Power transformer short-circuit impulse fault acoustic print recognition method based on translation-invariant CNN

Abstract

The application discloses a power transformer short-circuit impact fault soundprint recognition method based on a translation-invariant CNN. Firstly, the audio signal of the transformer short-circuit impact fault collected is preprocessed to obtain time-frequency information features. Then, the input fault audio signal is subjected to feature dimension reduction by using a mel-frequency spectrum to reduce the translation offset of the impact soundprint response time-frequency diagram. Next, the CNN is improved, the local features of the input full connection layer are globally fused, and the translation invariance of the fault sample is enhanced. Finally, the test sample is used for test recognition according to the trained translation-invariant CNN model, and the soundprint recognition result of the test sample is obtained. The actual measurement data result shows that the method improves the short-circuit impact fault recognition rate on the basis of guaranteeing the accurate recognition of the remaining fault categories, effectively verifies the robustness of the translation-invariant CNN to the short-circuit impact fault soundprint recognition, and has high engineering application value.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of transformer fault acoustic print recognition, and particularly relates to a power transformer short-circuit impact fault acoustic print recognition method based on a translation-invariant CNN. BACKGROUND

[0002] As the core of modern power systems, transformers are bound to fail due to long-term high-load operation. Timely detection of transformer short-circuit impact faults can prevent major accidents and reduce economic losses.

[0003] Common methods for diagnosing transformer short-circuit impact faults include oil chromatographic analysis and vibration analysis. For example, transformer turn-to-turn short-circuit faults are analyzed using transformer gas action and gas conditions; an index for quantitatively diagnosing winding vibration faults is proposed based on the correlation between vibration response and excitation signals; and a GBDT and K-means double-layer model is introduced based on oil chromatography. However, distribution transformers are widely distributed and numerous, and it is difficult to obtain vibration signals by sampling oil and gas during inspection, so oil chromatographic analysis and vibration analysis are not suitable for diagnosing faults in distribution transformers. In fact, transformers emit sound when operating normally, and both vibration signals and sound signals change with the operating state, so acoustic print recognition technology can be used to diagnose and identify transformer faults.

[0004] Acoustic print recognition first extracts acoustic features from acoustic signals. There are methods that directly use time-domain signals for VMD decomposition or combine EEMD and blind source separation to extract vibration signal and acoustic print signal features for fault monitoring, but all have the defect of high dimensionality of acoustic print features. Therefore, researchers have introduced speech feature extraction methods such as LPC, LPCC, and MFCC to achieve dimensionality reduction.

[0005] At the same time, deep learning has been introduced into the field of speech recognition to strengthen the recognition of features by speech recognition systems. CNN has certain translation invariance due to its local connection and weight sharing characteristics, making it suitable for fault sound recognition research. However, by extracting translation-invariant features from VGG16, ResNet50, and DenseNet201 network structures, it was found that CNN has a certain fault tolerance for translation invariance. As the translation scale of the features increases, the recognition rate of the model also decreases.

[0006] Therefore, improving the recognition rate of CNN and solving the problem of translation invariance are of great significance for the research and practical application of using acoustic print recognition technology to diagnose transformer faults. SUMMARY

[0007] The problem to be solved by the present application is to provide a power transformer short-circuit impact fault soundprint recognition method based on a translation-invariant CNN, which adopts a mel time-frequency spectrum to perform feature dimension reduction on an input fault sound signal, reduces the translation offset of the impact soundprint response time-frequency graph, and improves a traditional CNN network structure to realize the strengthening of the translation invariance of the CNN.

