CN113707176B - A Transformer Fault Detection Method Based on Acoustic Signal and Deep Learning Technology - Google Patents

Abstract

The invention relates to a transformer fault detection method based on acoustic signals and a deep learning technology, and compared with the prior art, the transformer fault detection method overcomes the defect that accurate detection of transformer faults is difficult to perform by using voiceprint signals. The invention comprises the following steps: acquiring sound data of the power transformer; preprocessing acoustic signals in a training sample set; extracting sound characteristics of the sound signal data; constructing a transformer fault detection model; training a transformer fault detection model; acquiring and preprocessing acoustic signal data of a transformer to be detected; and obtaining a fault detection result of the transformer to be detected. The invention can detect the fault of the transformer based on the voiceprint signal.

Description

A Transformer Fault Detection Method Based on Acoustic Signal and Deep Learning Technology

technical field

The invention relates to the technical field of transformer fault detection, in particular to a transformer fault detection method based on acoustic signals and deep learning technology.

Background technique

The transformers in the power grid have the characteristics of a large amount of use, a variety of types and specifications, and a long use time, which leads to an excessively high frequency of transformer failures in the power grid system. According to statistics, for every 200 transformers that have been in operation for more than 4 years, about 5 transformers fail. Therefore, the troubleshooting and maintenance of the transformer has become an important procedure for the stability of the power grid.

For transformer troubleshooting, the traditional processing method is to manually go to the location to be eliminated, and rely on manual experience to diagnose whether the transformer is faulty based on the sound emitted by the transformer. This method not only consumes a lot of time and energy, but also suffers from the interference of human factors, which may cause fault diagnosis errors.

In recent years, deep learning has developed rapidly, and the method of processing signals has the characteristics of high efficiency. If the deep learning method can be used to process the sound of the transformer, it will automatically analyze the characteristics of the sound emitted by the transformer, and classify the characteristics, so that the faulty transformer can be quickly diagnosed, which can greatly reduce the labor cost and improve the performance of the transformer. Quickly maintain the stability of the power grid.

In the prior art, although there are some voiceprint detection methods, most of them use mixed Gaussian models and hidden Markov models as effective acoustic models for voiceprint detection, and after the rise of deep learning, including convolutional neural networks, Autoencoder, Recurrent Neural Network (RNN), etc., more and more deep learning networks are applied in the field of sound detection. From the perspective of the accuracy and precision of voice detection, the voiceprint detection method based on deep learning is much better than the voiceprint detection method of traditional acoustic models, but there are many voiceprint detection methods based on deep learning in the existing technology. Using vibration data, no audio data has been used for fault detection.

Therefore, how to better apply deep learning technology to transformer voiceprint fault detection has become an urgent technical problem.

SUMMARY OF THE INVENTION

The purpose of the present invention is to solve the defect in the prior art that it is difficult to use voiceprint signals to accurately detect transformer faults, and to provide a transformer fault detection method based on acoustic signals and deep learning technology to solve the above problems.

In order to achieve the above object, technical scheme of the present invention is as follows:

A transformer fault detection method based on acoustic signal and deep learning technology, comprising the following steps:

11) Acquisition and acquisition of sound data of power transformers: The sound data of transformers are collected and obtained on the spot through the voiceprint acquisition sensor, and are marked into two categories: "normal" and "fault", and are defined as training sample sets;

12) Preprocessing of the acoustic signals in the training sample set: Denoising preprocessing is performed on the collected sound data of power transformers by using the preprocessing methods of segmentation, framing, sound windowing and adaptive filtering; Noise and tuning processing to enhance the sound signal;

13) Sound feature extraction of sound signal data: Extract the sound feature of the preprocessed power transformer sound data by using the Meral cepstral coefficient, and extract the MFCC coefficient;

14) Build a transformer fault detection model: build a transformer fault detection model based on the double-gated convolutional network model and the characteristics of the transformer acoustic signal;

15) Training of the transformer fault detection model: input the extracted MFCC coefficients into the transformer fault detection model for training;

16) Acquisition and preprocessing of the acoustic signal data of the transformer to be detected: Acquire the acoustic signal data of the transformer to be detected, perform denoising preprocessing, and extract the MFCC coefficients from the denoised preprocessed acoustic signal data of the transformer to be detected;

17) Obtaining the fault detection result of the transformer to be detected: Input the MFCC coefficients into the trained transformer fault detection model to obtain the transformer fault detection result.

