CN117195077A - Unsupervised detection method for fault of voiceprint signal of power transformer - Google Patents

Abstract

The invention discloses an unsupervised detection method for voiceprint signal faults of a power transformer, which is applied to the technical field of fault diagnosis of the power transformer. Comprising the following steps: the method comprises the steps of obtaining voiceprint signal data of each power transformer by using a sensor, forming a sample set, finding a density center of the sample set by using a density peak value algorithm, clustering by using a K nearest neighbor clustering algorithm to obtain a training sample set, putting the training sample set into an automatic encoder network to extract a low-dimensional sparse representation sequence sample set, preprocessing, taking the former part of a sequence as input of a gate control circulating unit GRU, and taking the latter part of the sequence as output for training; the voiceprint signal to be detected is sequentially output through the prediction sequence and the GRUNN target to obtain an abnormal score set, and the abnormal score set is processed by using a 3-sigma rule, so that faults are detected. The invention effectively improves the accuracy of fault detection and obtains lower false alarm rate and false missing rate.

Description

Unsupervised detection method for fault of voiceprint signal of power transformer

Technical Field

The invention relates to the technical field of power transformer fault detection, in particular to an unsupervised detection method for power transformer voiceprint signal faults.

Background

Along with the fact that a power transformer is one of important equipment in a power system, the power transformer is widely applied to power transmission and distribution systems and plays roles in voltage conversion, voltage regulation, power line isolation and power equipment protection. However, there is a risk of faults in the power transformer due to long-term operation and various external factors, and the faults may cause the equipment to fail to work or even cause accidents, which seriously affect the stable operation of the power system.

At present, the traditional power transformer fault detection method mainly depends on manual inspection and regular instrument detection, and the method needs to consume a great deal of manpower and material resources, can not monitor the state of equipment in real time, and can not perform real fault early warning. In addition, due to the complex electromagnetic field, thermal field and other environments in the power transformer, fault signals are often covered in background noise, so that fault detection becomes more difficult. The voiceprint signal is one of important indexes of fault detection of the power transformer, and whether the equipment is abnormal or not can be judged by analyzing acoustic characteristics. The voiceprint signal is a sound signal generated when the power transformer is operated, and sound of a specific frequency and amplitude is generated due to vibration of equipment, internal arc, and the like caused by a fault. Therefore, the fault type and the fault position can be accurately identified by analyzing the voiceprint signal of the transformer, and fault diagnosis and state monitoring of equipment are realized.

Research on power transformer voiceprint signal fault detection is mainly focused on two directions: feature extraction and fault identification. Feature extraction refers to extracting parameters or features that can represent sound features by performing mathematical processing and analysis on a voiceprint signal. Common features include frequency domain features (e.g., energy, magnitude spectrum, etc.), time domain features (e.g., duration, zero crossing rate, etc.), wavelet transform features, and the like. The fault identification is to match the extracted characteristics with a fault mode by using methods such as machine learning, deep learning and the like, and identify abnormal signals, so that the automatic detection of faults is realized. 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.

The prior main flow technology of fault diagnosis of the transformer voiceprint signal has a method based on machine learning and deep learning, a clustering-based K nearest neighbor algorithm (K-means), a Principal Component Analysis (PCA) algorithm, a Random Forest (RF), a Support Vector Machine (SVM) and other machine learning methods are applied to fault diagnosis, but the methods often depend on the quality of a feature extraction effect, before the methods are used for classified diagnosis, feature extraction is required, a method for manually extracting features such as peak-to-peak value, variance and frequency spectrum cannot be well applied to all scenes, a manual extraction mode is complex and complicated, some scholars propose a Fourier transform or wavelet transform mode to extract the features, however, the methods need to manually set parameters such as different applicable wavelet basis functions for different scenes, and the generalization capability is weak. When noise exists in the signal, the machine learning method cannot learn the data distribution characteristics well, and the nonlinear fitting effect on the noisy signal is poor, so that the effect is poor in test concentration. Deep learning is increasingly applied to solving the problem of complex nonlinear fitting, and for transformer fault classification, a neural network method is proposed, learning is performed on an existing data set, unknown samples are detected, fault signals are classified, but in actual situations, voiceprint signals cannot be acquired in advance, so that many neural network methods for supervised learning based on priori data do not have generalization. The long-short-term memory neural network (LSTM) is used as a variant of the cyclic neural network, so that the problem that long sequences are dependent can be solved, and meanwhile, abnormal long sequences in a transformer can be detected, but due to the complexity of the structure and the characteristic of circulation in training, the hardware is difficult, and the problem that the application is difficult can be solved better by adopting the improved LSTM network to predict the sequences.