[0008] The present application adopts the following technical scheme: a power transformer short-circuit impact fault soundprint recognition method based on a translation-invariant CNN, comprising the following steps:

[0009] S1, fault soundprint feature extraction: pre-processing the collected transformer short-circuit impact fault audio signal to obtain time-frequency information features, and converting the fault audio signal into positive half-frequency power spectrum data;

[0010] S2, fault soundprint mel dimension reduction feature extraction: performing dimension reduction on each frame of data of the positive half-frequency power spectrum through a mel filter, and splicing the signal output of each frame after dimension reduction into a dimension-reduced time-frequency spectrum image in time sequence;

[0011] S3, training set making: performing dimension reduction on each frame of each audio signal through the mel filter, dividing the obtained dimension-reduced time-frequency spectrum image into a matrix image as a training sample, and making a training set;

[0012] S4, construction of a translation-invariant CNN model: model training through the training set combined with the translation-invariant CNN model, to obtain a trained translation-invariant CNN model;

[0013] S5, test set making: selecting 8 consecutive frames from the middle position of each audio signal to perform dimension reduction through the mel filter, to obtain a test sample image, and forming a test set;

[0014] S6, recognition output: using the test sample to test and identify according to the trained translation-invariant CNN model, to obtain a soundprint recognition result of the test sample.

[0015] Preferably, in step S1, the fault soundprint feature extraction process comprises framing, windowing, fast Fourier transform, and power spectrum calculation, and the specific process is as follows:

[0016] S1.1, performing original signal extraction on the collected Q pieces of K kinds of transformer short-circuit impact fault audio signals to obtain a signal sequence X of length Lx1: train ={x1,x2,…,x i ,…x Q}, wherein the sampling frequency of each fault audio signal is f s , the time length is tx i , and xi represents the i-th signal sequence after signal extraction;

[0017] S1.2, frame and window operation is performed on the acquired signal sequence, and the signal distortion is reduced using the following Hamming window formula:

[0018]

[0019] wherein n is the index of the discrete sequence, N is the frame length of the signal sequence, w(n) is the value of the Hamming window, and the output sequence x of the windowed signal is: wi

[0020] wi i

[0021] S1.3, the frame and windowed signal is subjected to fast Fourier transform to obtain the sequence signal X i

[0022]

[0023] S1.4, the power spectrum P i xi

[0024]

[0025] Preferably, in step S2, the response formula of the Mel filter is:

[0026]

[0027] wherein 1≤m≤M, M is the number of Mel filters, and the function f(.) represents the maximum value of the frequency response corresponding to the Mel filter;

[0028] The power spectrum P xi of the signal output Y i (m) after dimensionality reduction by the Mel filter is:

[0029]

[0030] After dimensionality reduction by the Mel filter feature, the signal output y i of each frame is stacked in time sequence to obtain a reduced dimension time-frequency spectrum image.

[0031] Preferably, in step S3, a training set is prepared, and for each audio signal x i , the output y i of each frame after dimensionality reduction by the Mel filter is stacked in time sequence to form a reduced dimension time-frequency spectrum image with a dimension of 28×58, and each continuous 8 frames of the reduced dimension time-frequency spectrum image is divided into a group of matrix images to obtain a training sample with an image size of 28×8, and all transformer short-circuit impact fault audio signals X train ​​​​​​​= {x1, x2, …, x i ,…x Q} be made into a training set Y train = {y1, y2, …, y i ,…,y Q}.

[0032] Preferably, in step S4, the translation-invariant CNN model comprises two convolutional layers, one maximum pooling layer, a global fusion layer and one fully connected layer, wherein both of the two convolutional layers are not zero-padded, the kernel size of both of the two convolutional layers is 3x3, the step size of both of the two convolutional layers is 1, both of the two convolutional layers are connected with a batch normalization layer, and the batch normalization layer performs batch normalization on the input signal through the following normalization formula:

[0033]

[0034] wherein x is the input signal, k is the dimension of the signal, is the normalized output signal;

[0035] After the batch normalization layer, a ReLU function is used for activation, and then a pooling layer is used to further reduce the feature dimension, a dropout layer is added after the fully connected layer to reduce overfitting, and a classifier is used to classify the output features.

[0036] The classifier uses softmax to make a probability determination on the fault sound, and is used in combination with a classification cross-entropy loss function.

[0037] Further, in step S4, the data label label = {label1, label2, …, label K} is corresponded to the training set sample Y train of one kind, and then the model is trained, and the method is as follows:

[0038] S4.1, input the training set Y train , and reduce the feature sample image size to 28x8x1, and then pass through two convolutional layers, the number of convolutional kernels of the first convolutional layer is 64, and the number of convolutional kernels of the second convolutional layer is 128, so that the first layer feature output size is 26x6x64, and the second layer feature output size is 24x4x128.

