A transformer acoustic feature extraction and fault identification method

Abstract

The application provides a transformer acoustic feature extraction and fault recognition method, and belongs to the field of power industry and transformer detection, comprising the following steps: collecting sound signals of a transformer under different working states by using a microphone; extracting MFCC features, Delta MFCC features and Delta Delta MFCC features after pretreatment; combining the three kinds of features; using ResNet as a classifier and training the classifier by using fusion feature data; and finally repeating the feature extraction operation on the sound data of the transformer to be detected in the current period and inputting the sound data into the feature classification model to obtain the current state of the transformer to be detected.

Description

Technical Field

[0001] The present invention relates to the fields of electric power industry and transformer detection, and in particular to a transformer acoustic feature extraction and fault identification method. Background Art

[0002] Transformer operating status identification is a key research area for ensuring the safe and reliable operation of power systems. Transformer fault or anomaly detection primarily involves methods based on acoustic signals and gas. The mechanical structure of a transformer dictates that its acoustic signal is closely related to its mechanical state. Different sounds produced during operation can reflect different abnormal transformer conditions, such as overload, overloading, and discharge in high-voltage porcelain bushing leads. Furthermore, long-term transformer operation can also cause changes in the mechanical structure, which can affect the audible sound signal. Compared to other methods, acoustic signal-based detection offers numerous advantages: strong real-time performance, easy data acquisition, accurate detection results, and non-contact detection.

[0003] Transformer fault detection methods based on sound signals can be categorized into those based on manual recognition and those based on signal feature extraction. Manual recognition methods primarily rely on listening to the sound signals generated by the transformer and performing fault diagnosis and classification based on experience and knowledge. While this method is intuitive and simple, it is highly subjective, and diagnostic accuracy and efficiency are difficult to guarantee. Signal feature extraction methods, on the other hand, extract characteristic parameters that reflect transformer faults through sound signal acquisition, processing, and feature extraction. These parameters are then used to classify and diagnose power equipment faults using a classifier. In terms of feature extraction, commonly used feature parameters in existing research include Mel-frequency cepstral coefficients (MFCC), linear prediction cepstral coefficients (LPCC), and gammatone filter cepstral coefficients (GFCC). Different features have their own unique characteristics. Extracting different features and using them simultaneously as classifier inputs can, to a certain extent, address the inaccurate and incomplete summaries associated with using a single feature as the recognition basis. Using different extracted features as classifier inputs can also help improve the generalization and robustness of the overall method.

[0004] And the classifier as the implementation part of the classification and recognition, it also has a key role in deciding whether the transformer sound signal can be accurately identified, the current field uses the classifier mainly has the traditional machine learning algorithm such as support vector machine (Support Vector Machine, SVM) and decision tree, and with the continuous development of research, these traditional machine learning algorithms show weak generalization ability and other problems.In recent years, deep learning has made more progress in the diagnosis technology of other fields, and deep learning technology shows strong generalization ability and high accuracy in sound-based fault recognition.

[0005] Therefore, a method of extracting multiple feature parameters and combining them into fusion features as identification basis, combining the convolutional neural network in deep learning technology as the classifier is proposed to realize the transformer working state recognition. SUMMARY

[0006] The present application aims at the deficiencies of the prior art, and provides a transformer acoustic feature extraction and fault recognition method, a transformer acoustic feature extraction and fault recognition method based on MFCC and CNN.

[0007] To achieve the above object, the present application adopts the following technical scheme:

[0008] The present application provides a transformer acoustic feature extraction and fault recognition method, comprising the following steps:

[0009] S1, obtaining sound data of three states of transformer normal operation, transformer discharge fault and transformer overload;

[0010] S2, data preprocessing: preprocessing the data obtained in S1 respectively, preprocessing includes downsampling and pre-emphasis;

[0011] S3, feature extraction of sound data: the data that has been preprocessed is sequentially framed, windowed, Fourier transformed, power spectral density calculated, mel frequency power spectrum density converted, discrete cosine transformed, MFCC extracted, first and second order differences of MFCC calculated to obtain ΔMFCC and ΔΔMFCC respectively;