Therefore, an unsupervised detection method for voiceprint signal faults of a power transformer is provided to solve the problems existing in the prior art, which are needed to be solved by those skilled in the art.

Disclosure of Invention

In view of the above, the invention provides an unsupervised detection method for voiceprint signal faults of a power transformer, which is used for solving the technical problems existing in the prior art.

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

an unsupervised detection method for voiceprint signal faults of a power transformer comprises the following steps:

s1, collecting voiceprint signals of each power transformer by using a sensor, extracting and collecting the voiceprint signals based on time sequence by a reading circuit connected with the sensor, and forming a sample set S;

s2, defining a window, and performing sliding slicing on each sequence sample in the sample set S by using the window to obtain a sliced sample set S';

s3, adopting a density peak algorithm to the sliced sample set S' to obtain a sequence serving as a density center;

s4, using the sequence obtained in the S3 as a central point, using a K-neighbor-based clustering method to find K nearest neighbors of the central point, and jointly forming a training sample set X by the neighbors and the central point;

s5, constructing an AE-GRUNN model consisting of an automatic encoder network AEN, a gate control circulation unit GRU and a fully connected neural network NN, taking a training sample set X as input and output of the automatic encoder network to train, obtaining a trained automatic encoder, extracting an encoder part, outputting a low-dimensional sparse representation sequence sample set X', namely extracting a bottleneck layer vector of the automatic encoder network;

s6, taking a low-dimensional sparse representation sequence sample set X' as input and target output of a GRUNN model formed by a gating circulation unit GRU and a fully connected neural network NN, and training the constructed GRUNN model to obtain a trained GRUNN model;

s7, inputting a low-dimensional sparse representation sequence sample set X' into a GRUNN model and outputting a predicted sequence;

s8, calculating an anomaly score scr value through a prediction sequence and target output of a GRUNN model to obtain an anomaly score set A of all samples in the samples to be detected;

s9, using a 3-sigma rule to eliminate abnormal points in the abnormal score set A, extracting new voiceprint signals to be detected in 2-4 periods, and repeating S5-S6 to obtain a new abnormal score set.

Optionally, S1 is specifically:

sample set s= { S ₁ ，s ₂ ，..，s _num Numerical value of 200-3000 s, wherein num is the number of samples _i ＝(s _i ¹ ，s _i ² ，..，s _i ^L ) I=1, 2, …, num, which is the time series of the i-th sample, L is the length of the sequence, and the value range is 50 to 500.

Optionally, S2 is specifically:

s21, defining the length range of a window to be win=30-300;

s22, setting the initial position of the sliding window as the initial position of the sequence;

s23, taking a 1 st sample time sequence of the sample set S, taking the size of a sliding window as a unit, starting from the initial position of the sequence, gradually moving the sliding windows, and cutting out each sliding window;

s24, repeating S23, taking the time series of samples from the 2 nd to num, and finally slicing the samples to form a training sample set S' = { S_win ¹ ，S_win ² …S_win ^num }。