[0039] S4.2, the feature map of the output convolutional layer is passed through a maximum pooling layer, and the layer output feature map size is 12x2x128; then, the feature map output by the maximum pooling layer is subjected to a global fusion operation, and a plurality of local features are fused into one global feature point, from 12x2x128 to 1x128, and after convolution, pooling and global fusion, 128 feature points of each training sample y i are obtained.

[0040] S4.3. Flatten all the feature points after global fusion into a one-dimensional sequence and input it into the fully connected layer; then, the fully connected layer performs classification and output through a softmax classifier. The softmax function formula is:

[0041]

[0042] The Sotfmax classifier converts the layer output scores into probabilities and determines the classification result based on the probability magnitude.

[0043] S4.4 Complete the training of the translation-invariant CNN model, obtain the trained translation-invariant CNN model and save it.

[0044] Preferably, in step S5, a test set Y is created. test From each audio signal x i Eight consecutive frames are selected from the 29th frame in the middle position. Only these eight frames are processed by a Mel filter for dimensionality reduction, resulting in a 28×8 test sample image. The test sample images of all audio signals together constitute the test set Y. test .

[0045] Preferably, in step S6, the identification output will be the test sample Y. test The trained translation-invariant CNN model is input, passed through two convolutional layers, one max-pooling layer, a global fusion layer, and one fully connected layer, and then classified by a trained classifier to obtain the test sample Y. test The recognition results.

[0046] The present invention also provides: an electronic device, comprising:

[0047] One or more processors;

[0048] A storage device on which one or more programs are stored;

[0049] When the one or more programs are executed by the one or more processors, the one or more processors implement any of the above-described methods for voiceprint recognition of short-circuit impact faults in power transformers based on translation-invariant CNNs.

[0050] The present invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps in any of the above-mentioned methods for voiceprint recognition of short-circuit impact faults in power transformers based on translation-invariant CNN.

[0051] Compared with the prior art, the present invention, employing the above technical solution, has the following technical effects:

[0052] The application is based on a power transformer short-circuit impact fault soundprint recognition method of a translation-invariant CNN, takes power transformer short-circuit impact faults as the main research object, considers the problems of random translation and too high feature recognition dimension of the impact signal, uses a mel-frequency spectrum to perform feature dimension reduction on the input fault sound signal to reduce the translation offset of the impact soundprint response time-frequency graph, and improves the traditional CNN network structure to realize the strengthening of the translation invariance of the CNN, thereby improving the short-circuit impact fault recognition rate on the basis of ensuring the accurate recognition of the remaining fault categories, effectively verifying the robustness of the translation-invariant CNN for short-circuit impact fault soundprint recognition, and having high engineering application value. BRIEF DESCRIPTION OF DRAWINGS

[0053] Figure 1 A power transformer short-circuit impact fault soundprint recognition method flowchart of the application;

[0054] Figure 2 A power transformer short-circuit impact fault time-frequency spectrum of the application;

[0055] Figure 3 A power transformer short-circuit impact fault mel dimension reduction time-frequency spectrum of the application;

[0056] Figure 4 A CNN convolution recognition process chart of the application;

[0057] Figure 5 A translation-invariant CNN global fusion process chart of the application;

[0058] Figure 6 A different kind of audio accuracy comparison chart of the application. DETAILED DESCRIPTION

[0059] In order to make the purpose, technical scheme and advantages of the application more clear, the technical scheme of the application will be further described in detail below with reference to the drawings, and the described embodiments are only a part of the embodiments involved in the application. All non-innovative embodiments of other researchers in the art on the embodiments belong to the protection scope of the application. Meanwhile, the step numbers in the embodiments are only set for the convenience of description and explanation, and the order between the steps is not limited in any way, and the execution order of each step in the embodiments can be adaptively adjusted according to the understanding of those skilled in the art.