The preprocessing of the acoustic signal in the training sample set includes the following steps:

21) Perform a segmented operation on the collected power transformer sound data s(t):

segment the obtained power transformer sound data,

Divide into s(t)={s ₁ (t), s ₂ (t),...,s _q (t),...s _r (t)}, and calculate the total length L of the voiceprint data, which is Calculated as follows:

L=time×fS _Sample =r×r _L ,

Among them, fs _sample is the sampling frequency of the sound, time is the sampling time, r is the number of segments, and r _L is the length of the segment;

22) Framing processing of the segmented transformer sound data s _q (t):

Set the transformer voiceprint frame length to 500ms, and perform framing processing as

s _q (t)={s _q1 (t),s _q2 (t),...,s _qp (t),...s _qLength (t)},

Among them, the length of each frame is set to 500ms, and each segment is divided into Length frames;

23) Windowing the transformer sound after framing:

The end-point smoothing window processing is performed on the framed data, and the frame is windowed by the Hamming window. The function w(t) of the Hamming window is expressed as follows:

where M is the frame length and t is the time;

The time domain signal f _qp (t) is obtained for each frame, and its expression is as follows:

f _qp (t)=s _qp (t)*w(n),

Wherein, _fqp (t) is the time domain signal of the pth frame of the qth segment signal, w(n) is the windowing function, and _sqp (t) is the signal value of the pth frame of the qth segment signal;

24) Use an adaptive filter to denoise the windowed sound:

Sampling f _qp (t) to obtain the X _i (n) digital signal sequence, and set the initial value of the filter as follows:

Among them, W(k) is the optimal weight coefficient, μ is the convergence factor, and λ is the maximum eigenvalue of the correlation matrix;

Compute the actual output estimate of the filter output:

y(k)=W ^T (k)Xi(n),

Among them, y(k) is the actual output estimated value, W ^T (k) is the transpose of the optimal weight coefficient, and X _i (n) is the input signal sequence;

e(k) is the error signal, which is the calculation error:

e(k)=d(k)-y(k),

Among them, d(k) is the expected output value, and the filter coefficients at time k+1 are updated:

W(k+1)=W(k)+μe(k)X(k),

Among them, W(k+1) is the optimal weight coefficient at time k+1, W(k) is the optimal weight coefficient at time k, e(k) is the error at time k, and X(k) is the error at time k input sequence;

Use the steepest descent method to iteratively solve the problem to minimize the error signal, and obtain the output y(k) after adaptive filtering and denoising;

25) Perform data enhancement on the acoustic signal by cutting, adding noise, and tuning the output y(k) of the adaptive filtering and denoising.

The sound feature extraction of the sound signal data includes the following steps:

31) Perform inverse transformation on y(k) after data enhancement to generate s(t), and perform pre-emphasis operation on s(t), and the calculation formula is as follows:

y(z)=s(z)H(z),

where y(z) is the output of the pre-emphasis, s(z) is the Z-domain representation of the signal s(t), and H(z) is the transfer function of the high-pass filter, which is:

This high-pass filter is in the form of the z-domain, where the value of μ is between 0.9-1.0;

32) Perform frame division processing on the pre-emphasized transformer sound data s _q (t):

Set the frame length of the transformer voiceprint to 30ms, and perform frame-by-frame processing of the transformer voice data. The processed structure is 30ms per frame;

33) Windowing the sound data after framing: use the hamming window to window the frame. Assuming that the output obtained after the first two steps of preprocessing is S(n) and the window function is W(n), the calculation formula is :

S'(n)=S(n)×W(n);

where the window function W(n) is:

In this formula, n represents time n, and its value range is 0≤n≤N-1, N is the number of sampling points, and different Hamming windows are obtained according to different a values, n represents the sequence independent variable, and a is the set constant ;

34) Perform fast Fourier transform on the windowed data:

Perform FFT on each frame signal to get the frequency domain signal X(k):

where X(k) represents the frequency domain output, x(n) represents the time domain input, and N is the number of sampling points;

35) Perform mel filtering on the data after the fast Fourier transform, and the conversion formula of mel filtering is:

Among them, f represents the physical frequency, and Mel(f) represents the mel frequency; the obtained mel frequency is filtered through M triangular filter banks of mel scale, and the formula of the filter bank is:

In this formula:

Then calculate the logarithmic energy of each filter bank output:

Among them, E(m) is the logarithmic energy, and _Hm (k) is the filter bank;

36) Perform cepstral analysis on the Mel-filtered data to extract the MFCC coefficient C(n), and the MFCC coefficient C(n) is extracted by performing discrete cosine transform DCT:

Among them, M represents the number of filters.

The construction of the transformer fault detection model includes the following steps:

41) A transformer fault detection model is set based on a double-gated convolutional neural network, which includes two convolution gating layers, two pooling layers, a fully connected layer and an output layer;

42) Setting the convolution gating layer: the convolution gating layer extracts features through the convolution operation of the input data and the convolution kernel to obtain a feature map, and the depth of the feature map depends on the number of the set convolution kernels;

Assuming that the input is X∈R ^A×B , where A and B represent the length and width of the input data, respectively, the convolution operation is defined as:

代表第l层卷积层的第j个特征平面图，

代表第l-1层卷积层的第i个平面特征图，M代表平面特征图的个数，

代表卷积核，

代表偏置，

和

的具体参数通过训练优化来确定，f代表激励层函数，常用的激励函数有ReLu与sigmoid函数，其表达式如下：in:

represents the jth feature plane of the lth convolutional layer,

represents the ith plane feature map of the l-1th convolutional layer, M represents the number of plane feature maps,

represents the convolution kernel,

stands for bias,

and

The specific parameters of is determined by training optimization, f represents the excitation layer function, the commonly used excitation functions are ReLu and sigmoid functions, and their expressions are as follows:

The gated output is represented as:

表示矩阵的元素对应相乘，σ表示sigmoid门控函数，X表示上一层的输出；Among them: h(X) represents the output of Gated CNN, W and V represent different convolution kernels, b and c represent different biases,

Represents the corresponding multiplication of the elements of the matrix, σ represents the sigmoid gating function, and X represents the output of the previous layer;

43) Design the pooling layer: use the maximum pooling or average pooling method to set the pooling layer;

44) Build a fully connected layer and use the softmax classifier as the output layer for final classification output.

The training of the transformer fault detection model includes the following steps:

51) Input the MFCC coefficients extracted from the training sample set into the transformer fault detection model;

52) The MFCC coefficients are continuously iterated through forward propagation and backward propagation in the transformer fault detection model, and parameter optimization is performed to obtain the trained transformer fault detection model.

beneficial effect

A transformer fault detection method based on acoustic signals and deep learning technology of the present invention, compared with the prior art, can perform fault detection of transformers based on voiceprint signals (non-vibration signals); Two types of “normal” and “fault” are used to construct the sound data set of transformer working conditions; for the interference of environmental noise, the sound data set of transformer working conditions is pre-processed for denoising; considering the practicability of the training model, in order to enhance the subsequent construction Due to the robustness of the neural network, a series of data enhancement operations are performed on the data; after that, the transformer sound data set is sent to the network for training through MFCC feature extraction; when the basic CNN network is used for complex fault detection, due to the amount of audio data Very few or the audio duration is relatively long, there are low training efficiency and overfitting, and the detection accuracy is low.

Aiming at the transformer fault detection effect, the invention designs a double-gated convolutional neural network transformer sound detection model, which provides a favorable method basis and detection model realization for realizing the transformer fault online monitoring.

Description of drawings

Fig. 1 is the method sequence diagram of the present invention;

Fig. 2 is the method realization logic flow chart of the present invention;

FIG. 3 is a structural diagram of a double-gated convolutional neural network involved in the present invention.

