An unsupervised fault detection method for power transformer voiceprint signals

Abstract

The present invention discloses an unsupervised detection method for the acoustic fingerprint signal faults of power transformers, which is applied to the technical field of power transformer fault diagnosis. The method includes: acquiring the acoustic fingerprint signal data of each power transformer by using sensors to form a sample set, using the density peak algorithm to find the density center of the sample set and using the K-nearest neighbor clustering algorithm for clustering to obtain a training sample set, putting it into an autoencoder network to extract a low-dimensional sparse representation sequence sample set, performing preprocessing, using the first part of the sequence as the input of the gated recurrent unit (GRU) and the second part as the output for training; the acoustic fingerprint signal to be measured passes through the prediction sequence and the GRU target output in sequence to obtain an abnormal score set, and the 3-sigma rule is used to process the abnormal score set, so as to detect faults. The present invention effectively improves the accuracy of fault detection and obtains a lower false alarm rate and missed alarm rate.

Description

Technical Field

[0001] The present invention relates to the technical field of power transformer fault detection, and more specifically, to an unsupervised detection method for power transformer acoustic signal faults. Background Art

[0002] The power transformer is one of the important devices in the power system, widely used in power transmission and distribution systems, and plays the roles of voltage transformation, voltage regulation, power line isolation, and power equipment protection. However, due to the influence of long-term operation and various external factors, there is a risk of faults in power transformers, and these faults may cause the equipment to malfunction or even cause accidents, seriously affecting the stable operation of the power system.

[0003] Currently, traditional power transformer fault detection methods mainly rely on manual inspections and regular instrument detections. This method requires a large amount of manpower and material resources, and cannot monitor the status of the equipment in real time, so it is impossible to carry out true fault warnings. In addition, due to the complex electromagnetic field and thermal field environments inside the power transformer, fault signals are often masked by background noise, making fault detection more difficult. The acoustic signal is one of the important indicators for power transformer fault detection, and the presence of abnormalities in the equipment can be judged by analyzing the acoustic characteristics. The acoustic signal refers to the sound signal generated when the power transformer is operating. Due to equipment vibrations or internal arcs caused by faults, sounds with specific frequencies and amplitudes will be generated. Therefore, by analyzing the acoustic signals of the transformer, the types and locations of faults can be accurately identified, and fault diagnosis and condition monitoring of the equipment can be realized.

[0004] Research on power transformer acoustic signal fault detection mainly focuses on two directions: feature extraction and fault identification. Feature extraction refers to extracting parameters or features that can represent the sound characteristics by mathematically processing and analyzing the acoustic signal. Commonly used features include frequency domain features (such as energy, amplitude spectrum, etc.), time domain features (such as duration, zero-crossing rate, etc.), and wavelet transform features. Fault identification uses methods such as machine learning and deep learning to match the extracted features with fault patterns to identify abnormal signals, thereby realizing automatic fault detection. However, current fault diagnosis methods often require a large number of manually labeled samples for supervised learning, which limits their flexibility and scalability in practical applications.

[0005] The mainstream technologies for fault diagnosis of transformer acoustic fingerprint signals currently include methods based on machine learning and deep learning. Machine learning methods such as the K-nearest neighbor algorithm (K-means) based on clustering, principal component analysis (PCA) algorithm, random forest (RF), and support vector machine (SVM) have been applied to fault diagnosis. However, these methods often rely on the quality of feature extraction. Before using these methods for classification diagnosis, feature extraction is required. The method of manually extracting features such as peak-to-peak value, variance, and spectrum cannot be well applied to all scenarios, and the manual extraction method is complex and cumbersome. Some scholars have proposed methods such as Fourier transform or wavelet transform for feature extraction. However, these methods require manually setting different applicable wavelet basis function parameters for different scenarios, and the generalization ability is weak. When there is noise in the signal, machine learning methods cannot well learn the data distribution characteristics, and the non-linear fitting effect on noisy signals is poor, resulting in poor performance in the test set. Deep learning is increasingly applied to solve complex non-linear fitting problems. For transformer fault classification, some people have proposed neural network methods to learn from existing datasets, detect unknown samples, and classify fault signals. However, in actual situations, we cannot obtain acoustic fingerprint signals in advance. Therefore, many neural network methods based on supervised learning with prior data do not have generalization. The long short-term memory neural network (LSTM), as a variant of the recurrent neural network, can better handle the problem of dependencies in long sequences and can detect abnormal long sequences in transformers. However, due to the complexity of its own structure and the characteristics of loops in training, it is difficult to be hardwareized. The present invention uses an improved LSTM network for sequence prediction, which can better solve the problem of difficult application.