[0060] The application is based on a power transformer short-circuit impact fault soundprint recognition method of a translation-invariant CNN, takes power transformer short-circuit impact faults as the main research object, considers the problems of random translation and too high feature recognition dimension of the impact signal, uses a mel-frequency spectrum to perform feature dimension reduction on the input fault sound signal to reduce the translation offset of the impact soundprint response time-frequency graph, and improves the traditional CNN network structure to realize the strengthening of the translation invariance of the CNN.

[0061] The overall flow of the voiceprint recognition method of the present application is divided into two parts: fault voiceprint Mel dimension reduction feature extraction and translation invariant CNN model training and recognition, as shown in Figure 1 The steps include the following steps:

[0062] S1, fault voiceprint feature extraction: pre-process the collected audio signal of the transformer short-circuit impact fault, obtain time-frequency information features, and convert the fault audio signal into positive half-frequency power spectrum data;

[0063] S2, fault voiceprint Mel dimension reduction feature extraction: each frame of data of the positive half-frequency power spectrum is dimensionally reduced through a Mel filter, and the signal output of each frame after dimension reduction is spliced into a dimensionally reduced time-frequency spectrum image in time sequence;

[0064] S3, making a training set: each frame of each audio signal is dimensionally reduced through a Mel filter, the obtained dimensionally reduced time-frequency spectrum image is segmented into matrix images as training samples, and a training set is made;

[0065] S4, constructing a translation invariant CNN model: training the model through the training set combined with the translation invariant CNN model, and obtaining the trained translation invariant CNN model;

[0066] S5, making a test set: selecting 8 consecutive frames from the middle position of each audio signal, dimensionally reducing them through a Mel filter, obtaining a test sample image, and forming a test set;

[0067] S6, recognition output: using the test sample to test and recognize according to the trained translation invariant CNN model, and obtaining the voiceprint recognition result of the test sample.

[0068] In an embodiment of the present application, the transformer fault audio is divided into 8 types, of which 3 are normal and 5 are abnormal. Due to the relatively small amount of data, the amount of data is increased by audio framing, taking N=8192 points as a frame, each frame is windowed and then FFT transformed to obtain the power spectrum, and then passed through a Mel filter to obtain each frame output. The time sequence is stacked into a Mel dimensionally reduced time-frequency spectrum image.

[0069] When making a training data set, the 28x58 dimensionally reduced time-frequency spectrum image is segmented into a group of matrix images every 8 consecutive frames; when making a test data set, only 8 consecutive frames of the measured audio are randomly taken.

[0070] In order to verify the effectiveness of the translation invariant CNN, the short-circuit impact fault signal test set contains 9 kinds of continuous translation and position different impact signals.

[0071] The power transformer short-circuit impact fault voiceprint recognition method of the present application is as follows:

[0072] Step S1: fault voiceprint feature extraction

[0073] As a time-varying and non-stationary signal, the transformer short-circuit impulse fault needs to be preprocessed to obtain time-frequency information features, and the fault signal is converted into a two-dimensional energy spectrum suitable for CNN processing. The fault voiceprint feature extraction process is: framing, windowing, fast Fourier transform (FFT), and power spectrum.

[0074] First, a fault audio signal with a sampling frequency of f s = 48KHz and a time length of t = 10s is subjected to original signal extraction, obtaining an i-th signal sequence x i with a length of L x 1.

[0075] Then, the obtained signal sequence is subjected to framing and windowing operation, and N is the frame length. Each audio signal can be divided into 58 frames. In particular, in this embodiment, N is 8192.

[0076] A Hamming window with large sidelobe attenuation is selected to reduce signal distortion. The Hamming window formula is:

[0077]

[0078] In the formula, n is the index of the discrete sequence, and w(n) is the value of the Hamming window.

[0079] The windowed signal output sequence x wi is:

[0080] x wi = x i x w(n)

[0081] Next, the framed and windowed signal is subjected to FFT transform, and the transformed sequence signal X i (k) is:

[0082]

[0083] The power spectrum P i (k) of the positive half frequency of the sequence X xi is:

[0084]

[0085] The fault voiceprint time-frequency graph feature before dimension reduction, as shown in Figure 2 , has a dimension of 4096, and the main frequency components of the transformer operation are basically below 1kHz, and there are basically no frequency components at high frequencies.