Optionally, S3 is specifically:

s31, defining S' = { s_win ¹ ，S_win ² …S_win ^num }＝{S_t ₁ ，S_t ₂ ，…，S_t _N 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 of all sequences, where win is the sequence length, i.e. the length of the window;

s32, calculating the Euclidean distance d between any two sequences in the S _ij ，i＝1，2，…，N，j＝1，2，…，N：

d _ij ＝||S_t _i -S_t _j || ₂

Wherein d _ij Is Euclidean distance, S_t _i The length of the ith sequence in the sequences obtained after slicing is win, S_t _j The length of the j sequence in the sequence obtained after slicing is win;

s33, calculating the local density of each sequence:

wherein the cutoff distance d _c For the cut-off distance as the super-parameter d _c The value of (1) is taken as ρ _i I=1, 2..when the average value of N is 0.2 x NThe value is in the range of 0.1 to 0.8, d _ij Is Euclidean distance ρ _i Is local density, χ is a judgment function;

s34, counting local density, taking a sequence sample with the maximum local density as a density center, wherein the density center sequence is S_t _po ：

po＝arg maxρ _i

Where po is the sequence number of the sequence with the greatest local density, argmax ρ _i Is when ρ _i When the maximum value is taken, i takes on a value, max (ρ _i ) Is the local density maximum.

Optionally, S5 is specifically:

the number of neurons of the input layer of the automatic encoder network is the length win of the sequence in the training sample set X, the number of neurons of the hidden layer ranges from 10 to 500, and the number of neurons of the bottleneck layer ranges from n=12 to 256.

Optionally, in S8, calculating an anomaly score scr value, where the formula is:

wherein y is _pre For GRUNN predictive vector, y _true N_ner is the predicted sequence length, i.e., NN neuron number, for the true vector of samples; y_pre _i I=1, 2 …, n_er is the i-th time point value, y_true, in the grenn predicted sequence sample _i I=1, 2 …, n_er is the true value of the i-th time point in the sequence sample, n_er=n_ner, E is the average calculation function, and T is the matrix transpose symbol.

Optionally, the 3-sigma rule in S9 specifically includes the steps of:

s91, calculating the average value and standard deviation of the anomaly score set A;

s92, calculating all scr values of the anomaly score set A, and judging that a sequence corresponding to the scr value is an anomaly sequence, namely an anomaly voiceprint signal, when the scr value is larger than the average value plus 3 times of standard deviation or smaller than the average value minus 3 times of standard deviation.

Compared with the prior art, the invention discloses an unsupervised detection method for the voiceprint signal fault of the power transformer, which has the beneficial effects that: by using the thought of a clustering algorithm, the model automatically searches for the training set without manual definition, so that unsupervised training is realized, and the data set obtained by density clustering is used as the training set to better reveal the characteristics of the sequence, so that the method is more suitable for training of a later network; the method has the advantages that the model of the automatic encoder network AEN combined with the gate control circulating unit GRU and the fully connected neural network is designed, the characteristics extracted by the automatic encoder network AEN can be used for revealing the change characteristics of the sequence, and the method avoids the defect that extraction functions are respectively defined for data with different characteristics; finally, an AE-GRUNN model with a simple structure is designed, the prediction effect is better, the report missing rate and the false report rate are both lower, an unsupervised model learning mode and a simple prediction network model structure are realized in the whole process, and the method can be applied to follow-up better hardware and actual scenes.

Drawings

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings that are required to be used in the embodiments or the description of the prior art will be briefly described below, and it is obvious that the drawings in the following description are only embodiments of the present invention, and that other drawings can be obtained according to the provided drawings without inventive effort for a person skilled in the art.

FIG. 1 is a flow chart of an unsupervised detection method for voiceprint signal faults of a power transformer;

fig. 2 is a diagram of an AE-GRUNN detection model according to the present invention.

Detailed Description

The following description of the embodiments of the present invention will be made clearly and completely with reference to the accompanying drawings, in which it is apparent that the embodiments described are only some embodiments of the present invention, but not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.

Referring to fig. 1, the invention discloses an unsupervised detection method for voiceprint signal faults of a power transformer, which comprises the following steps:

s1, collecting voiceprint signals of each power transformer by using a sensor, extracting and collecting the voiceprint signals based on time sequence by a reading circuit connected with the sensor, and forming a sample set S;

s2, defining a window, and performing sliding slicing on each sequence sample in the sample set S by using the window to obtain a sliced sample set S';

s3, adopting a density peak algorithm to the sliced sample set S' to obtain a sequence serving as a density center;

s4, using the sequence obtained in the S3 as a central point, using a K-neighbor-based clustering method to find K nearest neighbors of the central point, and jointly forming a training sample set X by the neighbors and the central point;

s5, constructing an AE-GRUNN model consisting of an automatic encoder network AEN, a gate control circulation unit GRU and a fully connected neural network NN, taking a training sample set X as input and output of the automatic encoder network to train, obtaining a trained automatic encoder, extracting an encoder part, outputting a low-dimensional sparse representation sequence sample set X', namely extracting a bottleneck layer vector of the automatic encoder network;

s6, taking a low-dimensional sparse representation sequence sample set X' as input and target output of a GRUNN model formed by a gating circulation unit GRU and a fully connected neural network NN, and training the constructed GRUNN model to obtain a trained GRUNN model;

s7, inputting a low-dimensional sparse representation sequence sample set X' into a GRUNN model and outputting a predicted sequence;

s8, calculating an anomaly score scr value through a prediction sequence and target output of a GRUNN model to obtain an anomaly score set A of all samples in the samples to be detected;

s9, using a 3-sigma rule to eliminate abnormal points in the abnormal score set A, extracting new voiceprint signals to be detected in 2-4 periods, and repeating S5-S6 to obtain a new abnormal score set.

Specifically, the automatic encoder network in S5 is composed of an encoder and a decoder.

Specifically, abnormal points are removed by using a 3-sigma rule, so that fault voiceprint signals are detected as completely as possible, and new faults can be timely detected when the transformer runs, so that lower false alarm rate and lower false alarm rate are obtained.

Further, S1 is specifically:

sample set s= { S ₁ ，s ₂ ，..，s _num Numerical value of 200-3000 s, wherein num is the number of samples _i ＝(s _i ¹ ，s _i ² ，..，s _i ^L ) I=1, 2, …, num, which is the time series of the i-th sample, L is the length of the sequence, and the value range is 50 to 500.

Further, S2 is specifically:

s21, defining the length range of a window to be win=30-300;

s22, setting the initial position of the sliding window as the initial position of the sequence;

s23, taking a 1 st sample time sequence of the sample set S, taking the size of a sliding window as a unit, starting from the initial position of the sequence, gradually moving the sliding windows, and cutting out each sliding window;

s24, repeating S23, taking the time series of samples from the 2 nd to num, and finally slicing the samples to form a training sample set S' = { S_win ¹ ，S_win ² …S_win ^num }。

Specifically, the moving step length of the sliding window in S23 is 1, which means that 1 data point is slid every time, and the 1 st sample time series of sliced samples s_win is obtained ¹ ＝{S_win ₁ ¹ ，S_win ₂ ¹ ，…S_win _nw ¹ N, wherein nw is the number of sequences obtained after sliding window slicing, S_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 from the 1 st sample slice, wherein the length of the sequence is win, i.e. the size of the window.

Further, S3 is specifically:

s31, defining S' = { s_win ¹ ，S_win ² …S_win ^num }＝{S_t ₁ ，S_t ₂ ，…，S_t _N }, it

Wherein 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 of all sequences, where win is the sequence length, i.e. the length of the window;

s32, calculating the Euclidean distance d between any two sequences in the S _ij ，i＝1，2，…，N，j＝1，2，…，N：

d _ij ＝||S_t _i -S_t _j || ₂

Wherein d _ij Is Euclidean distance, S_t _i The length of the ith sequence in the sequences obtained after slicing is win, S_t _j The length of the j sequence in the sequence obtained after slicing is win;

s33, calculating the local density of each sequence:

wherein the cutoff distance d _c For the cut-off distance as the super-parameter d _c The value of (1) is taken as ρ _i I=1, 2..the average value of N is 0.2 x N, ranging from 0.1 to 0.8, d _ij Is Euclidean distance ρ _i Is local density, χ is a judgment function;

s34, counting local density, taking a sequence sample with the maximum local density as a density center, wherein the density center sequence is S_t _po ：

po＝arg max ρ _i

Where po is the sequence number of the sequence with the greatest local density, argmax ρ _i Is when ρ _i When the maximum value is taken, i takes on a value, max (ρ _i ) Is the local density maximum.

Specifically, the distance in S32 is calculated using the L2 norm.

Specifically, in S4, clustering is performed by using a K-nearest neighbor method to obtain s_t _po As the center, k distance density centers S_t are extracted from S _po The nearest sequence sample, wherein k has a value range of 100-5000, and the Euclidean distance is used for distance calculation.

Further, S5 is specifically:

the number of neurons of the input layer of the automatic encoder network is the length win of the sequence in the training sample set X, the number of neurons of the hidden layer ranges from 10 to 500, and the number of neurons of the bottleneck layer ranges from n=12 to 256.

Specifically, the structure of the Automatic Encoder (AEN) is: 1 layer of input layer, 6 layers of hidden layer and 1 layer of output layer;

the number of neurons from input to output is 200, 100, 80, 64, 80, 100, 200 for each layer, wherein the input layer and the three hidden layer constituent encoders connected later, the number of neurons is 200, 100, 80, 64;

the extraction bottleneck layer is a 64-length sequence, the hyper-parameter selection of AEN is obtained through grid search and 10-fold cross validation, the activation functions of the hidden layer are all selected to be linear rectification functions (ReLu), the input layer and the output layer have no activation functions, the back propagation optimizer is selected to be Adam optimizer, the training batch is selected to be 256, and the training times are 5000.

Specifically, in S6, the input time of the GRUNN is the length of the front part of the sequence in the low-dimensional sparse representation sequence sample set X', the interval is 12-256, the characteristic is defined as 1 dimension, the number interval of the neurons in the gate control circulation unit GRU is 20-100, and the output is connected with the fully connected neural network, wherein the number interval of the neurons is 1-20, namely the length of the rear part of the sequence.

Specifically, in the structure of the gate control circulation unit GRU, the number of neurons is 30;

the sequences in the low-dimensional sparse representation sequence sample set X' with the length of 64 are divided into a front part and a rear part, and the lengths are respectively 60 and 4. The gate control circulation unit GRU inputs the front part of the sequence with the length of 60, and outputs the front part of the sequence to be connected with the fully-connected neural network, and the number of neurons of the fully-connected neural network is 4, namely the rear part of the sequence. The values of the last 4 points are predicted by the AE-GRUNN model from the first 60 points in a sequence of length 64.

The super-parameter selection of the AE-GRUNN model is obtained through grid search and 10-fold cross validation, the finally obtained activation functions are all linear rectification functions (ReLu), the fully-connected neural network is used as output, no activation function exists, the back propagation optimizer is used for selecting an Adam optimizer, the training batch is used for selecting 256, and the training times are 5000.

Further, in S8, the anomaly score scr value is calculated according to the formula:

wherein y is _pre For GRUNN predictive vector, y _true N_ner is the predicted sequence length, i.e., NN neuron number, for the true vector of samples; y_pre _i I=1, 2 …, n_er is the i-th time point value, y_true, in the grenn predicted sequence sample _i I=1, 2 …, n_er is the true value of the i-th time point in the sequence sample, n_er=n_ner, E is the average calculation function, and T is the matrix transpose symbol.

Further, the 3-sigma rule in S9 specifically includes the steps of:

s91, calculating the average value and standard deviation of the anomaly score set A;

s92, calculating all scr values of the anomaly score set A, and judging that a sequence corresponding to the scr value is an anomaly sequence, namely an anomaly voiceprint signal, when the scr value is larger than the average value plus 3 times of standard deviation or smaller than the average value minus 3 times of standard deviation.

In one particular embodiment:

referring to fig. 2, the structure of the prediction model is shown, a training set is obtained through a clustering algorithm, the AEN model is trained by the training set, after a bottleneck layer is extracted, a sequence is divided into a front part and a rear part, the grenn model is trained, finally a complete AE-grenn model capable of predicting the sequence is formed, the sequence to be tested firstly passes through an encoder part in the AEN, then the grenn model is used for dividing the sequence, the anomaly score is calculated by the real value and the predicted value of the divided rear part, and finally anomaly information is obtained by a 3-sigma rule, so that the functions of an unsupervised training model and anomaly signal detection are achieved.

The following table shows performance indexes of model detection in a sample to be detected, and TP is a real example. The representation model correctly marks positive samples as the number of positive samples. FP is a false positive example, indicating the number of negative samples that the model wrongly marked as positive samples. TN is a true negative example, indicating that the model correctly marks the negative as the number of negative samples. FN is a false negative example, indicating the number of positive samples that the model wrongly marks as negative samples. The formulas of the accuracy acc, the accuracy pre, the false alarm rate FRR and the missing alarm rate MRR are as follows:

acc＝(TP+TN)/(TP+FP+TN+FN)

prec＝TP/(TP+FP)

FRR＝FP/TN+FP

MRR＝FN/FN+TP

TABLE 1 Performance index

Index (I)	acc	prec	FRR	MRR
					Percentage (%)	98.2	93％	1.88％	1.5％

The invention has the advantages that the accuracy rate of detecting abnormal signals is higher and reaches 98.2%, meanwhile, the false alarm rate and the false alarm rate are lower and are respectively 1.88% and 1.5%, which shows that the false judgment of the model on normal data is less, the normal data can be accurately distinguished from the abnormal data, the model can capture most of the abnormal data, and the real abnormal situation is rarely missed, so that the invention has higher detection efficiency, and the false alarm rate are kept at a lower level, so that the invention has higher reliability when the abnormal data is processed.

In the present specification, each embodiment is described in a progressive manner, and each embodiment is mainly described in a different point from other embodiments, and identical and similar parts between the embodiments are all enough to refer to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant points refer to the description of the method section.

The previous description of the disclosed embodiments is provided to enable any person skilled in the art to make or use the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the generic principles defined herein may be applied to other embodiments without departing from the spirit or scope of the invention. Thus, the present invention is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims (7)

1. An unsupervised detection method for voiceprint signal faults of a power transformer is characterized by comprising the following steps:

s1, collecting voiceprint signals of each power transformer by using a sensor, extracting and collecting the voiceprint signals based on time sequence by a reading circuit connected with the sensor, and forming a sample set S;

s2, defining a window, and performing sliding slicing on each sequence sample in the sample set S by using the window to obtain a sliced sample set S';

s3, adopting a density peak algorithm to the sliced sample set S' to obtain a sequence serving as a density center;

s4, using the sequence obtained in the S3 as a central point, using a K-neighbor-based clustering method to find K nearest neighbors of the central point, and jointly forming a training sample set X by the neighbors and the central point;

s5, constructing an AE-GRUNN model consisting of an automatic encoder network AEN, a gate control circulation unit GRU and a fully connected neural network NN, taking a training sample set X as input and output of the automatic encoder network to train, obtaining a trained automatic encoder, extracting an encoder part, outputting a low-dimensional sparse representation sequence sample set X', namely extracting a bottleneck layer vector of the automatic encoder network;

s6, taking a low-dimensional sparse representation sequence sample set X' as input and target output of a GRUNN model formed by a gating circulation unit GRU and a fully connected neural network NN, and training the constructed GRUNN model to obtain a trained GRUNN model;

s7, inputting a low-dimensional sparse representation sequence sample set X' into a GRUNN model and outputting a predicted sequence;

s8, calculating an anomaly score scr value through a prediction sequence and target output of a GRUNN model to obtain an anomaly score set A of all samples in the samples to be detected;

s9, using a 3-sigma rule to eliminate abnormal points in the abnormal score set A, extracting new voiceprint signals to be detected in 2-4 periods, and repeating S5-S6 to obtain a new abnormal score set.

