CN116778956A - Transformer acoustic feature extraction and fault identification method - Google Patents
Transformer acoustic feature extraction and fault identification method Download PDFInfo
- Publication number
- CN116778956A CN116778956A CN202310537586.4A CN202310537586A CN116778956A CN 116778956 A CN116778956 A CN 116778956A CN 202310537586 A CN202310537586 A CN 202310537586A CN 116778956 A CN116778956 A CN 116778956A
- Authority
- CN
- China
- Prior art keywords
- data
- matrix
- transformer
- mfcc
- characteristic data
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 35
- 238000000605 extraction Methods 0.000 title claims abstract description 18
- 238000013145 classification model Methods 0.000 claims abstract description 27
- 230000005236 sound signal Effects 0.000 claims abstract description 26
- 230000004927 fusion Effects 0.000 claims abstract description 15
- 238000007781 pre-processing Methods 0.000 claims abstract description 13
- 239000011159 matrix material Substances 0.000 claims description 86
- 230000003595 spectral effect Effects 0.000 claims description 18
- 238000012549 training Methods 0.000 claims description 18
- 230000006870 function Effects 0.000 claims description 15
- 238000005070 sampling Methods 0.000 claims description 15
- 238000004364 calculation method Methods 0.000 claims description 12
- 230000000694 effects Effects 0.000 claims description 12
- 238000009432 framing Methods 0.000 claims description 9
- 238000001228 spectrum Methods 0.000 claims description 9
- 230000003068 static effect Effects 0.000 claims description 8
- 238000012545 processing Methods 0.000 claims description 7
- 230000008569 process Effects 0.000 claims description 6
- 230000004044 response Effects 0.000 claims description 6
- 238000012360 testing method Methods 0.000 claims description 6
- 230000001131 transforming effect Effects 0.000 claims description 6
- 238000012795 verification Methods 0.000 claims description 6
- 239000000284 extract Substances 0.000 claims description 5
- 238000011156 evaluation Methods 0.000 claims description 3
- 238000013507 mapping Methods 0.000 claims description 3
- 230000008447 perception Effects 0.000 claims description 3
- 238000001514 detection method Methods 0.000 abstract description 10
- 238000013527 convolutional neural network Methods 0.000 description 5
- 238000013135 deep learning Methods 0.000 description 5
- 238000010801 machine learning Methods 0.000 description 3
- 238000011160 research Methods 0.000 description 3
- 238000003745 diagnosis Methods 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 238000012706 support-vector machine Methods 0.000 description 2
- 241001480010 Cestrum Species 0.000 description 1
- 208000037170 Delayed Emergence from Anesthesia Diseases 0.000 description 1
- 230000002159 abnormal effect Effects 0.000 description 1
- 230000005856 abnormality Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000007796 conventional method Methods 0.000 description 1
- 238000003066 decision tree Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000012631 diagnostic technique Methods 0.000 description 1
- 230000008030 elimination Effects 0.000 description 1
- 238000003379 elimination reaction Methods 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 238000004880 explosion Methods 0.000 description 1
- WABPQHHGFIMREM-UHFFFAOYSA-N lead(0) Chemical compound [Pb] WABPQHHGFIMREM-UHFFFAOYSA-N 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 229910052573 porcelain Inorganic materials 0.000 description 1
- 230000035945 sensitivity Effects 0.000 description 1
Landscapes
- Complex Calculations (AREA)
Abstract
The invention provides a transformer acoustic feature extraction and fault identification method, which belongs to the field of power industry and transformer detection, and comprises the steps of utilizing a microphone to collect sound signals of a transformer in different working states, extracting MFCC features, ΔMFCC features and ΔΔMFCC features of the transformer after preprocessing, combining the three features, using ResNet as a classifier and using fusion feature data to train the classifier, and finally repeating feature extraction operation on sound data of the detected transformer in the current period and inputting the sound data into a feature classification model to obtain the current state of the detected transformer.
Description
Technical Field
The invention relates to the fields of power industry and transformer detection, in particular to a transformer acoustic feature extraction and fault identification method.
Background
The working state identification of the transformer is one of important research fields for guaranteeing safe and reliable operation of a power system, and the fault or abnormality detection of the transformer mainly comprises a detection method based on sound signals, a detection method based on gas and the like. The mechanical structure of the transformer determines that the sound signal is closely related to the mechanical state of the transformer, and different sounds occurring in operation can reflect different abnormal states of the transformer, such as overload, high-voltage porcelain bushing lead wire discharge and the like. Compared with other methods, the detection method based on the sound signal has a plurality of advantages: strong real-time performance, easy data acquisition, accurate detection result, non-contact detection and the like.