[0086] Step S2: fault voiceprint mel dimension reduction feature extraction

[0087] The power spectrum P xiThen, each frame of data is reduced in dimensionality using a Mel filter. The response formula of the Mel filter is:

[0088]

[0089] In the formula, 1≤m≤M, where M is the number of Mel filters. In this embodiment, M = 28.

[0090] Signal power spectrum P xi The signal output Y after dimension reduction by the Mel filter i (m) is:

[0091]

[0092] A Mel-Filter was used to linearly extract the main frequency components below 1kHz from the transformer, and to efficiently extract the frequency components above 1kHz nonlinearly, effectively reducing the input sample size. After dimensionality reduction, each frame's output is stitched together in chronological order to form a dimensionality-reduced time-frequency image. The image size after dimensionality reduction feature extraction for each audio segment is 28×58. The dimensionality-reduced time-frequency image is shown below. Figure 3 As shown, the dimensions are... Figure 2 The number dropped from 4096 to Figure 3 28.

[0093] Step S3: Training Set Creation

[0094] In this embodiment, Q K types of pure transformer fault sound signals X were pre-collected. train ={x1,x2,…,x i ,…x Q After dimensionality reduction using a Mel filter, each 10-second audio signal x i Each frame outputs y i The images are stacked in chronological order to form a dimensionality-reduced time-spectral image with a dimension of 28×58.

[0095] After the input fault sound data is dimensionality-reduced using a Mel filter, the spectral image is divided into a matrix image every 8 consecutive frames. One audio signal can be divided into 7 (rounded to the nearest integer) training samples with an image size of 28×8. All pre-acquired transformer fault sound signals X train Made into training set Y train ={y1,y2,…,y i ,…,y Q}

[0096] Step S4: Based on the translation-invariant CNN model

[0097] CNN is divided into two parts: the filtering stage and the classification stage. The filtering stage mainly consists of convolutional layers, normalization layers, activation layers, pooling layers, etc.

[0098] The input image data is a Mel frequency spectrum after dimension reduction, which is convolved with multiple convolution kernels in the convolution layer and added with a bias to obtain a series of feature maps. A batch normalization layer BN is added after the convolution layer. The BN layer normalizes the input signal, and the normalization formula is:

[0099]

[0100] In the formula, x is the input signal, k is the dimension of the signal, is the normalized output signal.

[0101] In order to optimize the BN layer, two parameters are added for scaling operation and moving operation to improve the nonlinear representation ability of the model. The parameter optimization formula is:

[0102]

[0103] In the formula, p is the scaling parameter, q is the moving parameter, and y is the optimized output signal.

[0104] After the BN layer, a ReLU activation is added and then a pooling layer. Since average pooling causes the features to be not obvious, maximum pooling is selected to react to local features and further reduce the feature dimension to prevent model overfitting.

[0105] The classification level part is mainly composed of a full connection layer and a Softmax layer, etc. The output features are classified.

[0106] The full connection layer flattens the last layer of the filtering level into a one-dimensional feature vector as input, integrates all feature information of the filtering level, extracts global features, and makes a probability determination of the fault sound in combination with Softmax. The Softmax classification function is:

[0107]

[0108] In the formula, y is the predicted probability, x i is the output logic value of the full connection layer. The loss function used in combination with Softmax is the classification cross-entropy loss function, which is:

[0109]

[0110] In the formula, L is the loss function calculation value, j is the number of classification labels, is the predicted value, is the output target value, and m is the batch sample processing size.

[0111] In this embodiment, the improved CNN process is as shown in Figure 4 As the CNN level goes deeper, the extracted features become more and more abstract while still retaining the feature position information.

[0112] In fact, the convolutional layer in the CNN convolution process extracts some feature position information, which is essentially used to help the model identify. For the two-dimensional input features of the transformer fault in the case of small samples, the impact signal of the short-circuit impact fault also has random translation, and when the feature scale is translated too much, the recognition rate will decrease. Therefore, improvement measures need to be taken to reduce the extraction of position information by the model during training.

[0113] Therefore, the embodiment proposes a strategy of global fusion of local features, the principle of which is that before the input two-dimensional mel dimension reduction matrix feature map is output into the full connection layer by the model down-sampling layer, the local features are globally fused to eliminate the influence of relative position movement in multiple feature maps on identification, so that the model learns more abstract global features, such as Figure 5 .

[0114] The local features before entering the full connection layer are globally fused, and the global fusion formula is:

[0115]

[0116] In the formula, j l is the weight size in the current down-sampling layer, x l+1 is the input weight of the next full connection layer.

[0117] The position information of the impact signal is changed from the local feature of the down-sampling layer to a global feature point of a superimposed feature, eliminating the problem of relative translation change of the feature map impact signal. However, in the fusion process, multiple local features are fused into a global feature point, resulting in a decrease in the relative number of features input into the full connection layer. Therefore, the number of filters for extracting features by the convolutional layer is appropriately increased to increase the number of feature maps, and the original input is described and extracted from multiple angles.

[0118] The shift-invariant CNN adds a global fusion operation before the full connection layer on this basis, and the output size of different parts of the CNN and the shift-invariant CNN network structures is compared as shown in Table 1.

[0119] Table 1 Comparison of CNN and shift-invariant CNN layer outputs

[0120]

[0121] To avoid the sharp decrease in the number of feature points affecting the recognition rate of the model, the method of the embodiment appropriately increases the number of channels of the convolutional layer on the basis of adding the global fusion operation, and the shift-invariant CNN network structure after increasing the number of feature points is shown in Table 2.

[0122] Table 2 Shift-invariant CNN network structure

[0123]

[0124] The translation-invariant CNN network structure is composed of two convolutional layers, one maximum pooling layer, a global fusion layer, and one fully connected layer. Among them, the two convolutional layers are not zero-padded, the kernel size is 3*3, the step is 1, and the BN layer is connected after batch normalization, and the ReLU function is activated. The dropout layer is added after the fully connected layer to reduce overfitting, the initial rate is 0.3, the learning rate size is set to 0.001, and the softmax classifier is used with the classification cross-entropy loss function.

[0125] The data label label={label1, label2, …, label K} is corresponded to the training set sample Y train , and the model is trained. First, the training set Y train is input, and the dimension-reduced feature sample image size is 28*8*1. After two convolutional layers, the number of convolutional kernels of the convolutional layer 1 is 64, and the number of convolutional kernels of the convolutional layer 2 is 128. The first layer feature output size is 26*6*64, and the second layer feature output size is 24*4*128. Then, the feature map output by the convolutional layer is processed through a maximum pooling layer, and the layer output feature map size is 12*2*128. Then, the feature map output by the maximum pooling layer is subjected to a global fusion operation, and a plurality of local features are fused into a global feature point, which is fused from 12*2*128 to 1*128. Finally, all the feature points after global fusion are flattened and expanded into a one-dimensional sequence to input the fully connected layer. Then, the classification output is completed by the softmax classifier through the fully connected layer, and the translation-invariant CNN model is trained and saved.

[0126] Step S5: Test set preparation

[0127] In this embodiment, the actual test data is also an audio signal with a sampling frequency of f s = 48KHz and a time length of t = 10s. The test set Y test is prepared in a similar manner as the training set Y train . Both of them are subjected to dimension reduction by the Mel filter to extract the dimension-reduced features.

[0128] However, the training set Y train is divided into 58 frames, each of which is subjected to dimension reduction preprocessing, while the actual test set Y test is selected from the 29th frame to the 36th frame (29-36) in the middle of each audio, and only the 8 frames are subjected to dimension reduction preprocessing to form a 28*8 test sample image.

[0129] Step S6: recognition output

[0130] The effectiveness and robustness of the transformer short-circuit impact fault voiceprint recognition method based on the translation-invariant CNN are verified, and test samples Y test are used for test recognition test are input into the improved pre-CNN, the translation-invariant CNN only with global fusion, and the translation-invariant CNN with the number of feature points increased on the basis of global fusion, and the recognition rates of the three for different types of audio are as shown in Figure 6

[0131] Under the optimal network structure shown in Table 2, the translation-invariant CNN and the CNN are used for fault voiceprint model training, and the test results are compared as shown in Table 3.