[0006] Therefore, it is an urgent problem for those skilled in the art to propose an unsupervised detection method for transformer acoustic fingerprint signal faults to solve the difficulties existing in the prior art. Summary of the Invention

[0007] In view of this, the present invention provides an unsupervised detection method for transformer acoustic fingerprint signal faults to solve the technical problems existing in the prior art.

[0008] In order to achieve the above object, the present invention provides the following technical solutions:

[0009] An unsupervised detection method for transformer acoustic fingerprint signal faults includes the following steps:

[0010] S1. Use a sensor to collect each transformer acoustic fingerprint signal, extract and collect the time-series-based acoustic fingerprint signal through a reading circuit connected to the sensor, and form a sample set S;

[0011] S2. Define a window and use the window to perform sliding slicing on each sequence sample in the sample set S to obtain the sliced sample set S'.

[0012] S3. Apply the density peak algorithm to the sliced sample set S' to obtain a sequence as the density center.

[0013] S4. Use the sequence obtained in S3 as the center point and use the K-nearest neighbor clustering method to find the K nearest neighbors of the center point. The neighbors and the center point together form the training sample set X.

[0014] S5. Construct an AE-GRUNN model composed of an autoencoder network AEN, a gated recurrent unit GRU, and a fully connected neural network NN. Use the training sample set X as the input and output of the autoencoder network for training to obtain the trained autoencoder. Extract the encoder part and output the low-dimensional sparse representation sequence sample set X', that is, extract the bottleneck layer vector of the autoencoder network.

[0015] S6. Use the low-dimensional sparse representation sequence sample set X' as the input and target output of the GRUNN model composed of the gated recurrent unit GRU and the fully connected neural network NN, and train the constructed GRUNN model to obtain the trained GRUNN model.

[0016] S7. Input the low-dimensional sparse representation sequence sample set X' into the GRUNN model and output the prediction sequence.

[0017] S8. Calculate the anomaly score scr value through the prediction sequence and the target output of the GRUNN model to obtain the anomaly score set A of all samples in the sample to be tested.

[0018] S9. Use the 3-sigma rule to exclude the anomaly points in the anomaly score set A, extract the new sample to be tested voiceprint signal in 2-4 cycles, and repeat S5-S6 to obtain a new anomaly score set.

[0019] Optionally, S1 is specifically:

[0020] The sample set S = {s1, s2,.., s L , 2 , 1 , i ,

[0021] ,

[0020] ,

[0019] , i , i , i , num , , , }, where num is the number of samples, and the value range is 200 - 3000. s i = (s i 1 , s i 2 ,.., s i L ), i = 1, 2,..., num, is the time series of the i-th sample, and L is the length of the sequence, and the value range is 50 - 500.

[0021] Optionally, S2 is specifically:

[0022] S21. Define the length range of the window as win = 30 - 300;

[0023] S22. Set the starting position of the sliding window as the starting position of the sequence;

[0024] S23. Take the first sample time series of the sample set S. Starting from the starting position of the sequence, gradually move the sliding window in units of the sliding window size, and cut out each sliding window;

[0025] S24. Repeat S23, take the 2nd to num-th sample time series, and finally the samples after slicing form the training sample set S’ = {S_win 1 , S_win 2 …S_win num}.