The method for detecting the faults of the transformer based on the sound signals can be divided into a method based on manual identification and a method based on signal feature extraction, and the method for manually identifying is characterized in that the sound signals generated by the transformer are manually heard, and fault diagnosis and classification are carried out according to experience and knowledge. The method based on signal feature extraction extracts feature parameters capable of reflecting faults of the transformer through collection, processing and feature extraction of sound signals, and then performs fault classification and diagnosis on the power equipment through a classifier. In terms of feature extraction, characteristic parameters commonly used in the existing research include Mel frequency cepstrum coefficient (Mel-frequency cepstral coefficients, MFCC), linear prediction cepstrum coefficient (Linear Prediction Cestrum Coefficient, LPCC), gammatine filter cepstrum coefficient (Gammatone Filter Cepstral Coefficient, GFCC) and the like, different features have respective characteristics, if different features can be extracted and simultaneously used as input of a classifier, various problems of inaccurate generalization, incomplete generalization and the like caused by using a single feature as a recognition basis can be solved to a certain extent, and meanwhile, the extracted different features are used as input of the classifier to help improve generalization capability and robustness of the whole method.
The classifier is used as an implementation part of classification and identification, has a key role in determining whether the transformer sound signal can be accurately identified, and the classifier used in the current field mainly comprises a support vector machine (Support Vector Machine, SVM), a decision tree and other traditional machine learning algorithms, and the traditional machine learning algorithms show the problems of weak generalization capability and the like along with the continuous development of research. In recent years, deep learning has been greatly advanced in diagnostic techniques in other fields, and the deep learning technique has the advantages of strong generalization capability, high accuracy and the like in fault recognition based on sound.
Therefore, the method for extracting various characteristic parameters and combining the characteristic parameters into the fusion characteristic is provided as the recognition basis, and the working state recognition of the transformer is realized by combining a convolutional neural network in the deep learning technology as a classifier.
Disclosure of Invention
The invention aims to overcome the defects of the prior art, and provides a transformer acoustic feature extraction and fault identification method based on MFCC and CNN.
In order to achieve the above purpose, the invention adopts the following technical scheme:
the invention provides a transformer acoustic feature extraction and fault identification method, which comprises the following steps:
s1, acquiring sound data of three states of normal operation of a transformer, discharge fault of the transformer and overload of the transformer;
s2, preprocessing data: respectively preprocessing the data acquired in the step S1, wherein the preprocessing comprises downsampling and pre-emphasis;
s3, extracting characteristics of the sound data: framing, windowing, fourier transforming, calculating power spectrum density, converting Mel frequency power spectrum density, and discrete cosine transform, extraction of MFCC calculating first and second order differences of MFCC thereby obtaining Δmfcc and ΔΔmfcc, respectively;
the MFCC of the data is used as a static feature, the ΔMFCC and ΔΔMFCC of the data are used as dynamic features, and the static features and the dynamic features of the data are collectively called feature data;
s4, combining the characteristic data: combining the characteristic data obtained in the step S3 to obtain fusion characteristic data;
s5, constructing a classification model: dividing the fusion characteristic data into a training set, a verification set and a test set, using a ResNet50 model as a classifier, and training the ResNet50 model to obtain a characteristic classification model;
s6, deploying the feature classification model, extracting acoustic features of the transformer in the current period, and identifying faults: and acquiring sound data of the detected transformer with fixed duration in the current period, carrying out preprocessing, feature extraction and feature data combination which are the same as those of the S2, the S3 and the S4 on the sound data of the detected transformer to obtain fused feature data in the current period, inputting the fused feature data as the feature classification model, and finally outputting a prediction result, namely the predicted current state of the detected transformer by the feature classification model.
Further, the S2 specifically is:
s201, voice data downsampling: in order to reduce the data volume, the calculation volume of the subsequent algorithm and improve the operation speed of the whole method, the voice data in the three states acquired by the S1 needs to be downsampled, and in the process, the Nyquist sampling theorem needs to be satisfied, namely the sampling rate after downsampling is more than twice the highest frequency of the signal;
s202, voice data pre-emphasis: in order to enhance the energy of the high-frequency signal and reduce the influence of the low-frequency signal, pre-emphasis processing is carried out on the down-sampled sound data obtained in the step S201; in audio signals, the energy of the high frequency signal is typically low, while the energy of the low frequency signal is high, which, if left untreated, can cause the information of the high frequency signal to be masked.
