first sample time series is obtained. 1 ={M_win1 1 ,M_win2 1 …M_win nw 1}, where nw is the number of sequences obtained after sliding window slicing, and its value is 3000, M_win i 1 ={ss i 1 (1),ss i 1 (2),…,ss i 1 (win)}, i = 1, 2, ..., nw, is the i-th sequence obtained by slicing the first sample, where the length of the sequence is win, i.e., the size of the window;

[0059] Step 2.4, repeat step 2.3, take the 2nd to num = 100 sample time series, and finally slice the sample composition to obtain the training sample set M' = {M_win 1 ,M_win 2 …M_win num}.

[0060] Step 3: Construct a noisy autoencoder neural network (DAE), where DAE consists of an encoder and a decoder. The middle layer is the bottleneck layer. The DAE encoder is from the input layer to the bottleneck layer, and the DAE decoder is from the bottleneck layer to the output layer. The specific structure is:

[0061] 1 input layer, the number of neurons is the length of the sliced ​​sequence win = 100;

[0062] 5 hidden layers, the number of neurons is 80, 60, 50, 60, 80 respectively;

[0063] 1 output layer, the number of neurons is the length of the sliced ​​sequence win = 100;

[0064] During the training process, the sample set S' is used as the input and output of DAE to train DAE and obtain the trained DAE model, in which the hyperparameter selection is obtained through grid search and 10-fold cross validation. Finally, the activation function of the hidden layer is the linear rectifier function (ReLu), the input layer and the output layer have no activation function, the back propagation optimizer is the Adam optimizer, the training batch is 512, and the number of training times is 3000.

[0065] Extract the encoder part and output the extracted feature sequence sample set X', which is the bottleneck layer vector of the DAE after training.

[0066] Step 4: Design a DBSCAN model to cluster the extracted feature sequence sample set X', specifically:

[0067] Step 4.1, initialize parameters, define the density threshold ε = 0.6, sample number MinPts = 200 hyperparameters, and form a DBSCAN clustering model;

[0068] Step 4.2: Input the sample set X', and after DBSCAN clustering, obtain the core points, cluster points and noise points;

[0069] Step 4.3, remove the noise points, extract the cluster points and use them as the training samples X;

[0070] Step 5: Build a VAE model. The training sample X is used as the input and output of the VAE model to train the VAE. The specific structure is:

[0071] 1 input layer with 50 neurons;

[0072] 5 hidden layers, the number of neurons is 80, 80, 40, 80, 80;

[0073] 1 output layer with 50 neurons;

[0074] The trained VAE model was obtained, in which the hyperparameters were selected through grid search and 10-fold cross validation. Finally, the activation function of the hidden layer was the linear rectifier function (ReLu), the input layer and the output layer had no activation function, the back propagation optimizer was the Adam optimizer, the training batch was 512, and the number of training times was 3000.

[0075] Step 6: Connect the encoder part of the trained DAE model and the VAE model together, and use the output of the encoder of the DAE model as the input of the VAE to form the trained D-Vae model; input the voiceprint signal to be tested into the trained D-Vae model, and output the generated sequence sample;

[0076] Step 7: Generate sequence samples and compare them with the real feature samples output by the encoder of the trained DAE model, calculate the anomaly score, and obtain the anomaly score set R. The calculation formula is:

[0077]

[0078] Among them, y gen is the generated vector output by the D-Vae model, y true is the true vector of the sample, n_er is the length of the sequence sample, that is, the number of features extracted by the bottleneck layer in DAE; y_gen i ,i=1,2…,n_er is the i-th value in the generated vector output by the D-Vae model, y_true i ,i=1,2…,n_er is the i-th value in the sample true vector. Finally, all anomaly scores ab form the anomaly score set R.

[0079] Step 8: Use the 3-sigma rule to eliminate abnormal points in the abnormal score set R. The specific steps are as follows:

[0080] Step 8.1, calculate the mean and standard deviation of the anomaly score set R;

[0081] Step 8.2, traverse all ab values ​​of the abnormal score set A. When the ab value is greater than the average plus 3 times the standard deviation or less than the average minus 3 times the standard deviation, the sequence corresponding to the ab value is determined to be an abnormal sequence, that is, an abnormal voiceprint signal. Use the 3-sigma rule to eliminate abnormal points so that the fault voiceprint signals can be detected as much as possible. When new faults exist during the operation of the transformer, they can also be detected in time, resulting in a lower false alarm rate and missed alarm rate.

[0082] After 2 to 4 cycles, a new voiceprint signal to be tested is extracted and input into the trained D-Vae model to obtain a new set of anomaly scores. The 3-sigma rule is used to eliminate anomalies so that all fault voiceprint signals can be detected as much as possible. When new faults occur during the operation of the transformer, they can also be detected in time, resulting in lower false alarm and missed alarm rates.

[0083] Figure 2The detection process of the faulty transformer voiceprint signal by the proposed D-Vae model is demonstrated. The voiceprint signal to be tested is sent into the D-Vae model, the anomaly score is calculated based on the reconstructed vector and the true vector, and the 3-sigma rule is used to detect abnormal sequences.

[0084] The following table shows the performance indicators of the model detection in the samples to be tested. TP is a true positive sample. It indicates the number of positive samples correctly marked as positive samples by the model. FP is a false positive sample, which indicates the number of negative samples incorrectly marked as positive samples by the model. TN is a true negative sample, which indicates the number of negative samples correctly marked as negative samples by the model. FN is a false negative sample, which indicates the number of positive samples incorrectly marked as negative samples by the model. The formulas for accuracy acc, precision prec, false positive rate FRR, and false negative rate MRR are as follows:

[0085] acc=(TP+TN) / (TP+FP+TN+FN)

[0086] prec=TP / (TP+FP)

[0087] FRR=FP / TN+FP

[0088] MRR=FN / FN+TP

[0089] Table 1. Performance indicators

[0090]

[0091] It can be seen that the results of the invention show a high-accuracy abnormal signal detection capability, reaching an accuracy of 98.6%. At the same time, it has a high precision rate of 99.6%. The false alarm rate and missed alarm rate are relatively low, 1.34% and 1.42% respectively, indicating that the model has fewer misjudgments in judging normal data and can accurately distinguish normal data from abnormal data. In addition, the model can capture most of the abnormal data and rarely misses real abnormal situations. Therefore, it has efficient detection capabilities and is highly applicable in actual scenarios.

Claims

1. An unsupervised detection method for power transformer voiceprint signal fault based on feature clustering combined with generative learning, characterized in that: The following steps are involved: S1. Place a voiceprint signal sensor on the transformer, record the voiceprint signal of each transformer on the spot through the voiceprint signal sensor, use the readout circuit to collect and extract the signal, and obtain a sample set M; S2, preprocessing the sample set M, using a sliding window to slice the voiceprint signal extracted by each sensor with a certain step length to obtain a training sample set M'; S3, construct a noisy autoencoder neural network DAE, the DAE consists of an encoder and a decoder, use the sample set M' as the input and output of the DAE to train the DAE, obtain the trained DAE model, extract the encoder part, output the extracted feature sequence sample set X', as the extracted bottleneck layer vector of the trained DAE; S4. Design a DBSCAN clustering model to cluster the extracted feature sequence sample set X' to obtain the sample X to be trained; S5. Build a VAE model, use the training sample X as the input and output of the VAE model, train the VAE, and obtain the trained VAE model; S6. Connect the encoder part of the trained DAE model and the trained VAE model together, and use the output of the encoder of the DAE model as the input of the VAE to form the trained D-Vae model; input the voiceprint signal to be tested into the trained D-Vae model, and output the generated sequence sample; S7, comparing the generated sequence sample with the real feature sample output by the encoder of the trained DAE model, calculating the anomaly score, and obtaining an anomaly score set R; S8. Use the 3-sigma rule to eliminate abnormal points in the abnormal score set R, and extract new voiceprint signals to be tested after 2 to 4 cycles, and input them into the trained D-Vae model to obtain a new abnormal score set. Use the 3-sigma rule to eliminate abnormal points so that all fault voiceprint signals can be detected as much as possible. When new faults exist during the operation of the transformer, they can also be detected in time, resulting in lower false alarm and missed alarm rates.

2. According to claim 1, the unsupervised detection method of power transformer voiceprint signal fault based on feature clustering combined with generative learning is characterized in that: The sample set M in S1 is {m1,m2,..,m num }, where num is the number of samples, ranging from 100 to 3000, and m i =(m i 1 ,m i 2 ,..,m i L ),i=1,2,…,num,is the time series of the ith sample, where L is the length of the sequence, ranging from 50 to 500.

3. According to claim 1, the unsupervised detection method of power transformer voiceprint signal fault based on feature clustering combined with generative learning is characterized in that: The step S2 is specifically as follows: S21, define the length range of the window as win, with a value range of 10 to 300; S22, setting the starting position of the sliding window to the starting position of the sequence; S23, take the first sample time series in the sample set M, take the sliding window size as the unit, start from the starting position of the sequence, gradually move the sliding window, and cut each sliding window out. The moving step of the sliding window is 2, which means that 2 data points are slid each time, and the sliced ​​sample M_win of the first sample time series is obtained. 1 ={M_win1 1 ,M_win2 1 …M_win nw 1 }, where nw is the number of sequences obtained after sliding window slicing, M_win i 1 ={ss i 1 (1),ss i 1 (2),…,ss i 1 (win)}, i = 1, 2, ..., nw, is the i-th sequence obtained by slicing the first sample, where the length of the sequence is win, i.e., the size of the window; S24, repeat step 2.3, take the 2nd to numth sample time series, and finally slice the sample composition to obtain the training sample set M'={M_win 1 ,M_win 2 …M_win num }.

4. According to claim 1, the unsupervised detection method of power transformer voiceprint signal fault based on feature clustering combined with generative learning is characterized in that: The structure of each layer of the noisy autoencoder neural network DAE described in S3 includes 1 input layer, the number of hidden layers ranges from 3 to 11, and 1 output layer; the number of neurons includes 1 input layer, the number of neurons is the length of the sequence after slicing win; the number of neurons in the hidden layer ranges from 5 to 200; the number of neurons in 1 output layer is the length of the sequence after slicing win.

5. According to claim 1, the unsupervised detection method of power transformer voiceprint signal fault based on feature clustering combined with generative learning is characterized in that: The step S4 is specifically as follows: Step 4.1, initialize parameters, define the density threshold ε and the minimum number of samples MinPts hyperparameters, and form the DBSCAN clustering model; Step 4.2: Input the sample set X' and obtain the core points, clustering points and noise points after the DBSCAN clustering model; Step 4.3: Remove the noise points, extract the cluster points and use them as the training samples X.

6. The unsupervised detection method for power transformer voiceprint signal fault based on feature clustering combined with generative learning according to claim 5 is characterized in that: The range of the hyperparameters: the density threshold ε is defined as having a value interval of 0.1 to 0.8; the minimum number of samples MinPts is defined as having a value interval of 50 to 500.

7. The unsupervised detection method for power transformer voiceprint signal fault based on feature clustering combined with generative learning according to claim 1 is characterized in that: The structure of each layer of the VAE model described in S5 is: 1 input layer, the number of hidden layers ranges from 3 to 11, and 1 output layer; the number of neurons is: the number of neurons in 1 input layer ranges from 5 to 200; the number of neurons in the hidden layer ranges from 5 to 200; the number of neurons in 1 output layer ranges from 5 to 200.

8. The unsupervised detection method for power transformer voiceprint signal fault based on feature clustering combined with generative learning according to claim 1 is characterized in that: In step S7, the anomaly score is calculated using the following formula: Among them, y gen is the generated vector output by the trained D-Vae model, y true is the true vector of the sample, n_er is the length of the sequence sample, that is, the number of features extracted by the bottleneck layer in DAE; y_gen i ,i=1,2…,n_er is the i-th value in the generated vector output by the D-Vae model, y_true i ,i=1,2…,n_er is the i-th value in the sample true vector; Finally, all anomaly scores ab constitute the anomaly score set R.

9. The unsupervised detection method for power transformer voiceprint signal fault based on feature clustering combined with generative learning according to claim 1 is characterized in that: The specific steps of the 3-sigma rule in step S8 are: Step 8.1, calculate the mean and standard deviation of the anomaly score set R; Step 8.2, traverse all ab values ​​of the abnormal score set A. When the ab value is greater than the average plus 3 times the standard deviation or less than the average minus 3 times the standard deviation, determine that the sequence corresponding to the ab value is an abnormal sequence, that is, an abnormal voiceprint signal. Use the 3-sigma rule to eliminate abnormal points so that all fault voiceprint signals can be detected as much as possible; Step 8.3: When a new fault occurs during the operation of the transformer, it can be detected in time, resulting in a lower false alarm rate and missed alarm rate.

Images