[0026] Optionally, S3 is specifically as follows:

[0027] S31. Define S’ = {S_win 1 , S_win 2 …S_win num} = {S_t1, S_t2, …, S_t N}, where N is the total number of sequences obtained by slicing the num sequences, S_t i = {S_t i 1 , S_t i 2 , …, S_t i win}, i = 1, 2, …, N, is the i-th sequence in all sequences, where win is the sequence length, that is, the length of the window;

[0028] S32. Calculate the Euclidean distance d ij , i = 1, 2, …, N, j = 1, 2, …, N:

[0029] d ij = ||S_t i - S_t j ||2

[0030] Where, d ij is the Euclidean distance, S_t i is the i-th sequence in the sequences obtained after slicing, with a length of win, and S_t j is the j-th sequence in the sequences obtained after slicing, with a length of win;

[0031] S33. Calculate the local density of each sequence:

[0032]

[0033]

[0034] Among them, the truncation distance d c is the truncation distance, and as a hyperparameter d c takes the value of ρ i , when the average value of ρ for i = 1, 2,..., N is 0.2 * N, the range is 0.1 to 0.8, and d ij is the Euclidean distance, ρ i is the local density, and χ is the judgment function;

[0035] S34. Calculate the local density, and take the sequence sample with the largest local density as the density center. The density center sequence is S_t po :

[0036] po = arg maxρ i

[0037] Among them, po is the serial number of the sequence with the largest local density, and argmaxρ i is the value of i when ρ i takes the maximum value, and max(ρ i ) is the maximum value of the local density.

[0038] Optionally, S5 is specifically:

[0039] The number of neurons in the input layer of the autoencoder network is the length win of the sequences in the training sample set X. The number of neurons in the hidden layer ranges from 10 to 500, and the number of neurons in the bottleneck layer ranges from n = 12 to 256.

[0040] Optionally, in S8, calculate the abnormal score scr value, and the formula is:

[0041]

[0042] Among them, y pre is the GRUNN prediction vector, y true is the true vector of the sample, n_ner is the length of the predicted sequence, that is, the number of NN neurons; y_pre i , i = 1, 2…, n_er is the value at the i-th time point in the GRUNN predicted sequence sample, and y_true i , i = 1, 2…, n_er is the true value at the i-th time point in this sequence sample, n_er = n_ner, E is the average value calculation function, and T is the matrix transpose symbol.

[0043] Optionally, the specific steps of the 3-sigma rule in S9 are:

[0044] S91. Calculate the mean and standard deviation of the abnormal score set A;

[0045] S92. Calculate all scr values of the abnormal score set A. When the scr value is greater than the mean plus 3 times the standard deviation or less than the mean minus 3 times the standard deviation, determine the sequence corresponding to the scr value as an abnormal sequence, that is, an abnormal voiceprint signal.

[0046] As can be seen from the above technical solutions, compared with the prior art, the present invention discloses a method for unsupervised detection of power transformer voiceprint signal faults, and its beneficial effects are as follows: By using the idea of the clustering algorithm, the model can automatically find the training set without manual definition, thereby realizing unsupervised training. And using the data set obtained by density clustering as the training set can better reveal the characteristics of the sequence, so it is more suitable for the training of the subsequent network; Design an autoencoder network AEN combined with a gated recurrent unit GRU and a fully connected neural network model. The features extracted by the autoencoder network AEN can better reveal the change characteristics of the sequence. This method avoids the disadvantage of separately defining extraction functions for data with different characteristics; Finally, design a simple-structured AE-GRUNN model with better prediction effect and lower false alarm rate and missed alarm rate. The whole process realizes an unsupervised model learning method and a simple prediction network model structure, which can be better hardwareized and applied in actual scenarios in the future. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only the embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained according to the provided drawings without creative efforts.

[0048] Figure 1 It is a flowchart of a method for unsupervised detection of power transformer voiceprint signal faults provided by the present invention;

[0049] Figure 2 It is a structural diagram of the AE-GRUNN detection model provided by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0050] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0051] See Figure 1As shown in the figure, the present invention discloses an unsupervised detection method for power transformer acoustic fingerprint signal faults, including the following steps:

[0052] S1. Use a sensor to collect each power transformer acoustic fingerprint signal, extract and collect the time-sequence-based acoustic fingerprint signal through a reading circuit connected to the sensor, and form a sample set S.