[0132] Table 3 Comparison of test results of CNN and translation-invariant CNN

[0133]

[0134]

[0135] In summary, the power transformer short-circuit impact fault is taken as the main research object, the problems of random translation and too high feature recognition dimension of the impact signal are considered, and an efficient fault recognition method based on the dimension-reduced Mel time-frequency spectrum and the translation-invariant CNN is established. On the basis of ensuring accurate recognition of the remaining fault categories, the short-circuit impact fault recognition rate reaches 97.92%, which is improved by 18.67%, effectively verifying the robustness of the translation-invariant CNN for short-circuit impact fault voiceprint recognition, and having high engineering application value.

[0136] In the embodiments of the present application, an electronic device is also provided, which includes one or more processors, a storage device having one or more programs stored thereon, and when the one or more programs are executed by the one or more processors, the one or more processors implement the power transformer short-circuit impact fault voiceprint recognition method based on the translation-invariant CNN described in any of the above embodiments.

[0137] In the embodiments of the present application, a computer readable storage medium having a computer program stored thereon is also provided, and when the program is executed by a processor, the steps in the power transformer short-circuit impact fault voiceprint recognition method based on the translation-invariant CNN in any of the above embodiments are implemented.

[0138] The above description is only the preferred embodiments of the present application, and it should be pointed out that for ordinary skilled persons in the art, without departing from the principles of the present application, a number of improvements and refinements can be made, and these improvements and refinements should also be regarded as the protection scope of the present application.​

Claims

1. A power transformer short-circuit impulse fault acoustic print recognition method based on a shift-invariant CNN, characterized in that, The method comprises the following steps: S1, fault voiceprint feature extraction: pre-processing the collected audio signal of transformer short-circuit impact fault, obtaining time-frequency information feature, and converting the fault audio signal into positive half-frequency power spectrum data; S2, fault voiceprint mel dimension reduction feature extraction: each frame of data of the positive half-frequency power spectrum is dimensionally reduced through a mel filter, and the signal output of each frame after dimension reduction is spliced into a dimensionally reduced time-frequency spectrum image in time sequence; S3, making a training set: each frame of each audio signal is dimensionally reduced through a mel filter, the obtained dimensionally reduced time-frequency spectrum image is segmented into matrix images as training samples, and a training set is made; S4, constructing a translation invariant CNN model: training the model through the training set combined with the translation invariant CNN model, and obtaining the trained translation invariant CNN model; Tagging data with the training set samples After one-to-one correspondence of the categories, model training is performed as follows: S4.1, input the training set The dimension-reduced feature sample image size is 28x8x1, two convolutional layers are passed, the number of convolutional kernels of the first convolutional layer is 64, the number of convolutional kernels of the second convolutional layer is 128, the first layer feature output size is 26x6x64, and the second layer feature output size is 24x4x128. S4.2, the feature map outputted by the convolutional layer is inputted into a max-pooling layer, and the size of the layer output feature map is 12x2x128; then, the feature map outputted by the max-pooling layer is subjected to a global fusion operation, and a plurality of local features are fused into one global feature point, i.e., 12x2x128 is fused into 1x128; after the convolution, the pooling and the global fusion, the 128 feature points of each training sample are obtained . S4.3, all feature points after global fusion are flattened and expanded into a one-dimensional sequence and input into a fully connected layer; then the fully connected layer is classified and output through a softmax classifier, the softmax classifier converts the layer output score into a probability, and the classification result is determined by the probability size; S4.4, the training of the translation invariant CNN model is completed, and the trained translation invariant CNN model is obtained and saved; S5, making a test set: selecting 8 consecutive frames from the middle position of each audio signal, dimensionally reducing the 8 consecutive frames through a mel filter, obtaining a test sample image, and forming a test set; S6, recognition output: using the test sample to test and identify according to the trained translation invariant CNN model, and obtaining the voiceprint recognition result of the test sample.