[0012] The MFCC of the data is the static feature, the ΔMFCC and ΔΔMFCC of the data are the dynamic features, and the static and dynamic features of the data are collectively referred to as feature data;

[0013] S4, combination of the feature data: combining the feature data obtained in S3 to obtain fusion feature data;

[0014] S5. Construction of a classification model: Divide the fused feature data into a training set, a validation set, and a test set, use a ResNet50 model as a classifier, and train the ResNet50 model to obtain a feature classification model;

[0015] S6. Deployment of the feature classification model and acoustic feature extraction and fault identification of the transformer in the current period: Acquire sound data of the detected transformer of a fixed duration in the current period, and perform the same preprocessing, feature extraction, and feature data combination on the sound data of the detected transformer as in S2, S3, and S4 to obtain fused feature data for the current period, which is used as input to the feature classification model. Finally, the feature classification model outputs a prediction result, i.e., the predicted current state of the detected transformer.

[0016] Furthermore, the S2 is specifically:

[0017] S201, sound data downsampling: To reduce the amount of data, the amount of computation required for subsequent algorithms, and to increase the speed of the entire method, the sound data collected in the three states in S1 need to be downsampled. During this process, it is important to ensure that the Nyquist sampling theorem is met, i.e., the sampling rate after downsampling must be greater than twice the highest frequency of the signal.

[0018] S202, sound data pre-emphasis: To enhance the energy of high-frequency signals and reduce the impact of low-frequency signals, pre-emphasize the downsampled sound data obtained in S201. In audio signals, the energy of high-frequency signals is usually low, while the energy of low-frequency signals is high. If no pre-emphasis is performed, the information of the high-frequency signals will be obscured.

[0019] Furthermore, the S3 is specifically:

[0020] S301, framing:

[0021] The pre-processed audio signal x(n) of the mth frame obtained by S2 is divided into frames. The number of sampling points of each frame, i.e., the length of each frame is N. A 50% overlap is set between every two frames. Assuming the sampling rate is fs, the time length of each frame is:

[0022]

[0023] S302, windowing:

[0024] The signal of each sampling point n is windowed using a Hamming window to reduce the spectrum leakage effect of the discrete signal. Assuming that the window function of the mth frame is w(m,n), the signal after the weighted window function of the mth frame is:

[0025] xw(m,n)=x(n)×w(m,n),0≤n<N,0≤m<M (2)

[0026] S303, Fourier transform:

[0027] Perform Fourier transform on the windowed signal xw(m,n) and find the frequency domain representation:

[0028] X(m,k)=FFT[xw(m,n)],0≤k<K (6)

[0029] S304, calculate power spectrum density

[0030] Power spectral density is a function that describes how the signal power distribution changes with frequency. It is the square of the Fourier transform and represents the signal power at different frequencies. The power spectral density PSD(m,k) is calculated as:

[0031]

[0032] S305, convert Mel frequency power spectrum density:

[0033] It is usually implemented using a triangular filter bank, where each triangular filter corresponds to a Mel frequency and the shape of the triangular filter is similar to a triangle. The triangular filter filters and downsamples the signal on the Mel frequency scale to extract audio features that are closer to the frequency response perceived by the human ear.

[0034] Define a filter bank consisting of I triangular filters, each with a center frequency of f(i), and its frequency response is:

[0035]

[0036] Map PSD(m,k) to the Mel frequency scale using a triangular filter bank to obtain the Mel frequency power spectral density (MPSD):

[0037]

[0038] S306, Discrete Cosine Transform:

[0039] Perform discrete cosine transform (DCT) on the Mel frequency power spectral density to obtain the MFCC features of the sound data:

[0040]

[0041] Where m represents the number of frames, j represents the dimension of the MFCC coefficients of the current frame, and I represents the number of triangular filters;

[0042] S307, calculate the first-order difference and second-order difference of MFCC:

[0043] Performing first-order and second-order difference calculations on the MFCC (m, j) obtained in S306 . This step is used to enhance the dynamics of the feature to better reflect the time-varying nature of the transformer sound signal.