[0053] S2. Define a window, and use the window to perform sliding slicing processing on each sequence sample in the sample set S to obtain the sliced sample set S'.

[0054] S3. Apply the density peak algorithm to the sliced sample set S' to obtain a sequence as the density center.

[0055] S4. Take the sequence obtained in S3 as the center point, and use the K-nearest neighbor clustering method to find the K nearest neighbors of the center point. The neighbors and the center point jointly form the training sample set X.

[0056] S5. Construct an AE-GRUNN model composed of an autoencoder network AEN, a gated recurrent unit GRU, and a fully connected neural network NN. Use the training sample set X as the input and output of the autoencoder network for training to obtain the trained autoencoder. Extract the encoder part and output the low-dimensional sparse representation sequence sample set X', that is, extract the bottleneck layer vector of the autoencoder network.

[0057] S6. Use the low-dimensional sparse representation sequence sample set X' as the input and target output of the GRUNN model composed of the gated recurrent unit GRU and the fully connected neural network NN, and train the constructed GRUNN model to obtain the trained GRUNN model.

[0058] S7. Input the low-dimensional sparse representation sequence sample set X' into the GRUNN model and output the prediction sequence.

[0059] S8. Calculate the anomaly score scr value through the prediction sequence and the target output of the GRUNN model to obtain the anomaly score set A of all samples in the sample to be tested.

[0060] S9. Use the 3-sigma rule to exclude the anomaly points in the anomaly score set A, extract the new acoustic fingerprint signal to be tested in 2-4 cycles, and repeat S5-S6 to obtain a new anomaly score set.

[0061] Specifically, the autoencoder network in S5 is composed of an encoder and a decoder.

[0062] Specifically, the 3-sigma rule is used to exclude abnormal points, so that all fault voiceprint signals can be detected as much as possible. When there are new faults during the operation of the transformer, they can be detected in time, resulting in a lower false alarm rate and missed alarm rate.

[0063] Further, S1 is specifically as follows:

[0064] The sample set S = {s1, s2,.., s num}, where num is the number of samples, and the value range is 200 - 3000. s i = (s i 1 , s i 2 ,.., s i L ), i = 1, 2,..., num, is the time series of the i-th sample, and L is the length of the sequence, with the value range of 50 - 500.

[0065] Further, S2 is specifically as follows:

[0066] S21. Define the window length range as win = 30 - 300;

[0067] S22. Set the starting position of the sliding window as the starting position of the sequence;

[0068] S23. Take the first sample time series of the sample set S. Starting from the starting position of the sequence, move the sliding window step by step with the size of the sliding window as the unit, and cut out each sliding window;

[0069] S24. Repeat S23, take the 2nd to num-th sample time series, and finally the samples after slicing form the training sample set S' = {S_win 1 , S_win 2 …S_win num}.

[0070] Specifically, the moving step of the sliding window in S23 is 1, indicating that each time 1 data point is slid, and the sliced samples S_win 1 = {S_win1 1 , S_win2 1 , …S_win nw 1} are obtained for the first sample time series, where nw is the number of sequences obtained after slicing the sliding window, and S_win i 1 = {ss i 1 (1), ss i 1 (2), …, ss i1 (win)}, i = 1, 2, ..., nw, is the i-th sequence obtained by slicing the first sample, where the length of the sequence is win, that is, the size of the window.

[0071] Furthermore, S3 is specifically:

[0072] S31, define S' = {S_win 1 , S_win 2 …S_win num}={S_t1,S_t2,…,S_t N},That

[0073] Where N is the total number of sequences obtained after slicing num sequences, S_t i ={S_t i 1 , S_t i 2 ,…,S_t i win}, i = 1, 2, ..., N, is the i-th sequence among all sequences, where win is the sequence length, that is, the length of the window;

[0074] S32. Calculate the Euclidean distance d between any two sequences in S' ij ,i=1,2,…,N,j=1,2,…,N:

[0075] d ij =||S_t i -S_t j ||2

[0076] Among them, d ij is the Euclidean distance, S_t i is the i-th sequence in the sequence obtained after slicing, with a length of win, S_t j is the jth sequence in the sequence obtained after slicing, with a length of win;