Further, the step S3 specifically includes:
s301, framing:
framing the preprocessed m-th frame audio signal x (N) obtained in the step S2, wherein the sampling point number of each frame, namely the length of each frame is N, 50% of overlapping is set between every two frames, and the time length of each frame is as follows when the sampling rate is fs:
s302, windowing:
windowing the signal of each sampling point n by using a Hamming window to reduce the frequency spectrum leakage effect of the discrete signal; assuming that the window function of the mth frame is w (m, n), the weighted signal of the window function of the mth frame is:
xw(m,n)＝x(n)×w(m,n)，0≤n＜N，0≤m＜M (2)
s303, fourier transform:
fourier transforming the windowed signal xw (m, n) to obtain a frequency domain representation:
X(m,k)＝FFT[xw(m,n)]，0≤k＜K (6)
s304, calculating the power spectral density
The power spectral density is a function describing the power distribution of a signal as a function of frequency; is the square of the fourier transform, representing the magnitude of the signal power at different frequencies, calculating the power spectral density PSD (m, k):
s305, converting the Mel frequency power spectral density:
typically implemented with a bank of triangular filters, each triangular filter corresponding to a mel frequency, the triangular filter being shaped like a triangle; the triangular filter filters and downsamples the signal on a Mel frequency scale, and extracts audio characteristics which are more similar to human ear perception frequency response;
defining a filter bank consisting of I triangular filters, each filter having a center frequency f (I), the frequency response being:
mapping the PSD (m, k) onto a Mel frequency scale using a triangular filter bank to obtain a Mel frequency power spectral density (MPSD):
s306, discrete cosine transform:
discrete Cosine Transform (DCT) is performed on the Mel frequency power spectral density to obtain MFCC characteristics of sound data:
wherein m represents the number of frames, j represents the dimension of the MFCC coefficient of the current frame, and I represents the number of triangular filters;
s307, a first-order difference and a second-order difference of the MFCC are calculated:
performing first-order difference and second-order difference calculation on the MFCC (m, j) obtained in the step S306, wherein the effect of this step is to enhance the dynamics of the characteristics so as to better reflect the time-varying properties of the transformer sound signal;
let ΔMFC (C, m) be obtained by first-order differential and second-order differential calculations, respectively 1 ) And ΔΔmfcc (m, j 2 )；
m represents the number of frames, j 1 And j 2 The dimensions of the current frame Δmfcc coefficients and Δmfcc coefficients are represented, respectively;
MFCC, Δmfcc, and ΔΔmfcc characteristic data are obtained through the S306 and the S307.
Further, the specific steps of S4 are as follows:
setting a space matrix as a (224 ), and filling the MFCCs, Δmfccs and ΔΔmfccs obtained in the step S3 into a respectively, wherein the specific process is as follows:
firstly, taking an MFCC characteristic data matrix, a delta MFCC characteristic data matrix and delta MFCC characteristic data (the total number of characteristic data M is more than or equal to 1344) with M equal to 1344, namely 1344 frames, extracting each characteristic data as a matrix with 1344 columns, setting the MFCC to take the first 12 dimensions (excluding the 0 th dimension) as 12 rows (the matrix size is 12 multiplied by 1344), the respective feature data matrices of the Δmfcc feature data matrix and the ΔΔmfcc feature data matrix are respectively 11 rows and 10 rows (feature of differential calculation) (Δmfcc matrix size 11×1344, ΔΔmfcc matrix size 10×1344);
then, cutting and filling the first 224 columns of the MFCC feature data matrix into a matrix A, wherein the first 12 rows in the matrix A are filled, the rest 225-1344 columns of the MFCC feature data matrix are cut in sequence (the next cutting 225-448 columns) in a similar operation, filling downwards in the matrix A from the thirteenth row (cutting and filling 13-24 rows), and so on, and finally, the MFCC feature data matrix is fully filled into the first 72 rows of the matrix A;
performing similar operation on the ΔMFCC characteristic data matrix, cutting and filling 1-224 columns of characteristic data (11 multiplied by 224) into 73 rd to 83 rd rows of the matrix A, and finally filling all the ΔMFCC characteristic data matrix into 73 to 138 rows of the matrix A, wherein the total number of the rows is 66;
and similarly operating the delta MFCC characteristic data matrix, cutting and filling 1-224 columns of characteristic data (10 multiplied by 224) into 139 th to 148 th rows of the matrix A, and finally filling the delta MFCC characteristic data matrix into 139 to 198 rows of the matrix A, wherein 60 rows are all formed.
The remaining 199 to 224 rows of matrix a that are not filled are all filled with 0 s.
The matrix A after filling operation contains 1344 frames of MFCC, ΔMFCC and ΔΔMFCC characteristic data, the step operation is carried out on the data of all frames, 1344 frames are sequentially cut (1345-2688 frames are filled into the matrix B,2689-4032 frames are filled into the matrix C until the filling of the total number M of the characteristic data is completed), and finally all the matrices obtain the fused characteristic data.
Further, the step S5 includes the steps of:
s501, data set division:
corresponding the fusion characteristic data according to the working states of the corresponding transformer, forming a data set by each working state and the corresponding fusion characteristic data, dividing the data set into a training set, a verification set and a test set according to 70%, 20% and 10%, and respectively representing the normal state of the transformer, the discharge fault of the transformer and the overload state of the transformer by labels a, b and c;
s502, model training:
using a ResNet50 model as a classifier, and training the ResNet50 model to obtain a characteristic classification model; to prevent overfitting, dropout is used as a loss function, and the optimizer is SGD (random gradient descent). The Accuracy (Accuracy), the Precision (Precision), the Recall (Recall), and the F1 value (F1 score) are introduced as evaluation indexes, wherein the Accuracy is the proportion of the number of samples correctly classified by the model to the total number of samples, the Precision is the proportion of the samples truly positive among all the samples predicted to positive, the Recall is the proportion of the samples accurately predicted to positive among all the samples truly positive, and the F1 value is the harmonic average of the Accuracy and the Recall, so that the prediction capability and the classification effect of the model can be comprehensively reflected to a certain extent.