[0044] Assume that ΔMFC(C,m1) and ΔΔMFCC(m,j2) are obtained by first-order difference and second-order difference calculation respectively;

[0045] m represents the number of frames, j1 and j2 represent the dimensions of the ΔMFCC coefficients and ΔΔMFCC coefficients of the current frame, respectively;

[0046] After S306 and S307 , MFCC, ΔMFCC and ΔΔMFCC feature data are obtained.

[0047] Furthermore, the specific steps of S4 are:

[0048] Assume that an empty matrix is ​​A(224, 224), and fill the MFCC, ΔMFCC, and ΔΔMFCC obtained in S3 into A respectively. The specific process is as follows:

[0049] First, take m equal to 1344, that is, 1344 frames of MFCC feature data matrix, ΔMFCC feature data matrix and ΔΔMFCC feature data (total feature data M≥1344), and each extracted feature data is a matrix with 1344 columns. Set the first 12 dimensions of MFCC (excluding the 0th dimension), that is, the MFCC feature data matrix is ​​12 rows (matrix size is 12×1344), then the feature data matrices of the ΔMFCC feature data matrix and the ΔΔMFCC feature data matrix are 11 rows and 10 rows respectively (characteristics of differential calculation) (ΔMFCC matrix size is 11×1344, ΔΔMFCC matrix size is 10×1344);

[0050] Then the first 224 columns of the MFCC feature data matrix are cut and filled into matrix A. At this time, the first 12 rows in matrix A have been filled. The remaining 225-1344 columns of the MFCC feature data matrix are cut in sequence by similar operations (the next time the 225-448 columns are cut) and filled downward from the 13th row in matrix A (after the 13th-24 rows are cut and filled), and so on. Finally, the MFCC feature data matrix is ​​fully filled into the first 72 rows of matrix A.

[0051] A similar operation is performed on the ΔMFCC feature data matrix. The feature data of columns 1-224 (11×224) are trimmed and filled into rows 73-83 of matrix A. Finally, the ΔMFCC feature data matrix is ​​filled into rows 73 to 138 of matrix A, for a total of 66 rows.

[0052] A similar operation is performed on the ΔΔMFCC feature data matrix. The feature data of columns 1-224 (10×224) are trimmed and filled into rows 139-148 of matrix A. Finally, the ΔΔMFCC feature data matrix is ​​filled into rows 139 to 198 of matrix A, for a total of 60 rows.

[0053] The remaining unfilled rows 199 to 224 in matrix A are filled with 0.

[0054] After the padding operation, matrix A contains 1344 frames of MFCC, ΔMFCC, and ΔΔMFCC feature data. This step is performed on the data of all frames, and 1344 frames are trimmed in sequence (frames 1345-2688 are filled into matrix B, and frames 2689-4032 are filled into matrix C, until the total number of feature data M is filled). Finally, all matrices obtain fused feature data.

[0055] Furthermore, the step S5 includes the following steps:

[0056] S501. Dataset division:

[0057] The fused feature data are mapped according to the corresponding transformer working state. Each working state and its corresponding fused feature data constitute a data set. The data set is divided into training set, validation set and test set according to 70%, 20% and 10% respectively. The normal state of the transformer, the transformer discharge fault and the transformer overload state are represented by labels a, b and c respectively.

[0058] S502, model training:

[0059] A ResNet50 model was used as a classifier and trained to obtain a feature classification model. To prevent overfitting, dropout was used as the loss function, and SGD (stochastic gradient descent) was used as the optimizer. Accuracy, precision, recall, and F1 score were introduced as evaluation metrics. Accuracy is the ratio of samples correctly classified by the model to the total number of samples. Precision is the ratio of samples predicted as positive to those that are actually positive. Recall is the ratio of samples correctly predicted as positive to those that are actually positive. The F1 score is the harmonic mean of precision and recall, which can comprehensively reflect the model's predictive ability and classification effect to a certain extent.