[0077] S33. Calculate the local density of each sequence:

[0078]

[0079]

[0080] Among them, the cutoff distance d c is the cutoff distance, as a hyperparameter d c The value of ρ i ,i=1,2,...,the average value of N is 0.2*N, ranging from 0.1 to 0.8, d ij is the Euclidean distance, ρ iρ is the local density, and χ is the judgment function;

[0081] S34. Calculate the local density, and take the sequence sample with the maximum local density as the density center. The density center sequence is S_t po :

[0082] po = arg max ρ i

[0083] where po is the sequence number of the sequence with the maximum local density, and argmaxρ i is the value of i when ρ i takes the maximum value, and max(ρ i ) is the maximum value of the local density.

[0084] Specifically, the L2 norm is used for distance calculation in S32.

[0085] Specifically, the K-nearest neighbor method is used for clustering in S4. With S_t po as the center, k sequence samples closest to the density center S_t po are extracted from S'. The value range of k is 100 to 5000, and the Euclidean distance is used for distance calculation.

[0086] Furthermore, S5 is specifically as follows:

[0087] The number of neurons in the input layer of the autoencoder network is the length win of the sequences in the training sample set X. The number of neurons in the hidden layer ranges from 10 to 500, and the number of neurons in the bottleneck layer ranges from n = 12 to 256.

[0088] Specifically, the structure of the autoencoder (AEN) is: 1 input layer, 6 hidden layers, and 1 output layer;

[0089] The number of neurons from the input to the output in each layer is 200, 100, 80, 64, 80, 100, 200 respectively. The input layer and the three subsequent connected hidden layers form the encoder, with the number of neurons 200, 100, 80, 64;

[0090] Among them, the bottleneck layer is extracted as a sequence of length 64. The hyperparameters of AEN are obtained through grid search and 10-fold cross-validation. Finally, the activation function of the hidden layer is selected as the rectified linear unit (ReLu), and there is no activation function in the input layer and output layer. The Adam optimizer is selected for backpropagation optimization, the training batch is selected as 256, and the number of training times is 5000.

[0091] Specifically, the input time of GRUNN in S6 is the length of the previous part of the sequence in the low-dimensional sparse representation sequence sample set X', with the interval of 12 to 256, the feature is defined as 1D, the number of neurons in the gated recurrent unit GRU is in the interval of 20 to 100, and the output is connected to a fully connected neural network, where the number of neurons is in the interval of 1 to 20, that is, the length of the latter part of the sequence.

[0092] Specifically, in the structure of the gated recurrent unit GRU, the number of neurons is 30;

[0093] The sequence in the low-dimensional sparse representation sequence sample set X' with a length of 64 is divided into the previous part and the latter part of the sequence, with lengths of 60 and 4 respectively. The gated recurrent unit GRU inputs the previous part of the sequence with a length of 60, and the output is connected to a fully connected neural network. The number of neurons in the fully connected neural network is 4, which is the latter part of the sequence. Through the AE-GRUNN model, in the sequence with a length of 64, the values of the first 60 points are used to predict the values of the last 4 points.

[0094] The hyperparameter selection of the AE-GRUNN model is obtained through grid search and 10-fold cross-validation. Finally, the activation function is selected as the rectified linear unit (ReLu), the fully connected neural network is used as the output without an activation function, the backpropagation optimizer is selected as the Adam optimizer, the training batch is selected as 256, and the number of training times is 5000.

[0095] Further, in S8, the abnormal score scr value is calculated, and the formula is:

[0096]

[0097] where y pre is the GRUNN prediction vector, y true is the true vector of the sample, n_ner is the predicted sequence length, that is, the number of NN neurons; y_pre i , i = 1, 2..., n_er is the value at the i-th time point in the sequence sample predicted by GRUNN, y_true i , i = 1, 2..., n_er is the true value at the i-th time point in this sequence sample, n_er = n_ner, E is the average value calculation function, and T is the matrix transpose symbol.

[0098] Further, the specific steps of the 3-sigma rule in S9 are as follows:

[0099] S91. Calculate the average value and standard deviation of the abnormal score set A;

[0100] S92. Calculate all scr values of the abnormal score set A. When the scr value is greater than the mean plus 3 times the standard deviation or less than the mean minus 3 times the standard deviation, determine the sequence corresponding to the scr value as an abnormal sequence, that is, an abnormal voiceprint signal.

[0101] In a specific embodiment:

[0102] See Figure 2 As shown, the structure of the prediction model is presented. First, a training set is obtained through a clustering algorithm. After that, the AEN model is trained with this training set. After extracting the bottleneck layer, the sequence is divided into a first part and a second part, and the GRUNN model is trained. Finally, a complete AE-GRUNN model that can predict sequences is formed. The sequence to be tested first passes through the encoder part in AEN, and then after splitting the sequence, it passes through the GRUNN model. Calculate the abnormal score for the true value and the predicted value of the second part after splitting. Finally, the 3-sigma rule obtains the abnormal information, so as to achieve the function of unsupervised training of the model and detecting abnormal signals.

[0103] The following table shows the performance indicators of the model detection in the samples to be tested. TP is the true positive. It represents the number of positive samples that the model correctly labels as positive samples. FP is the false positive, which represents the number of negative samples that the model wrongly labels as positive samples. TN is the true negative, which represents the number of negative samples that the model correctly labels as negative samples. FN is the false negative, which represents the number of positive samples that the model wrongly labels as negative samples. The formulas for accuracy acc, precision prec, false positive rate FRR, and false negative rate MRR are as follows:

[0104] acc = (TP + TN) / (TP + FP + TN + FN)

[0105] prec = TP / (TP + FP)

[0106] FRR = FP / (TN + FP)

[0107] MRR = FN / (FN + TP)

[0108] Table 1. Performance indicators

[0109] Index acc prec FRR MRR Percentage (%) 98.2 93％ 1.88％ 1.5％

[0110] It can be seen that the invention has a relatively high accuracy in detecting abnormal signals, reaching 98.2%. At the same time, it has relatively low false positive rate and false negative rate, which are 1.88% and 1.5% respectively, indicating that the model has fewer misjudgments of normal data, can accurately distinguish normal data from abnormal data, and the model can capture most abnormal data and rarely miss real abnormal situations. Therefore, it has a relatively high detection efficiency, and both of them maintain a relatively low level, which can have higher reliability when dealing with abnormal data.

[0111] The various embodiments in this specification are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. For the same or similar parts among the embodiments, reference can be made to each other. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple. For the relevant parts, reference can be made to the description in the method section.

[0112] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present invention. Various modifications to these embodiments will be obvious to those skilled in the art. The general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. An unsupervised detection method for acoustic fingerprint signal faults of a power transformer, characterized in that, It includes the following steps: S1. Use sensors to collect the acoustic fingerprint signals of each power transformer, extract and collect the time-sequence-based acoustic fingerprint signals through a reading circuit connected to the sensors, and form a sample set S; S2. Define a window, and use the window to perform sliding slicing on each sequence sample in the sample set S to obtain the sliced sample set S'; S3. Apply the density peak algorithm to the sliced sample set S' to obtain a sequence as the density center; S4. Use the sequence obtained in S3 as the center point, and use the K-nearest neighbor clustering method to find the K nearest neighbors of the center point. The neighbors and the center point jointly form the training sample set X; S5. Construct an AE-GRUNN model composed of an autoencoder network AEN, a gated recurrent unit GRU, and a fully connected neural network NN. Use the training sample set X as the input and output of the autoencoder network for training to obtain the trained autoencoder. Extract the encoder part and output the low-dimensional sparse representation sequence sample set X', that is, extract the bottleneck layer vector of the autoencoder network; S6. Use the low-dimensional sparse representation sequence sample set X' as the input and target output of the GRUNN model composed of the gated recurrent unit GRU and the fully connected neural network NN, and train the constructed GRUNN model to obtain the trained GRUNN model; S7. Input the low-dimensional sparse representation sequence sample set X' into the GRUNN model and output the prediction sequence; S8. Calculate the anomaly score scr value through the prediction sequence and the target output of the GRUNN model to obtain the anomaly score set A of all samples in the sample to be tested; S9. Use the 3-sigma rule to exclude the anomaly points in the anomaly score set A, extract the new acoustic fingerprint signals to be tested in 2-4 cycles, and repeat S5-S6 to obtain a new anomaly score set; S3 specifically is: S31. Define S’ = {S_win 1 , S_win 2 …S_win num} = {S_t1, S_t2, …, S_t N}, where N is the total number of sequences obtained by slicing num sequences, S_t i = {S_t i 1 , S_t i 2 , …, S_t i win}, i = 1, 2, …, N, is the i-th sequence in all sequences, where win is the sequence length, that is, the length of the window; S32. Calculate the Euclidean distance d between any two sequences in S'. ij , where i = 1, 2, …, N and j = 1, 2, …, N: d ij = ||S_t i - S_t j ||² where d ij is the Euclidean distance, and S_t i is the i-th sequence in the sequence obtained after slicing, with a length of win, and S_t j is the j-th sequence in the sequence obtained after slicing, with a length of win; S33. Calculate the local density of each sequence: Among them, the truncation distance d c is the truncation distance, which is the hyperparameter d c takes the value when the average value of ρi is 0.2*N, where i = 1, 2,..., N, and the range is 0.1 to 0.8, and d ij is the Euclidean distance, and ρ i is the local density, and χ is the judgment function; S34. Statistically calculate the local density, and take the sequence sample with the maximum local density as the density center. The density center sequence is S_t po : po=argmaxρi where po is the sequence number of the sequence with the largest local density, argmaxρ i is the value of i when ρ i takes the maximum value, and max(ρ i ) is the maximum value of the local density.

2. The unsupervised detection method for power transformer voiceprint signal fault according to claim 1 is characterized in that: S1 specifically is: Sample set S = {s1, s2,.., s num}, where num is the number of samples, and its value range is 200 to 3000. s i = (s i 1 , s i 2 ,.., s i L ), i = 1, 2,..., num, is the time series of the i-th sample, L is the length of the sequence, and its value range is 50 to 500.

3. The unsupervised detection method for power transformer voiceprint signal fault according to claim 1 is characterized in that: S2 specifically is: S21. Define the length range of the window as win = 30 - 300; S22. Set the starting position of the sliding window as the starting position of the sequence; S23. Take the first sample time sequence of the sample set S, start from the starting position of the sequence with the sliding window size as the unit, gradually move the sliding window, and slice out each sliding window; S24. Repeat S23, take the 2nd to num-th sample time series, and the finally sliced samples form the training sample set S’ = {S_win 1 , S_win 2 …S_win num}.

4. A method for unsupervised detection of power transformer acoustic signal faults according to claim 1, characterized in that, S5 specifically is: The number of neurons in the input layer of the autoencoder network is the length win of the sequence in the training sample set X, the number of neurons in the hidden layer ranges from 10 to 500, and the number of neurons in the bottleneck layer ranges from n = 12 to 256.

5. The unsupervised detection method for power transformer voiceprint signal fault according to claim 1 is characterized in that: In S8, calculate the anomaly score scr value, and the formula is: Among them, y pre is the GRUNN prediction vector, and y true is the true vector of the sample, and n_ner is the predicted sequence length, that is, the number of NN neurons; y_pre i is the value at the i-th time point in the sequence sample predicted by GRUNN, where i = 1, 2..., n_er; y_true i is the true value at the i-th time point in this sequence sample, where i = 1, 2..., n_er; E is the average value calculation function, and T is the matrix transpose symbol.

6. The unsupervised detection method for power transformer voiceprint signal fault according to claim 1, characterized in that: The specific steps of the 3-sigma rule in S9 are: S91. Calculate the average value and standard deviation of the anomaly score set A; S92. Calculate all scr values of the anomaly score set A. When the scr value is greater than the average value plus 3 times the standard deviation or less than the average value minus 3 times the standard deviation, determine that the sequence corresponding to the scr value is an abnormal sequence, that is, an abnormal acoustic fingerprint signal.

Images