Further, the specific step of S6 is as follows: and acquiring sound data of the detected transformer in a fixed duration in the current period, performing the same processing as the S2, the S3 and the S4 to obtain fused characteristic data in the current period, inputting the fused characteristic data as a characteristic classification model, and outputting a prediction result, namely the predicted current state of the detected transformer by the final characteristic classification model.
The beneficial effects of the invention are as follows: the method comprises the steps of extracting the MFCC static characteristics of the transformer sound signals, calculating the dynamic characteristics of the transformer sound signals, adopting a convolutional neural network based on deep learning, adding the fusion characteristics of the dynamic characteristics for training and identifying, and finally realizing the characteristic extraction and fault identification of the transformer sound signals. The method provided by the invention can effectively identify the normal working state, the discharge fault and the overload state of the transformer according to the working sound of the transformer; compared with the traditional detection method, the non-contact detection method for diagnosing the faults of the transformer by utilizing the sound signals has the advantages of no need of disassembly, strong real-time performance and the like; the dynamic characteristics are introduced to enable the model to learn the dynamic information of the sound signals, so that the recognition effect of the model is improved, and the recognition effect of the MFCC static characteristics is better than the recognition effect by only using a single characteristic; the convolutional neural network based on deep learning is used as the classifier, so that the problems that the conventional method which uses a traditional machine learning algorithm as the classifier needs to continuously optimize and adjust the classifier, has weak generalization capability, needs to manually extract features and the like are solved; resNet introducing cross-layer connection and residual blocks is used, so that the problems of gradient elimination and gradient explosion by using a traditional convolutional neural network are solved.
Drawings
FIG. 1 is a flow chart of a method for extracting acoustic features and identifying faults of a transformer according to the present invention;
FIG. 2 is a diagram showing the characteristics of the sound signal in the normal state of the transformer;
FIG. 3 is a diagram of acoustic signal characteristics when the transformer is overloaded;
fig. 4 is a training-derived feature classification model confusion matrix.
Detailed Description
The present invention will be described in further detail with reference to the accompanying drawings, in order to make the objects, technical solutions and advantages of the present invention more apparent. It should be understood that the specific embodiments described herein are for purposes of illustration only and are not intended to limit the scope of the invention.
Referring to fig. 1, a method for extracting acoustic features and identifying faults of a transformer includes the following steps:
s1, acquiring sound data of three states of normal operation of a transformer, discharge fault of the transformer and overload of the transformer;
specifically, the structure of the transformer determines that the sound signal is closely related to the running state of the transformer, and the sound generated during the operation of the transformer has the characteristics of rich low-frequency components, high amplitude and the like, so that the microphone for collecting the sound data has the characteristics of at least low-frequency, correspondingly good, high sensitivity, high signal-to-noise ratio, wide dynamic range and the like, and can be used for collecting the sound generated during the operation of the transformer.
And selecting a proper microphone according to the conditions to acquire sound data of the transformer in three states of normal operation, discharge fault and overload.
S2, preprocessing data: respectively preprocessing the data acquired in the step S1, wherein the preprocessing comprises downsampling and pre-emphasis;
s3, extracting characteristics of the sound data: framing, windowing, fourier transforming, calculating power spectrum density, converting Mel frequency power spectrum density, and discrete cosine transform, extraction of MFCC calculating first and second order differences of MFCC thereby obtaining Δmfcc and ΔΔmfcc, respectively;
the MFCC of the data is used as a static feature, the ΔMFCC and ΔΔMFCC of the data are used as dynamic features, and the static features and the dynamic features of the data are collectively called feature data;
s4, combining the characteristic data: combining the characteristic data obtained in the step S3 to obtain fusion characteristic data;
s5, constructing a classification model: dividing the fusion characteristic data into a training set, a verification set and a test set, using a ResNet50 model as a classifier, and training the ResNet50 model to obtain a characteristic classification model;
s6, deploying the feature classification model, extracting acoustic features of the transformer in the current period, and identifying faults: and acquiring sound data of the detected transformer with fixed duration in the current period, carrying out preprocessing, feature extraction and feature data combination which are the same as those of the S2, the S3 and the S4 on the sound data of the detected transformer to obtain fused feature data in the current period, inputting the fused feature data as the feature classification model, and finally outputting a prediction result, namely the predicted current state of the detected transformer by the feature classification model.
The step S2 is specifically as follows:
s201, voice data downsampling: in order to reduce the data volume, the calculation volume of the subsequent algorithm and improve the operation speed of the whole method, the voice data in the three states acquired by the S1 needs to be downsampled, and in the process, the Nyquist sampling theorem needs to be satisfied, namely the sampling rate after downsampling is more than twice the highest frequency of the signal;
s202, voice data pre-emphasis: in order to enhance the energy of the high-frequency signal and reduce the influence of the low-frequency signal, pre-emphasis processing is carried out on the down-sampled sound data obtained in the step S201; in audio signals, the energy of the high frequency signal is typically low, while the energy of the low frequency signal is high, which, if left untreated, can cause the information of the high frequency signal to be masked.