[0060] Furthermore, the specific steps of S6 are: obtaining the sound data of the detected transformer of a fixed length in the current period, and performing the same processing as S2, S3 and S4 to obtain the fused feature data of the current period, and using it as the input of the feature classification model. Finally, the feature classification model outputs the prediction result, that is, the predicted current state of the detected transformer.

[0061] The beneficial effects of the present invention are as follows: extracting the MFCC static features of the transformer sound signal and calculating its dynamic features, and using a convolutional neural network based on deep learning, adding the fusion features of dynamic features for training and recognition, and finally realizing the feature extraction and fault recognition of the transformer sound signal. The method proposed by the present invention can effectively identify the normal working state, discharge fault and overload state of the transformer based on the working sound of the transformer; using sound signals for transformer fault diagnosis as a non-contact detection method has the advantages of no need for disassembly and strong real-time performance compared with traditional detection methods; the introduction of dynamic features enables the model to learn the dynamic information of the sound signal, thereby improving the recognition effect of the model, and the recognition effect of fusion of MFCC static features is better than that of recognition using only a single feature; using a convolutional neural network based on deep learning as a classifier solves the problems of the previous method using traditional machine learning algorithms as classifiers, such as the need for continuous optimization and adjustment of the classifier, weak generalization ability, and the need for manual feature extraction; using ResNet with cross-layer connection and residual block, thereby solving the gradient disappearance and gradient explosion problems of using traditional convolutional neural networks. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 This is a flow chart of a transformer acoustic feature extraction and fault identification method of the present invention;

[0063] Figure 2 This is the sound signal characteristic diagram when the transformer is in normal state;

[0064] Figure 3 This is the characteristic diagram of the sound signal when the transformer is overloaded;

[0065] Figure 4 is the confusion matrix of the trained feature classification model. DETAILED DESCRIPTION

[0066] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0067] Refer to 1, a transformer acoustic feature extraction and fault identification method, comprising the following steps:

[0068] S1. Acquisition of sound data of three states: normal operation of transformer, transformer discharge fault and transformer overload;

[0069] Specifically, the structure of the transformer determines that its sound signal is closely related to its operating state. The sound generated by the transformer when working has characteristics such as rich low-frequency components and high amplitude. Therefore, the microphone that collects sound data should have at least good low-frequency response, high sensitivity, high signal-to-noise ratio and wide dynamic range to be competent for collecting the sound generated by the transformer when working.

[0070] According to the above conditions, a suitable microphone is selected to collect the sound data of the transformer in normal operation, discharge fault and overload conditions.

[0071] S2. Data preprocessing: preprocessing the data acquired in S1, including downsampling and pre-emphasis;

[0072] S3. Feature extraction of sound data: The pre-processed data is sequentially framed, windowed, Fourier transformed, power spectral density calculated, converted to Mel-frequency power spectral density, discrete cosine transformed, MFCC extracted, and the first-order and second-order differences of MFCC calculated to obtain ΔMFCC and ΔΔMFCC respectively;

[0073] The MFCC of the data is used as the static feature, the ΔMFCC and ΔΔMFCC of the data are used as the dynamic features, and the static features and dynamic features of the data are collectively referred to as feature data;

[0074] S4, combining the feature data: combining the feature data obtained in S3 to obtain fused feature data;

[0075] S5. Construction of a classification model: Divide the fused feature data into a training set, a validation set, and a test set, use a ResNet50 model as a classifier, and train the ResNet50 model to obtain a feature classification model;

[0076] S6. Deployment of the feature classification model and acoustic feature extraction and fault identification of the transformer in the current period: Acquire sound data of the detected transformer of a fixed duration in the current period, and perform the same preprocessing, feature extraction, and feature data combination on the sound data of the detected transformer as in S2, S3, and S4 to obtain fused feature data for the current period, which is used as input to the feature classification model. Finally, the feature classification model outputs a prediction result, i.e., the predicted current state of the detected transformer.

[0077] The S2 is specifically:

[0078] S201, sound data downsampling: To reduce the amount of data, the amount of computation required for subsequent algorithms, and to increase the speed of the entire method, the sound data collected in the three states in S1 need to be downsampled. During this process, it is important to ensure that the Nyquist sampling theorem is met, i.e., the sampling rate after downsampling must be greater than twice the highest frequency of the signal.

[0079] S202, sound data pre-emphasis: To enhance the energy of high-frequency signals and reduce the impact of low-frequency signals, pre-emphasize the downsampled sound data obtained in S201. In audio signals, the energy of high-frequency signals is usually low, while the energy of low-frequency signals is high. If no pre-emphasis is performed, the information of the high-frequency signals will be obscured.

[0080] The S3 is specifically:

[0081] S301, framing:

[0082] The pre-processed audio signal x(n) of the mth frame obtained by S2 is divided into frames. The number of sampling points of each frame, i.e., the length of each frame is N. A 50% overlap is set between every two frames. Assuming the sampling rate is fs, the time length of each frame is:

[0083]

[0084] S302, windowing:

[0085] The signal of each sampling point n is windowed using a Hamming window to reduce the spectrum leakage effect of the discrete signal. Assuming that the window function of the mth frame is w(m,n), the signal after the weighted window function of the mth frame is:

[0086] xw(m,n)=x(n)×w(m,n),0≤n<N,0≤m<M (2)

[0087] S303, Fourier transform:

[0088] Perform Fourier transform on the windowed signal xw(m,n) and find the frequency domain representation:

[0089] X(m,k)=FFT[xw(m,n)],0≤k<K (11)

[0090] S304, calculate power spectrum density

[0091] Power spectral density is a function that describes how the signal power distribution changes with frequency. It is the square of the Fourier transform and represents the signal power at different frequencies. The power spectral density PSD(m,k) is calculated as:

[0092]

[0093] S305, convert Mel frequency power spectrum density:

[0094] It is usually implemented using a triangular filter bank, where each triangular filter corresponds to a Mel frequency and the shape of the triangular filter is similar to a triangle. The triangular filter filters and downsamples the signal on the Mel frequency scale to extract audio features that are closer to the frequency response perceived by the human ear.

[0095] Define a filter bank consisting of I triangular filters, each with a center frequency of f(i), and its frequency response is:

[0096]

[0097] Map PSD(m,k) to the Mel frequency scale using a triangular filter bank to obtain the Mel frequency power spectral density (MPSD):

[0098]

[0099] S306, Discrete Cosine Transform:

[0100] Perform discrete cosine transform (DCT) on the Mel frequency power spectral density to obtain the MFCC features of the sound data:

[0101]

[0102] Where m represents the number of frames, j represents the dimension of the MFCC coefficients of the current frame, and I represents the number of triangular filters;

[0103] S307, calculate the first-order difference and second-order difference of MFCC:

[0104] Performing first-order and second-order difference calculations on the MFCC (m, j) obtained in step S306. This step is used to enhance the dynamics of the features to better reflect the time-varying nature of the transformer sound signal.

[0105] Assume that ΔMFC(C,m1) and ΔΔMFCC(m,j2) are obtained by first-order difference and second-order difference calculation respectively;

[0106] m represents the number of frames, j1 and j2 represent the dimensions of the ΔMFCC coefficients and ΔΔMFCC coefficients of the current frame, respectively;

[0107] After S306 and S307 , MFCC, ΔMFCC and ΔΔMFCC feature data are obtained.

[0108] The three characteristics of the sound signal calculated when the transformer is in normal state are shown in the attached Figure 2 The three characteristics of the sound signal calculated when the transformer is overloaded are shown in the attached Figure 3.

[0109] The specific steps of S4 are:

[0110] Assume that an empty matrix is ​​A(224, 224), and fill the MFCC, ΔMFCC, and ΔΔMFCC obtained in S3 into A respectively. The specific process is as follows:

[0111] First, take m equal to 1344, that is, 1344 frames of MFCC feature data matrix, ΔMFCC feature data matrix and ΔΔMFCC feature data (total feature data M≥1344), and each extracted feature data is a matrix with 1344 columns. Set the first 12 dimensions of MFCC (excluding the 0th dimension), that is, the MFCC feature data matrix is ​​12 rows (matrix size is 12×1344), then the feature data matrices of the ΔMFCC feature data matrix and the ΔΔMFCC feature data matrix are 11 rows and 10 rows respectively (characteristics of differential calculation) (ΔMFCC matrix size is 11×1344, ΔΔMFCC matrix size is 10×1344);

[0112] Then the first 224 columns of the MFCC feature data matrix are cut and filled into matrix A. At this time, the first 12 rows in matrix A have been filled. The remaining 225-1344 columns of the MFCC feature data matrix are cut in sequence by similar operations (the next time the 225-448 columns are cut) and filled downward from the 13th row in matrix A (after the 13th-24 rows are cut and filled), and so on. Finally, the MFCC feature data matrix is ​​fully filled into the first 72 rows of matrix A.

[0113] A similar operation is performed on the ΔMFCC feature data matrix. The feature data of columns 1-224 (11×224) are trimmed and filled into rows 73-83 of matrix A. Finally, the ΔMFCC feature data matrix is ​​filled into rows 73 to 138 of matrix A, for a total of 66 rows.

[0114] A similar operation is performed on the ΔΔMFCC feature data matrix. The feature data of columns 1-224 (10×224) are trimmed and filled into rows 139-148 of matrix A. Finally, the ΔΔMFCC feature data matrix is ​​filled into rows 139 to 198 of matrix A, for a total of 60 rows.

[0115] The remaining unfilled rows 199 to 224 in matrix A are filled with 0.

[0116] After the padding operation, matrix A contains 1344 frames of MFCC, ΔMFCC, and ΔΔMFCC feature data. This step is performed on the data of all frames, and 1344 frames are trimmed in sequence (frames 1345-2688 are filled into matrix B, and frames 2689-4032 are filled into matrix C, until the total number of feature data M is filled). Finally, all matrices obtain fused feature data.

[0117] The S5 comprises the following steps:

[0118] S501. Dataset division:

[0119] The fused feature data are mapped according to the corresponding transformer working state. Each working state and its corresponding fused feature data constitute a data set. The data set is divided into training set, validation set and test set according to 70%, 20% and 10% respectively. The normal state of the transformer, the transformer discharge fault and the transformer overload state are represented by labels a, b and c respectively.

[0120] S502, model training:

[0121] A ResNet50 model was used as a classifier and trained to obtain a feature classification model. To prevent overfitting, dropout was used as the loss function, and SGD (stochastic gradient descent) was used as the optimizer. Accuracy, precision, recall, and F1 score were introduced as evaluation metrics. Accuracy is the ratio of samples correctly classified by the model to the total number of samples. Precision is the ratio of samples predicted as positive to those that are actually positive. Recall is the ratio of samples correctly predicted as positive to those that are actually positive. The F1 score is the harmonic mean of precision and recall, which can comprehensively reflect the model's predictive ability and classification effect to a certain extent.

[0122] After training, the feature classification model is obtained, and its confusion matrix is ​​attached. Figure 4 .

[0123] The specific steps of S6 are: obtaining the sound data of the transformer under test for a fixed time period in the current period, and performing the same processing as S2, S3 and S4 to obtain the fused feature data of the current period, and using it as the input of the feature classification model. Finally, the feature classification model outputs the prediction result, that is, the predicted current state of the transformer under test.

[0124] The above-described embodiments merely illustrate the implementation methods of the present invention. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that a person skilled in the art would be able to make numerous modifications and improvements without departing from the spirit of the present invention, all of which fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the appended claims.