2. The unsupervised detection method for voiceprint signal faults of power transformer according to claim 1, wherein S1 specifically comprises:

sample set s= { S ₁ ，s ₂ ，..，s _num Where num is the number of samples,the value range is 200-3000 s _i ＝(s _i ¹ ，s _i ² ，..，s _i ^L ) I=1, 2, …, num, which is the time series of the i-th sample, L is the length of the sequence, and the value range is 50 to 500.

3. The unsupervised detection method for voiceprint signal faults of power transformer according to claim 1, wherein S2 specifically comprises:

s21, defining the length range of a window to be win=30-300;

s22, setting the initial position of the sliding window as the initial position of the sequence;

s23, taking a 1 st sample time sequence of the sample set S, taking the size of a sliding window as a unit, starting from the initial position of the sequence, gradually moving the sliding windows, and cutting out each sliding window;

s24, repeating S23, taking the time series of samples from the 2 nd to num, and finally slicing the samples to form a training sample set S' = { S_win ¹ ，S_win ² …S_win ^num }。

4. The unsupervised detection method for voiceprint signal faults of power transformer according to claim 1, wherein S3 specifically comprises:

s31, defining S' = { s_win ¹ ，S_win ² …S_win ^num }＝{S_t ₁ ，S_t ₂ ，…，S_t _N 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 of all sequences, where win is the sequence length, i.e. the length of the window;

s32, calculating the Euclidean distance d between any two sequences in the S _ij ，i＝1，2，…，N，j＝1，2，…，N：

d _ij ＝||S_t _i -S_t _j || ₂

Wherein d _ij Is Euclidean distance, S_t _i The length of the ith sequence in the sequences obtained after slicing is win, S_t _j The length of the j sequence in the sequence obtained after slicing is win;

s33, calculating the local density of each sequence:

wherein the cutoff distance d _c For the cut-off distance as the super-parameter d _c The value of (1) is taken as ρ _i I=1, 2..the average value of N is 0.2 x N, ranging from 0.1 to 0.8, d _ij Is Euclidean distance ρ _i Is local density, χ is a judgment function;

s34, counting local density, taking a sequence sample with the maximum local density as a density center, wherein the density center sequence is S_t _po ：

po＝arg max ρ _i

Wherein po is the sequence number of the sequence with the greatest local density, arg max ρ _i Is when ρ _i When the maximum value is taken, i takes on a value, max (ρ _i ) Is the local density maximum.

5. The unsupervised detection method for voiceprint signal faults of power transformer according to claim 1, wherein S5 specifically comprises:

the number of neurons of the input layer of the automatic encoder network is the length win of the sequence in the training sample set X, the number of neurons of the hidden layer ranges from 10 to 500, and the number of neurons of the bottleneck layer ranges from n=12 to 256.

6. The unsupervised detection method for voiceprint signal faults of a power transformer according to claim 1, wherein the anomaly score scr value is calculated in S8 by the formula:

wherein y is _pre For GRUNN predictive vector, y _true N_ner is the predicted sequence length, i.e., NN neuron number, for the true vector of samples; y_pre _i I=1, 2 …, n_er is the i-th time point value, y_true, in the grenn predicted sequence sample _i I=1, 2 …, n_er is the true value of the i-th time point in the sequence sample, n_er=n_ner, E is the average calculation function, and T is the matrix transpose symbol.

7. The unsupervised detection method for voiceprint signal faults of a power transformer according to claim 1, wherein the 3-sigma rule in S9 specifically comprises the following steps:

s91, calculating the average value and standard deviation of the anomaly score set A;

s92, calculating all scr values of the anomaly score set A, and judging that a sequence corresponding to the scr value is an anomaly sequence, namely an anomaly voiceprint signal, when the scr value is larger than the average value plus 3 times of standard deviation or smaller than the average value minus 3 times of standard deviation.