The step S3 is specifically as follows:
s301, framing:
framing the preprocessed m-th frame audio signal x (N) obtained in the step S2, wherein the sampling point number of each frame, namely the length of each frame is N, 50% of overlapping is set between every two frames, and the time length of each frame is as follows when the sampling rate is fs:
s302, windowing:
windowing the signal of each sampling point n by using a Hamming window to reduce the frequency spectrum leakage effect of the discrete signal; assuming that the window function of the mth frame is w (m, n), the weighted signal of the window function of the mth frame is:
xw(m,n)＝x(n)×w(m,n)，0≤n＜N，0≤m＜M (2)
s303, fourier transform:
fourier transforming the windowed signal xw (m, n) to obtain a frequency domain representation:
X(m,k)＝FFT[xw(m,n)]，0≤k＜K (11)
s304, calculating the power spectral density
The power spectral density is a function describing the power distribution of a signal as a function of frequency; is the square of the fourier transform, representing the magnitude of the signal power at different frequencies, calculating the power spectral density PSD (m, k):
s305, converting the Mel frequency power spectral density:
typically implemented with a bank of triangular filters, each triangular filter corresponding to a mel frequency, the triangular filter being shaped like a triangle; the triangular filter filters and downsamples the signal on a Mel frequency scale, and extracts audio characteristics which are more similar to human ear perception frequency response;
defining a filter bank consisting of I triangular filters, each filter having a center frequency f (I), the frequency response being:
mapping the PSD (m, k) onto a Mel frequency scale using a triangular filter bank to obtain a Mel frequency power spectral density (MPSD):
s306, discrete cosine transform:
discrete Cosine Transform (DCT) is performed on the Mel frequency power spectral density to obtain MFCC characteristics of sound data:
wherein m represents the number of frames, j represents the dimension of the MFCC coefficient of the current frame, and I represents the number of triangular filters;
s307, a first-order difference and a second-order difference of the MFCC are calculated:
performing first-order difference and second-order difference calculation on the MFCC (m, j) obtained in the step S306, wherein the effect of this step is to enhance the dynamics of the characteristics so as to better reflect the time-varying properties of the transformer sound signal;
let ΔMFC (C, m) be obtained by first-order differential and second-order differential calculations, respectively 1 ) And ΔΔmfcc (m, j 2 )；
m represents the number of frames, j 1 And j 2 The dimensions of the current frame Δmfcc coefficients and Δmfcc coefficients are represented, respectively;
MFCC, Δmfcc, and ΔΔmfcc characteristic data are obtained through the S306 and the S307.
Three characteristics obtained by calculating the sound signal in the normal state of the transformer are shown in figure 2, and three characteristics obtained by calculating the sound signal in the overload state of the transformer are shown in figure 3.
The specific steps of the S4 are as follows:
setting a space matrix as a (224 ), and filling the MFCCs, Δmfccs and ΔΔmfccs obtained in the step S3 into a respectively, wherein the specific process is as follows:
firstly, taking an MFCC characteristic data matrix, a delta MFCC characteristic data matrix and delta MFCC characteristic data (the total number of characteristic data M is more than or equal to 1344) with M equal to 1344, namely 1344 frames, extracting each characteristic data as a matrix with 1344 columns, setting the MFCC to take the first 12 dimensions (excluding the 0 th dimension) as 12 rows (the matrix size is 12 multiplied by 1344), the respective feature data matrices of the Δmfcc feature data matrix and the ΔΔmfcc feature data matrix are respectively 11 rows and 10 rows (feature of differential calculation) (Δmfcc matrix size 11×1344, ΔΔmfcc matrix size 10×1344);
then, cutting and filling the first 224 columns of the MFCC feature data matrix into a matrix A, wherein the first 12 rows in the matrix A are filled, the rest 225-1344 columns of the MFCC feature data matrix are cut in sequence (the next cutting 225-448 columns) in a similar operation, filling downwards in the matrix A from the thirteenth row (cutting and filling 13-24 rows), and so on, and finally, the MFCC feature data matrix is fully filled into the first 72 rows of the matrix A;
performing similar operation on the ΔMFCC characteristic data matrix, cutting and filling 1-224 columns of characteristic data (11 multiplied by 224) into 73 rd to 83 rd rows of the matrix A, and finally filling all the ΔMFCC characteristic data matrix into 73 to 138 rows of the matrix A, wherein the total number of the rows is 66;
and similarly operating the delta MFCC characteristic data matrix, cutting and filling 1-224 columns of characteristic data (10 multiplied by 224) into 139 th to 148 th rows of the matrix A, and finally filling the delta MFCC characteristic data matrix into 139 to 198 rows of the matrix A, wherein 60 rows are all formed.
The remaining 199 to 224 rows of matrix a that are not filled are all filled with 0 s.
The matrix A after filling operation contains 1344 frames of MFCC, ΔMFCC and ΔΔMFCC characteristic data, the step operation is carried out on the data of all frames, 1344 frames are sequentially cut (1345-2688 frames are filled into the matrix B,2689-4032 frames are filled into the matrix C until the filling of the total number M of the characteristic data is completed), and finally all the matrices obtain the fused characteristic data.