Claims

1. A transformer acoustic feature extraction and fault identification method, characterized by: The steps include: S1. Acquisition of sound data of three states: normal operation of transformer, transformer discharge fault and transformer overload; S2. Data preprocessing: preprocessing the data acquired in S1, including downsampling and pre-emphasis; S3. Feature extraction of sound data: Frame the pre-processed data, perform windowing, Fourier transform, calculate power spectral density, convert Mel-frequency power spectral density, perform discrete cosine transform, extract MFCC, calculate first-order difference and second-order difference of MFCC to obtain ΔMFCC and ΔΔMFCC respectively; The MFCC of the data is used as the static feature, the ΔMFCC and ΔΔMFCC of the data are used as the dynamic features, and the static features and dynamic features of the data are collectively referred to as feature data; S4, combining the feature data: combining the feature data obtained in S3 to obtain fused feature data; S5. Construction of a classification model: Divide the fused feature data into a training set, a validation set, and a test set, use a ResNet50 model as a classifier, and train the ResNet50 model to obtain a feature classification model; S6, deployment of the feature classification model and acoustic feature extraction and fault identification of the transformer in the current period: obtaining sound data of the detected transformer of a fixed duration in the current period, and performing the same preprocessing, feature extraction, and feature data combination on the sound data of the detected transformer as in S2, S3, and S4, to obtain fused feature data of the current period, and use the fused feature data as input to the feature classification model. Finally, the feature classification model outputs a prediction result, i.e., the predicted current state of the detected transformer; The specific steps of S4 are: Assume an empty matrix A (224, 224), and fill the MFCC, ΔMFCC, and ΔΔMFCC obtained in S3 into A respectively. The specific process is as follows: First take Equal to 1344, that is, 1344 frames of MFCC feature data matrix, ΔMFCC feature data matrix and ΔΔMFCC feature data, the total number of feature data M ≥ 1344, each extracted feature data is a matrix with 1344 columns, MFCC is set to take the first 12 dimensions, excluding the 0th dimension, that is, the MFCC feature data matrix is ​​12 rows, the matrix size is 12 × 1344, then the ΔMFCC feature data matrix and the ΔΔMFCC feature data matrix each feature data matrix has 11 rows and 10 rows, respectively, the characteristics of the differential calculation, the ΔMFCC matrix size is 11 × 1344, the ΔΔMFCC matrix size is 10 × 1344; Then the first 224 columns of the MFCC feature data matrix are cut and filled into matrix A. At this time, the first 12 rows in matrix A have been filled. The remaining 225-1344 columns of the MFCC feature data matrix are cut in sequence by similar operations. The next time the 225-448 columns are cut, they are filled downward from the thirteenth row in matrix A. After the 13-24 rows are cut and filled, and so on, the final MFCC feature data matrix is ​​filled to the first 72 rows of matrix A. A similar operation is performed on the ΔMFCC feature data matrix. The feature data of columns 1-224 (11×224) are trimmed and filled into rows 73-83 of matrix A. Finally, the ΔMFCC feature data matrix is ​​filled into rows 73-138 of matrix A, for a total of 66 rows. A similar operation is performed on the ΔΔMFCC feature data matrix. The feature data of columns 1-224 (10×224) are trimmed and filled into rows 139-148 of matrix A. Finally, the ΔΔMFCC feature data matrix is ​​filled into rows 139 to 198 of matrix A, for a total of 60 rows. The remaining 199 to 224 rows in matrix A that are not filled are filled with 0; After the padding operation, matrix A contains 1344 frames of MFCC, ΔMFCC, and ΔΔMFCC feature data. This step is performed on the data of all frames, and 1344 frames are trimmed in sequence. Frames 1345-2688 are filled into matrix B, and frames 2689-4032 are filled into matrix C until the total number of feature data M is filled. Finally, all matrices obtain fused feature data.