2. The power transformer short circuit impulse fault acoustic print recognition method according to claim 1, characterized in that, In step S1, the fault voiceprint feature extraction process comprises framing, windowing, fast Fourier transform, and power spectrum calculation, and specifically comprises the following steps: S1.1, original signal extraction is performed on the collected Q pieces of K kinds of transformer short-circuit impact fault audio signals, to obtain a signal sequence with a length of , wherein a sampling frequency of each fault audio signal is , a time length is t, represents the i-th signal sequence after signal extraction;​ In step S2, the response formula of the mel filter is as follows: (1) wherein, is an index of the discrete sequence, is the signal sequence frame length, is the value of the Hamming window, the windowed signal output sequence is: (2) S1.3, performing fast Fourier transform on the framed and windowed signal to obtain a sequence signal : (3) S1.4, to the signal sequence positive half frequency power spectrum: (4) wherein is the positive half-power spectrum of the signal sequence .

3. The power transformer short circuit impulse fault acoustic print recognition method according to claim 2, characterized in that, In step S4, the translation invariant CNN model comprises two convolution layers, one maximum pooling layer, a global fusion layer, and one fully connected layer, wherein the two convolution layers are not zero-padded, the kernel size of the two convolution layers is 3*3, the step size of the two convolution layers is 1, the batch normalization layer is connected after the two convolution layers, the batch normalization layer performs batch normalization on the input signal through the following normalization formula: (5) wherein , is the number of mel-filters, the function denotes the maximum value of the frequency response of the corresponding mel-filter; positive half frequency power spectrum signal output after dimensionality reduction by mel filter is: (6) After the mel filter feature dimension reduction, the signal output of each frame is Stacked in time sequence, the dimension-reduced time-frequency spectrogram image is obtained.

4. The power transformer short circuit impulse fault acoustic print recognition method of claim 3, wherein, In step S3, a training set is made, and for each audio signal After dimension reduction by the Mel filter, the output of each frame is Stacked in time sequence into a dimension-reduced time-frequency spectrum image with a dimension of 28×58, each continuous 8 frames of the dimension-reduced time-frequency spectrum image is divided into a group of matrix images to obtain a training sample with an image size of 28×8, and all transformer short-circuit fault audio signals are made into a training set .

5. The power transformer short circuit impulse fault acoustic print recognition method according to claim 4, characterized in that, ReLU function is used to activate after the batch normalization layer, the feature dimension is further reduced through the pooling layer, the dropout layer is added after the fully connected layer to reduce overfitting, and the output features are classified through the classifier; (7) wherein is the input signal, is the dimension of the signal, is the normalized output signal; The classifier selects softmax to make a probability determination for the fault sound, and is used in combination with a classification cross-entropy loss function. In step S4.3, the softmax classifier is used to classify and output the fully connected layer, and the formula of the softmax function is as follows:

6. The power transformer short circuit impulse fault acoustic print recognition method of claim 5, wherein, One or more processors; (8)。 7. The power transformer short circuit impulse fault acoustic print recognition method according to claim 6, characterized in that, In step S5, the test set is made From each audio signal The 8 consecutive frames are selected from the 29th frame of the middle position, and only the 8 frames of signals are processed by the mel filter for dimension reduction, obtaining a 28x8 test sample image. The test sample images of all audio signals collectively constitute the test set .

8. The power transformer short circuit impulse fault acoustic print recognition method of claim 7, wherein, In step S6, the recognition output, the test sample The trained translation-invariant CNN model is input, and after two convolution layers, one maximum pooling layer, global fusion, and one fully connected layer, the classification output is obtained by the trained classifier to obtain the recognition result of the test sample .

9. An electronic device, comprising: A storage device having one or more programs stored thereon; When the one or more programs are executed by the one or more processors, the one or more processors implement the power transformer short-circuit impact fault voiceprint recognition method based on the translation invariant CNN according to any one of claims 1 to 8. ​ ​ 10. A computer-readable storage medium, characterized in that, A computer program product, comprising a computer readable medium having stored thereon computer program, the computer program comprising instructions executable by a processor to cause the processor to perform the steps of the method of any one of claims 1 to 8 when the computer program is run by the processor.

Images