The step S5 comprises the following steps:
s501, data set division:
corresponding the fusion characteristic data according to the working states of the corresponding transformer, forming a data set by each working state and the corresponding fusion characteristic data, dividing the data set into a training set, a verification set and a test set according to 70%, 20% and 10%, and respectively representing the normal state of the transformer, the discharge fault of the transformer and the overload state of the transformer by labels a, b and c;
s502, model training:
using a ResNet50 model as a classifier, and training the ResNet50 model to obtain a characteristic classification model; to prevent overfitting, dropout is used as a loss function, and the optimizer is SGD (random gradient descent). The Accuracy (Accuracy), the Precision (Precision), the Recall (Recall), and the F1 value (F1 score) are introduced as evaluation indexes, wherein the Accuracy is the proportion of the number of samples correctly classified by the model to the total number of samples, the Precision is the proportion of the samples truly positive among all the samples predicted to positive, the Recall is the proportion of the samples accurately predicted to positive among all the samples truly positive, and the F1 value is the harmonic average of the Accuracy and the Recall, so that the prediction capability and the classification effect of the model can be comprehensively reflected to a certain extent.
After training, a feature classification model is obtained, and the confusion matrix is shown in figure 4.
The specific steps of the S6 are as follows: and acquiring sound data of the detected transformer in a fixed duration in the current period, performing the same processing as the S2, the S3 and the S4 to obtain fused characteristic data in the current period, inputting the fused characteristic data as a characteristic classification model, and outputting a prediction result, namely the predicted current state of the detected transformer by the final characteristic classification model.
The foregoing examples merely illustrate embodiments of the invention and are described in more detail and are not to be construed as limiting the scope of the invention. It should be noted that it will be apparent to those skilled in the art that several variations and modifications can be made without departing from the spirit of the invention, which are all within the scope of the invention. Accordingly, the scope of protection of the present patent is to be determined by the appended claims.
Claims (6)
1. The method for extracting acoustic characteristics and identifying faults of the transformer is characterized by comprising the following steps of: the method comprises the following steps:
s1, acquiring sound data of three states of normal operation of a transformer, discharge fault of the transformer and overload of the transformer;
s2, preprocessing data: respectively preprocessing the data acquired in the step S1, wherein the preprocessing comprises downsampling and pre-emphasis;
s3, extracting characteristics of the sound data: framing, windowing, fourier transforming, calculating power spectrum density, converting Mel frequency power spectrum density, and discrete cosine transform, extraction of MFCC calculating first and second order differences of MFCC thereby obtaining Δmfcc and ΔΔmfcc, respectively;
the MFCC of the data is used as a static feature, the ΔMFCC and ΔΔMFCC of the data are used as dynamic features, and the static features and the dynamic features of the data are collectively called feature data;
s4, combining the characteristic data: combining the characteristic data obtained in the step S3 to obtain fusion characteristic data;
s5, constructing a classification model: dividing the fusion characteristic data into a training set, a verification set and a test set, using a ResNet50 model as a classifier, and training the ResNet50 model to obtain a characteristic classification model;
s6, deploying the feature classification model, extracting acoustic features of the transformer in the current period, and identifying faults: and acquiring sound data of the detected transformer with fixed duration in the current period, carrying out preprocessing, feature extraction and feature data combination which are the same as those of the S2, the S3 and the S4 on the sound data of the detected transformer to obtain fused feature data in the current period, inputting the fused feature data as the feature classification model, and finally outputting a prediction result, namely the predicted current state of the detected transformer by the feature classification model.
2. The method for extracting acoustic features and identifying faults of a transformer according to claim 1, wherein the step S2 is specifically:
s201, voice data downsampling: in order to reduce the data volume, the calculation volume of the subsequent algorithm and improve the operation speed of the whole method, the voice data in the three states acquired by the S1 needs to be downsampled, and in the process, the Nyquist sampling theorem needs to be satisfied, namely the sampling rate after downsampling is more than twice the highest frequency of the signal;
s202, voice data pre-emphasis: in order to enhance the energy of the high-frequency signal and reduce the influence of the low-frequency signal, pre-emphasis processing is carried out on the down-sampled sound data obtained in the step S201; in audio signals, the energy of the high frequency signal is typically low, while the energy of the low frequency signal is high, which, if left untreated, can cause the information of the high frequency signal to be masked.