2. A transformer acoustic feature extraction and fault identification method according to claim 1, characterized in that: The S2 is specifically: S201, sound data downsampling: To reduce the amount of data, the amount of computation required for subsequent algorithms, and to increase the speed of the entire method, the sound data collected in the three states in S1 need to be downsampled. During this process, it is important to ensure that the Nyquist sampling theorem is met, i.e., the sampling rate after downsampling must be greater than twice the highest frequency of the signal. S202, sound data pre-emphasis: To enhance the energy of high-frequency signals and reduce the impact of low-frequency signals, pre-emphasize the downsampled sound data obtained in S201. In audio signals, the energy of high-frequency signals is usually low, while the energy of low-frequency signals is high. If no pre-emphasis is performed, the information of the high-frequency signals will be obscured.

3. The transformer acoustic feature extraction and fault identification method according to claim 2, characterized in that: The S3 is specifically: S301, framing: The pre-processed Frame audio signal Frame division is performed, and the number of sampling points of each frame is the length of each frame. , set 50% overlap between every two frames, and set the sampling rate to , then the time length of each frame is: ； S302, adding window: For each sampling point The signal is windowed using a Hamming window to reduce the spectrum leakage effect of the discrete signal; assuming that The window function of the frame is , then The signal after the frame window function weighting is: ； S303, Fourier transform: The windowed signal Perform Fourier transform and find the frequency domain representation: ； S304, calculate power spectrum density Power spectral density is a function that describes how the signal power distribution changes with frequency. It is the square of the Fourier transform and represents the signal power at different frequencies. : ； S305, convert Mel frequency power spectrum density: It is usually implemented using a triangular filter bank, where each triangular filter corresponds to a Mel frequency and the shape of the triangular filter is similar to a triangle. The triangular filter filters and downsamples the signal on the Mel frequency scale to extract audio features that are closer to the frequency response perceived by the human ear. Define a A filter bank consisting of triangular filters, each with a center frequency of , then its frequency response is: ； Will The triangular filter bank is mapped to the Mel frequency scale to obtain the Mel frequency power spectral density MPSD: ； S306, Discrete Cosine Transform: Perform discrete cosine transform (DCT) on the Mel frequency power spectrum density to obtain the MFCC features of the sound data: ； in, Indicates the number of frames, Indicates the dimension of the current frame MFCC coefficient, Indicates the number of triangular filters; S307, calculate the first-order difference and second-order difference of MFCC: The S306 obtained Perform first-order and second-order difference calculations. This step is used to enhance the dynamics of the feature to better reflect the time-varying nature of the transformer sound signal. Assume that after the first-order difference and second-order difference calculation, we can get and ; Indicates the number of frames, and Represents the current frame Coefficient and the dimension of the coefficients; After S306 and S307 , MFCC, ΔMFCC and ΔΔMFCC feature data are obtained.

4. A transformer acoustic feature extraction and fault identification method according to claim 3, characterized in that: The S5 comprises the following steps: S501. Dataset division: The fused feature data are mapped according to the corresponding transformer working state. Each working state and its corresponding fused feature data constitute a data set. The data set is divided into training set, validation set and test set according to 70%, 20% and 10% respectively. The normal state of the transformer, the transformer discharge fault and the transformer overload state are represented by labels a, b and c respectively. S502, model training: The ResNet50 model is used as the classifier, and the ResNet50 model is trained to obtain a feature classification model; to prevent overfitting, dropout is used as the loss function, and the optimizer is SGD stochastic gradient descent; Accuracy, Precision, Recall, and F1 score are introduced as evaluation indicators. Accuracy is the proportion of samples correctly classified by the model to the total number of samples, precision is the proportion of samples that are truly positive among all samples predicted to be positive, recall refers to the proportion of samples that are correctly predicted to be positive among all samples that are actually positive, and F1 value is the harmonic mean of precision and recall, which can comprehensively reflect the prediction ability and classification effect of the model to a certain extent.

5. A transformer acoustic feature extraction and fault identification method according to claim 4, characterized in that: The specific steps of S6 are: obtaining the sound data of the transformer under test for a fixed time period in the current period, and performing the same processing as S2, S3 and S4 to obtain the fused feature data of the current period, and using it as the input of the feature classification model. Finally, the feature classification model outputs the prediction result, that is, the predicted current state of the transformer under test.

Images