Detailed ways

In order to have a further understanding and understanding of the structural features of the present invention and the effects achieved, the preferred embodiments and accompanying drawings are used in conjunction with detailed descriptions, and the descriptions are as follows:

As shown in Figure 1 and Figure 2, a transformer fault detection method based on acoustic signal and deep learning technology according to the present invention includes the following steps:

The first step is the acquisition of sound data of power transformers. The transformer sound data is collected on the spot by the voiceprint acquisition sensor, which is divided into two categories: "normal" and "fault" after labeling, and is defined as a training sample set. In the laboratory, the transformer sound data can be obtained and divided into two categories, normal working state and faulty working state, an experimental database can be constructed, and the training set and test set of the experimental database can be divided according to a certain proportion.

The second step is the preprocessing of the acoustic signals in the training sample set. The preprocessing methods of segmentation, framing, sound windowing and self-adaptive filtering are used to denoise and preprocess the collected power transformer sound data; and then the sound signal is enhanced by cutting, adding noise and tuning. The specific steps are as follows:

(1) Perform a segmented operation on the collected power transformer sound data s(t):

segment the obtained power transformer sound data,

Divide into s(t)={s ₁ (t), s ₂ (t),...,s _q (t),...s _r (t)}, and calculate the total length L of the voiceprint data, which is Calculated as follows:

L=time×fS _Sample =r×r _L ,

Among them, fs _sample is the sampling frequency of the sound, time is the sampling time, r is the number of segments, and r _L is the segment length.

(2) Framing processing of the segmented transformer sound data s _q (t):

Set the transformer voiceprint frame length to 500ms, and perform framing processing as

s _q (t)={s _q1 (t),s _q2 (t),...,s _qp (t),...s _qLength (t)},

Among them, the length of each frame is set to 500ms, and each segment is divided into Length frames.

(3) Windowing the transformer sound after framing:

The end-point smoothing window processing is performed on the framed data, and the frame is windowed by the Hamming window. The function w(t) of the Hamming window is expressed as follows:

where M is the frame length and t is the time;

The time domain signal f _qp (t) is obtained for each frame, and its expression is as follows:

f _qp (t)=s _qp (t)*w(n),

Among them, f _qp (t) is the time domain signal of the p-th frame of the q-th signal, w(n) is the windowing function, and s _qp (t) is the signal value of the p-th frame of the q-th signal.

(4) Use an adaptive filter to denoise the windowed sound:

Sampling f _qp (t) to obtain the X _i (n) digital signal sequence, and set the initial value of the filter as follows:

Among them, W(k) is the optimal weight coefficient, μ is the convergence factor, and λ is the maximum eigenvalue of the correlation matrix;

Compute the actual output estimate of the filter output:

y(k)=W ^T (k)Xi(n),

Among them, y(k) is the actual output estimated value, W ^T (k) is the transpose of the optimal weight coefficient, and X _i (n) is the input signal sequence;

e(k) is the error signal, which is the calculation error:

e(k)=d(k)-y(k),

Among them, d(k) is the expected output value, and the filter coefficients at time k+1 are updated:

W(k+1)=W(k)+μe(k)X(k),

Among them, W(k+1) is the optimal weight coefficient at time k+1, W(k) is the optimal weight coefficient at time k, e(k) is the error at time k, and X(k) is the error at time k input sequence;

The steepest descent method is used to iteratively solve the problem, so that the error signal is minimized, and the output y(k) after adaptive filtering and denoising is obtained.

(5) Perform data enhancement on the acoustic signal by cutting, adding noise, and tuning the output y(k) of the adaptive filtering and denoising.

The third step, sound feature extraction of sound signal data: same as speech recognition, for audio data, there are two application methods for fault detection, the first is to directly input one-dimensional signals for processing, and the second is to Convert it into an image to extract audio features for subsequent processing. One of the most commonly used features is Mel-scale Frequency Cepstral Coefficients, or MFCC for short. By using the Meral cepstral coefficients to extract the sound features of the preprocessed power transformer sound data, the MFCC coefficients are extracted, and the extracted MFCC coefficients are sent to the network later to generate feature maps. The specific steps are as follows:

(1) Perform inverse transformation on y(k) after data enhancement to generate s(t), and perform pre-emphasis on s(t). The calculation formula is as follows:

y(z)=s(z)H(z),

where y(z) is the output of the pre-emphasis, s(z) is the Z-domain representation of the signal s(t), and H(z) is the transfer function of the high-pass filter, which is:

This high-pass filter is in the form of the z-domain, where the value of μ is between 0.9-1.0. Pass the sound data through a high-pass filter to increase the energy of the high-frequency part, try to keep the energy of the high-frequency part and the low-frequency part consistent, and obtain the spectrum with the same signal-to-noise ratio in the entire frequency band.

(2) Framing processing of the pre-emphasized transformer sound data s _q (t):

The frame length of the transformer voiceprint is set to 30ms, and the transformer voice data is divided into frames. The processed structure is 30ms per frame.

(3) Windowing the framed sound data. The purpose of the windowing function is to increase the continuity between the left and right ends of each frame of sound by multiplying each frame of sound by the window function.

Use the hamming window to window the frame. Assuming that the output obtained after the first two steps of preprocessing is S(n) and the window function is W(n), the calculation formula is:

S'(n)=S(n)×W(n);

where the window function W(n) is:

In this formula, n represents time n, and its value range is 0≤n≤N-1, N is the number of sampling points, and different Hamming windows are obtained according to different a values, n represents the sequence independent variable, and a is the set constant .

(4) Fast Fourier transform is performed on the windowed data. Fast Fourier transform (fast Fourier transform, or FFT) is a commonly used way to transform a time domain signal into the frequency domain.

Perform FFT on each frame signal to get the frequency domain signal X(k):

Where X(k) represents the frequency domain output, x(n) represents the time domain input, and N is the number of sampling points.

(5) Perform mel filtering on the data after fast Fourier transform, and the conversion formula of mel filtering is:

Among them, f represents the physical frequency, and Mel(f) represents the mel frequency; the obtained mel frequency is filtered through M triangular filter banks of mel scale, and the formula of the filter bank is:

In this formula:

Then calculate the logarithmic energy of each filter bank output:

where E(m) is the logarithmic energy and _Hm (k) is the filter bank.

(6) Perform cepstral analysis on the Mel-filtered data, extract the MFCC coefficient C(n), and extract the MFCC coefficient by performing discrete cosine transform DCT:

Among them, M represents the number of filters. The larger M is, the more feature values are extracted in each frame, the more information is, and the more accurate the signal can be described.

The fourth step is to build a transformer fault detection model: build a transformer fault detection model based on the double-gated convolutional network model and the characteristics of the transformer acoustic signal.

The sound signal of the transformer can reflect the working status of the transformer. Traditional manual troubleshooting of transformer faults requires experienced manual on-site sound detection. This method of detecting transformer faults has a certain lag, that is, under normal circumstances, an experienced manual intervention is required to check the transformers one by one after a transformer fault occurs at a certain location. Manpower with rich experience in transformer troubleshooting requires years of practical training, which consumes a lot of time and cost. The transformer online monitoring system can quickly and easily monitor faults by using deep learning methods. At present, deep learning methods are similar to traditional fault detection methods. Most of them still use vibration data, and there are relatively few studies on fault detection using only audio data. . Compared with vibration data, audio data is easier to obtain and low cost, and is more suitable for large-scale practical applications. Therefore, the present invention proposes a method for fault detection using transformer audio data based on a double-gated convolutional neural network, and designs a model for transformer fault detection and identification based on acoustic signals.

Because the transformer is usually placed outdoors, its acoustic signal often has a lot of noise. When the energy of the noise is too large, the real acoustic signal emitted by the transformer itself will be submerged, resulting in a low effect of the trained detection model. The preceding steps of the present invention perform denoising preprocessing on the collected acoustic signals, and adopt adaptive filtering method to reduce noise energy and preserve the transformer's own acoustic signals. Transformer sound data, on the premise of removing noise, uses a series of data enhancement methods to expand the data set samples, so that the network can input more complex sample sets during training to enhance the robustness of the transformer detection model.

For the input processing of the audio data of the transformer, if the audio is directly input into the network training in a one-dimensional manner, the memory required for the training will be too large, which is physically difficult to achieve. If the one-dimensional sound data is directly input into the network as a time domain signal, the network can only extract a small part of the feature information, which is not conducive to the training of the network model. Therefore, before the data is input into the network, the MFCC coefficient feature extraction method is used to convert the time-domain signal into a time-frequency signal, which is input into the network in a two-dimensional signal manner. The MFCC coefficients are time-frequency signals, which contain more feature information. As a result, the network can extract more effective features and the trained model is better.

Due to the influence of the audio format, the general convolutional neural network often suffers from low training efficiency and overfitting, which leads to low fault detection accuracy. Different from the general convolutional neural network, the present invention adopts the network structure idea of gated convolution, and adds a gated switch to the output of the convolutional network. The function of the gating switch is to achieve a buffer control effect when the information is transmitted during training, so as to solve the overfitting phenomenon and improve the training efficiency. Aiming at the characteristics of the transformer sound, the present invention adds a convolution gating layer on the basis of the single-gated convolutional network, and designs a double-gated convolutional neural network. The double-gated convolutional neural network adopts two convolutional gating layers, which can effectively extract the deep features of the transformer's native sound, improve the training effect, and obtain a training model with excellent results.

The specific steps are as follows:

(1) As shown in Figure 3, a transformer fault detection model is set based on a double-gated convolutional neural network, which includes two convolution gating layers, two pooling layers, a fully connected layer and an output layer.

(2) Set the convolution gating layer: The convolution gating layer extracts features through the convolution operation between the input data and the convolution kernel, and obtains a feature map. The depth of the feature map depends on the number of set convolution kernels. ;

Assuming that the input is X∈R ^A×B , where A and B represent the length and width of the input data, respectively, the convolution operation is defined as:

代表第l层卷积层的第j个特征平面图，

代表第l-1层卷积层的第i个平面特征图，M代表平面特征图的个数，

代表卷积核，

代表偏置，

和

的具体参数通过训练优化来确定，f代表激励层函数，常用的激励函数有ReLu与sigmoid函数，其表达式如下：in:

represents the jth feature plane of the lth convolutional layer,

represents the ith plane feature map of the l-1th convolutional layer, M represents the number of plane feature maps,

represents the convolution kernel,

stands for bias,

and

The specific parameters of is determined by training optimization, f represents the excitation layer function, the commonly used excitation functions are ReLu and sigmoid functions, and their expressions are as follows:

The gated output is represented as:

表示矩阵的元素对应相乘，σ表示sigmoid门控函数，X表示上一层的输出。经过门控结构，降低了模型的梯度弥散，促进了模型的训练，简化了模型的结构，同时保留了模型的非线性表示能力；Among them: h(X) represents the output of Gated CNN, W and V represent different convolution kernels, b and c represent different biases,

The elements of the matrix represent the corresponding multiplication, σ represents the sigmoid gating function, and X represents the output of the previous layer. After the gated structure, the gradient dispersion of the model is reduced, the training of the model is promoted, the structure of the model is simplified, and the nonlinear representation ability of the model is retained;

(3) Design the pooling layer: When the convolution kernel is relatively small, the feature map after the convolution layer is still very large. At this time, a pooling operation is required to reduce the dimension, reduce the number of features, and enhance the network's ability to input features. Robustness, where max pooling or average pooling is used to set the pooling layer.

(4) Build a fully connected layer and use the softmax classifier as the output layer for the final classification output.

The fifth step is the training of the transformer fault detection model: the extracted MFCC coefficients are input into the transformer fault detection model for training. The specific steps are as follows:

(1) Input the MFCC coefficients extracted from the training sample set into the transformer fault detection model;

(2) The MFCC coefficients are iterated through forward propagation and back propagation in the transformer fault detection model, and the parameters are optimized to obtain the trained transformer fault detection model.

The sixth step, acquisition and preprocessing of the acoustic signal data of the transformer to be detected: Acquire the acoustic signal data of the transformer to be detected, perform denoising preprocessing, and extract MFCC coefficients from the denoised preprocessed acoustic signal data of the transformer to be detected.

Step 7: Obtaining the fault detection result of the transformer to be detected: Input the MFCC coefficients into the trained transformer fault detection model to obtain the transformer fault detection result.

The foregoing has shown and described the basic principles, main features and advantages of the present invention. It should be understood by those skilled in the art that the present invention is not limited by the above-mentioned embodiments. The above-mentioned embodiments and descriptions describe only the principles of the present invention. Without departing from the spirit and scope of the present invention, there are various Variations and improvements are intended to fall within the scope of the claimed invention. The scope of protection claimed by the present invention is defined by the appended claims and their equivalents.

Claims (3)

1. a transformer fault detection method based on acoustic signal and deep learning technology, is characterized in that, comprises the following steps: 11) Acquisition and acquisition of sound data of power transformers: The sound data of transformers are collected and obtained on the spot through the voiceprint acquisition sensor, and are marked into two categories: "normal" and "fault", and are defined as training sample sets; 12) Preprocessing of the acoustic signals in the training sample set: Denoising preprocessing is performed on the collected sound data of power transformers by using the preprocessing methods of segmentation, framing, sound windowing and adaptive filtering; Noise and tuning processing to enhance the sound signal; 13) Sound feature extraction of sound signal data: Extract the sound feature of the preprocessed power transformer sound data by using the Meral cepstral coefficient, and extract the MFCC coefficient; 14) Build a transformer fault detection model: build a transformer fault detection model based on the double-gated convolutional network model and the characteristics of the transformer acoustic signal; 15) Training of the transformer fault detection model: input the extracted MFCC coefficients into the transformer fault detection model for training; 16) Acquisition and preprocessing of the acoustic signal data of the transformer to be detected: Acquire the acoustic signal data of the transformer to be detected, perform denoising preprocessing, and extract the MFCC coefficients from the denoised preprocessed acoustic signal data of the transformer to be detected; 17) Obtaining the fault detection result of the transformer to be detected: input the MFCC coefficient into the trained transformer fault detection model to obtain the result of the transformer fault detection; The preprocessing of the acoustic signal in the training sample set includes the following steps: 21) Perform a segmented operation on the collected power transformer sound data s(t): segment the obtained power transformer sound data, Divide into s(t)={s ₁ (t), s ₂ (t),...,s _q (t),...s _r (t)}, and calculate the total length L of the voiceprint data, which is Calculated as follows: L=time×fS _Sample =r×r _L , Among them, fs _sample is the sampling frequency of the sound, time is the sampling time, r is the number of segments, and r _L is the length of the segment; 22) Framing processing of the segmented transformer sound data s _q (t): Set the transformer voiceprint frame length to 500ms, and perform framing processing as s _q (t)={s _q1 (t),s _q2 (t),...,s _qp (t),...s _qLength (t)}, Among them, the length of each frame is set to 500ms, and each segment is divided into Length frames; 23) Windowing the transformer sound after framing: The end-point smoothing window processing is performed on the framed data, and the frame is windowed by the Hamming window. The function w(t) of the Hamming window is expressed as follows:

where M is the frame length and t is the time; The time domain signal f _qp (t) is obtained for each frame, and its expression is as follows: f _qp (t)=s _qp (t)*w(n), Among them, f _qp (t) is the time domain signal of the p-th frame of the q-th segment signal, w(n) is the windowing function, and s _qp (t) is the p-th frame of the q-th segment signal. Signal value; 24) Use an adaptive filter to denoise the windowed sound: Sampling f _qp (t) to obtain the X _i (n) digital signal sequence, and set the initial value of the filter as follows:

Among them, W(k) is the optimal weight coefficient, μ is the convergence factor, and λ is the maximum eigenvalue of the correlation matrix; Compute the actual output estimate of the filter output: y(k)=W ^T (k)Xi(n), Among them, y(k) is the actual output estimated value, W ^T (k) is the transpose of the optimal weight coefficient, and X _i (n) is the input signal sequence; e(k) is the error signal, which is the calculation error: e(k)=d(k)-y(k), Among them, d(k) is the expected output value, and the filter coefficients at time k+1 are updated: W(k+1)=W(k)+μe(k)X(k), Among them, W(k+1) is the optimal weight coefficient at time k+1, W(k) is the optimal weight coefficient at time k, e(k) is the error at time k, and X(k) is the error at time k input sequence; Use the steepest descent method to iteratively solve the problem to minimize the error signal, and obtain the output y(k) after adaptive filtering and denoising; 25) Perform data enhancement on the acoustic signal by cutting, adding noise, and tuning the output y(k) of the adaptive filtering and denoising; The construction of the transformer fault detection model includes the following steps: 41) A transformer fault detection model is set based on a double-gated convolutional neural network, which includes two convolution gating layers, two pooling layers, a fully connected layer and an output layer; 42) Setting the convolution gating layer: the convolution gating layer extracts features through the convolution operation of the input data and the convolution kernel to obtain a feature map, and the depth of the feature map depends on the number of the set convolution kernels; Assuming that the input is X∈R ^A×B , where A and B represent the length and width of the input data, respectively, the convolution operation is defined as:

in:

represents the jth feature plane of the lth convolutional layer,

represents the ith plane feature map of the l-1th convolutional layer, M represents the number of plane feature maps,

represents the convolution kernel,

stands for bias,

and

The specific parameters of is determined by training optimization, f represents the excitation layer function, the excitation function is the sigmoid function, and its expression is as follows:

The gated output is represented as:

Among them: h(X) represents the output of GatedCNN, W and V represent different convolution kernels, b and c represent different biases,

Represents the corresponding multiplication of the elements of the matrix, σ represents the sigmoid gating function, and X represents the output of the previous layer; 43) Design the pooling layer: use the maximum pooling or average pooling method to set the pooling layer; 44) Build a fully connected layer and use the softmax classifier as the output layer for final classification output. 2. a kind of transformer fault detection method based on acoustic signal and deep learning technology according to claim 1, is characterized in that, the sound feature extraction of described acoustic signal data comprises the following steps: 31) Perform inverse transformation on y(k) after data enhancement to generate s(t), and perform pre-emphasis operation on s(t), and the calculation formula is as follows: y(z)=s(z)H(z), Among them, y(z) is the output of the pre-emphasis, s(z) is the Z domain representation of the signal s(t), H(z) is the transfer function of the high-pass filter, and the transfer function is: H(Z)=1 -|μz ^-1 , This high-pass filter is in the form of the z-domain, where the value of μ is between 0.9-1.0; 32) Perform frame division processing on the pre-emphasized transformer sound data s _q (t): Set the frame length of the transformer voiceprint to 30ms, and perform frame-by-frame processing of the transformer voice data. The processed structure is 30ms per frame; 33) Windowing the sound data after framing: use the hamming window to window the frame. Assuming that the output obtained after the first two steps of preprocessing is S(n) and the window function is W(n), the calculation formula is : S'(n)=S(n)×W(n); where the window function W(n) is:

In this formula, n represents time n, and its value range is 0≤n≤N-1, N is the number of sampling points, and different Hamming windows are obtained according to different a values, n represents the sequence independent variable, and a is the set constant ; 34) Perform fast Fourier transform on the windowed data: Perform FFT on each frame of signal to obtain the frequency domain signal W(k):

where X(k) represents the frequency domain output, x(n) represents the time domain input, and N is the number of sampling points; 35) Perform mel filtering on the data after the fast Fourier transform, and the conversion formula of mel filtering is:

Among them, f represents the physical frequency, and Mel(f) represents the mel frequency; the obtained mel frequency is filtered through M triangular filter banks of mel scale, and the formula of the filter bank is:

In this formula:

Then calculate the logarithmic energy of each filter bank output:

Among them, E(m) is the logarithmic energy, and _Hm (k) is the filter bank; 36) Perform cepstral analysis on the Mel-filtered data to extract the MFCC coefficient C(n), and the MFCC coefficient C(n) is extracted by performing discrete cosine transform DCT:

Among them, M represents the number of filters. 3. a kind of transformer fault detection method based on acoustic signal and deep learning technology according to claim 1, is characterized in that, the training of described transformer fault detection model comprises the following steps: 51) Input the MFCC coefficients extracted from the training sample set into the transformer fault detection model; 52) The MFCC coefficients are continuously iterated through forward propagation and backward propagation in the transformer fault detection model, and parameter optimization is performed to obtain the trained transformer fault detection model.

Images