3. The method for extracting acoustic features and identifying faults of a transformer according to claim 2, wherein the step S3 is specifically:
s301, framing:
framing the preprocessed m-th frame audio signal x (N) obtained in the step S2, wherein the sampling point number of each frame, namely the length of each frame is N, 50% of overlapping is set between every two frames, and the time length of each frame is as follows when the sampling rate is fs:
s302, windowing:
windowing the signal of each sampling point n by using a Hamming window to reduce the frequency spectrum leakage effect of the discrete signal; assuming that the window function of the mth frame is w (m, n), the weighted signal of the window function of the mth frame is:
xw(m,n)＝x(n)×w(m,n)，0≤n＜N，0≤m＜M (2)
s303, fourier transform:
fourier transforming the windowed signal xw (m, n) to obtain a frequency domain representation:
X(m,k)＝FFT[xw(m,n)]，0≤k＜K (1)
s304, calculating the power spectral density
The power spectral density is a function describing the power distribution of a signal as a function of frequency; is the square of the fourier transform, representing the magnitude of the signal power at different frequencies, calculating the power spectral density PSD (m, k):
s305, converting the Mel frequency power spectral density:
typically implemented with a bank of triangular filters, each triangular filter corresponding to a mel frequency, the triangular filter being shaped like a triangle; the triangular filter filters and downsamples the signal on a Mel frequency scale, and extracts audio characteristics which are more similar to human ear perception frequency response;
defining a filter bank consisting of I triangular filters, each filter having a center frequency f (I), the frequency response being:
mapping the PSD (m, k) onto a Mel frequency scale using a triangular filter bank to obtain a Mel frequency power spectral density (MPSD):
s306, discrete cosine transform:
discrete Cosine Transform (DCT) is performed on the Mel frequency power spectral density to obtain MFCC characteristics of sound data:
wherein m represents the number of frames, j represents the dimension of the MFCC coefficient of the current frame, and I represents the number of triangular filters;
s307, a first-order difference and a second-order difference of the MFCC are calculated:
performing first-order difference and second-order difference calculation on the MFCC (m, j) obtained in the step S306, wherein the effect of this step is to enhance the dynamics of the characteristics so as to better reflect the time-varying properties of the transformer sound signal;
let ΔMFC (C, m) be obtained by first-order differential and second-order differential calculations, respectively 1 ) And ΔΔmfcc (m, j 2 )；
m represents the number of frames, j 1 And j 2 The dimensions of the current frame Δmfcc coefficients and Δmfcc coefficients are represented, respectively;
MFCC, Δmfcc, and ΔΔmfcc characteristic data are obtained through the S306 and the S307.
4. The method for extracting acoustic features and identifying faults of a transformer according to claim 3, wherein the specific step of S4 is as follows:
setting a space matrix as a (224 ), and filling the MFCCs, Δmfccs and ΔΔmfccs obtained in the step S3 into a respectively, wherein the specific process is as follows:
firstly, taking an MFCC characteristic data matrix, a delta MFCC characteristic data matrix and delta MFCC characteristic data (the total number of characteristic data M is more than or equal to 1344) with M equal to 1344, namely 1344 frames, extracting each characteristic data as a matrix with 1344 columns, setting the MFCC to take the first 12 dimensions (excluding the 0 th dimension) as 12 rows (the matrix size is 12 multiplied by 1344), the respective feature data matrices of the Δmfcc feature data matrix and the ΔΔmfcc feature data matrix are respectively 11 rows and 10 rows (feature of differential calculation) (Δmfcc matrix size 11×1344, ΔΔmfcc matrix size 10×1344);
then, cutting and filling the first 224 columns of the MFCC feature data matrix into a matrix A, wherein the first 12 rows in the matrix A are filled, the rest 225-1344 columns of the MFCC feature data matrix are cut in sequence (the next cutting 225-448 columns) in a similar operation, filling downwards in the matrix A from the thirteenth row (cutting and filling 13-24 rows), and so on, and finally, the MFCC feature data matrix is fully filled into the first 72 rows of the matrix A;
performing similar operation on the ΔMFCC characteristic data matrix, cutting and filling 1-224 columns of characteristic data (11 multiplied by 224) into 73 rd to 83 rd rows of the matrix A, and finally filling all the ΔMFCC characteristic data matrix into 73 to 138 rows of the matrix A, wherein the total number of the rows is 66;
and similarly operating the delta MFCC characteristic data matrix, cutting and filling 1-224 columns of characteristic data (10 multiplied by 224) into 139 th to 148 th rows of the matrix A, and finally filling the delta MFCC characteristic data matrix into 139 to 198 rows of the matrix A, wherein 60 rows are all formed.
The remaining 199 to 224 rows of matrix a that are not filled are all filled with 0 s.
The matrix A after filling operation contains 1344 frames of MFCC, ΔMFCC and ΔΔMFCC characteristic data, the step operation is carried out on the data of all frames, 1344 frames are sequentially cut (1345-2688 frames are filled into the matrix B,2689-4032 frames are filled into the matrix C until the filling of the total number M of the characteristic data is completed), and finally all the matrices obtain the fused characteristic data.
5. The method for extracting acoustic features and identifying faults of a transformer according to claim 4, wherein the step S5 comprises the following steps:
s501, data set division:
corresponding the fusion characteristic data according to the working states of the corresponding transformer, forming a data set by each working state and the corresponding fusion characteristic data, dividing the data set into a training set, a verification set and a test set according to 70%, 20% and 10%, and respectively representing the normal state of the transformer, the discharge fault of the transformer and the overload state of the transformer by labels a, b and c;
s502, model training:
using a ResNet50 model as a classifier, and training the ResNet50 model to obtain a characteristic classification model; to prevent overfitting, dropout is used as a loss function, and the optimizer is SGD (random gradient descent). The Accuracy (Accuracy), the Precision (Precision), the Recall (Recall), and the F1 value (F1 score) are introduced as evaluation indexes, wherein the Accuracy is the proportion of the number of samples correctly classified by the model to the total number of samples, the Precision is the proportion of the samples truly positive among all the samples predicted to positive, the Recall is the proportion of the samples accurately predicted to positive among all the samples truly positive, and the F1 value is the harmonic average of the Accuracy and the Recall, so that the prediction capability and the classification effect of the model can be comprehensively reflected to a certain extent.
6. The method for extracting acoustic features and identifying faults of a transformer according to claim 5, wherein the specific step of S6 is as follows: and acquiring sound data of the detected transformer in a fixed duration in the current period, performing the same processing as the S2, the S3 and the S4 to obtain fused characteristic data in the current period, inputting the fused characteristic data as a characteristic classification model, and outputting a prediction result, namely the predicted current state of the detected transformer by the final characteristic classification model.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310537586.4A CN116778956A (en) | 2023-05-11 | 2023-05-11 | Transformer acoustic feature extraction and fault identification method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310537586.4A CN116778956A (en) | 2023-05-11 | 2023-05-11 | Transformer acoustic feature extraction and fault identification method |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN116778956A true CN116778956A (en) | 2023-09-19 |
Family
ID=87986907
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202310537586.4A Pending CN116778956A (en) | 2023-05-11 | 2023-05-11 | Transformer acoustic feature extraction and fault identification method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN116778956A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117131366A (en) * | 2023-10-26 | 2023-11-28 | 北京国电通网络技术有限公司 | Transformer maintenance equipment control method and device, electronic equipment and readable medium |
| CN117232644A (en) * | 2023-11-13 | 2023-12-15 | 国网吉林省电力有限公司辽源供电公司 | Transformer sound monitoring fault diagnosis method and system based on acoustic principle |
-
2023
- 2023-05-11 CN CN202310537586.4A patent/CN116778956A/en active Pending
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117131366A (en) * | 2023-10-26 | 2023-11-28 | 北京国电通网络技术有限公司 | Transformer maintenance equipment control method and device, electronic equipment and readable medium |
| CN117131366B (en) * | 2023-10-26 | 2024-02-06 | 北京国电通网络技术有限公司 | Transformer maintenance equipment control method and device, electronic equipment and readable medium |
| CN117232644A (en) * | 2023-11-13 | 2023-12-15 | 国网吉林省电力有限公司辽源供电公司 | Transformer sound monitoring fault diagnosis method and system based on acoustic principle |
| CN117232644B (en) * | 2023-11-13 | 2024-01-09 | 国网吉林省电力有限公司辽源供电公司 | Transformer sound monitoring fault diagnosis method and system based on acoustic principle |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108369813B (en) | Specific voice recognition method, apparatus and storage medium | |
| CN110534118A (en) | Transformer/reactor method for diagnosing faults based on Application on Voiceprint Recognition and neural network | |
| CN109856517B (en) | Method for distinguishing partial discharge detection data of extra-high voltage equipment | |
| CN116778956A (en) | Transformer acoustic feature extraction and fault identification method | |
| CN111653289B (en) | Playback voice detection method | |
| CN109473120A (en) | A kind of abnormal sound signal recognition method based on convolutional neural networks | |
| CN110570880B (en) | Snore signal identification method | |
| CN105206270A (en) | Isolated digit speech recognition classification system and method combining principal component analysis (PCA) with restricted Boltzmann machine (RBM) | |
| Zhang et al. | Fault identification based on PD ultrasonic signal using RNN, DNN and CNN | |
| CN111951824A (en) | Detection method for distinguishing depression based on sound | |
| CN111986699B (en) | Sound event detection method based on full convolution network | |
| CN112599134A (en) | Transformer sound event detection method based on voiceprint recognition | |
| CN101829689A (en) | Drift fault recognition method of hot-rolling strip steel based on sound signals | |
| Zhai et al. | Pattern recognition approach to identify loose particle material based on modified MFCC and HMMs | |
| Linh et al. | MFCC-DTW algorithm for speech recognition in an intelligent wheelchair | |
| CN115659248A (en) | Power equipment defect identification method, device, equipment and storage medium | |
| Dewi et al. | Analysis of LFCC feature extraction in baby crying classification using KNN | |
| Wiśniewski et al. | Automatic detection of disorders in a continuous speech with the hidden Markov models approach | |
| Rahman et al. | Dynamic time warping assisted svm classifier for bangla speech recognition | |
| CN111933186B (en) | Method, device and system for fault identification of on-load tap-changer | |
| Kharamat et al. | Durian ripeness classification from the knocking sounds using convolutional neural network | |
| CN112052712A (en) | Power equipment state monitoring and fault identification method and system | |
| Saritha et al. | Enhancement in speaker recognition using sincnet through optimal window and frame shift | |
| CN110931024B (en) | Audio-based prediction method and system for natural mating result of captive pandas | |
| CN107644651A (en) | Circuit and method for speech recognition |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |