Transformer fault diagnosis method based on VoaRT algorithm

Abstract

The invention discloses a transformer fault diagnosis method based on a VoaRT algorithm, and relates to the field of transformer fault diagnosis methods, and the method comprises the steps: collecting the partial discharge information in the operation process of a transformer; constructing a partial discharge information phase analysis spectrogram PDPR; feature dimension reduction is carried out under the condition that key information is not lost; cutting the spectrogram into a plurality of feature vector blocks, inputting the feature vector blocks into a convolutional attention network (CBAM), and obtaining weighted features; taking the feature vector as an input vector, carrying out learning training by using a CART decision tree after random sampling, and carrying out classification in combination with a strategy; and inputting a discharge information phase analysis spectrogram PDPR to obtain a diagnosis result. By constructing the convolutional attention decision tree network, the transformer diagnosis accuracy and speed are improved, diagnosis errors caused by sensor faults are reduced, and the stability of the system is enhanced.

Description

Technical Field

[0001] The present invention belongs to the technical field of transformer fault diagnosis methods, and specifically relates to a transformer fault diagnosis method based on a VoaRT algorithm. Background Art

[0002] Since the Second Industrial Revolution, the power system has become the lifeline of human society. As an important electrical equipment in the power system, the reliability and safety of the transformer are crucial to the stable operation of the power system. Since the transformer is subjected to multiple forces such as electromagnetic force, thermal force, and mechanical force during operation, various faults and defects may occur during use, which may cause serious hidden dangers in social life and production. Establishing an accurate and effective transformer fault diagnosis model can timely discover abnormal information inside the transformer equipment, accurately grasp the current operating status of the transformer and predict future development trends, and provide a reliable basis for maintenance personnel to formulate economical and efficient maintenance strategies. It is of great significance to the long-term stable operation of the entire power system and transformers.

[0003] However, the current transformer fault diagnosis methods have certain limitations, such as inaccurate fault diagnosis, strict conditions required for on-site testing, and incomplete fault information. Summary of the invention

[0004] In view of the deficiencies in the prior art, the present invention provides a transformer fault diagnosis method based on the VoaRT algorithm, which has the advantages of being able to timely discover abnormal information inside the transformer equipment, accurately grasp the current operating status of the transformer and predict future development trends.

[0005] To achieve the above object, the present invention provides the following technical solutions: a transformer fault diagnosis method based on the VoaRT algorithm, comprising: step 1: collecting partial discharge information during the operation of the transformer; step 2: constructing a partial discharge information phase analysis spectrum PDPR; step 3: performing feature dimension reduction without losing key information; step 4: cutting the spectrum into multiple feature vector blocks and inputting the convolutional attention network CBAM to obtain weighted features; step 5: taking the feature vector as the input vector, randomly sampling it and using the CART decision tree for learning and training, and classifying it by combining strategies; step 6: inputting the discharge information phase analysis spectrum PDPR to obtain a diagnosis result;

[0006] Preferably, in order to improve the quality of the spectrum signal and improve the signal characteristics, the spectrum signal needs to be preprocessed in advance. The spectrum phase shift, spectrum clipping, spectrum grayscale, and spectrum compression processing are performed in sequence to achieve spectrum preprocessing.

[0007] Perform spectrum phase shift. Partial discharge has strong randomness within the 360° phase cycle, and the number of discharges and pulse sizes are different. The PRPD spectra of some types with fewer numbers are phase shifted by 120°, 240° and 360° respectively to expand the type samples to achieve a balanced number; perform spectrum trimming, retain the areas closely related to the partial discharge information, and discard the annotation information such as discharge type, amplitude, peak value, etc.

[0008] Grayscale the spectrum. Use the grayscale formula to grayscale the spectrum according to the average value, operate pixel by pixel, and select the average value of the three components of R, G, and B as the grayscale value of the pixel each time, so as to achieve the purpose of concentrating information and strengthening features; finally, compress the spectrum, continue to process the spectrum using the same proportion of compression, reduce data, and reduce network calculations. The grayscale formula is:

[0009]

[0010] Preferably, during feature extraction, a feature extractor composed of convolution blocks and residual blocks is used to extract the features of various partial discharge spectra, and the hybrid attention mechanism CBAM optimization is used to obtain the best features; the spectrum data set is passed through a set of basic blocks to obtain data of 55×55×256, and then after passing through three sets of basic blocks, a partial discharge feature data set of 7×7×2048 is output, and the self-attention mechanism is combined to select the optimal partial discharge feature. The extracted partial discharge feature data set is transferred to the average pooling layer for pooling processing, and the output is 1×1×2048. The flattening layer is used to convert the multidimensional features into one dimension, and the output 2048 is transferred to the fully connected layer, and the one-dimensional optimal feature data is output to the random forest classifier.

[0011] A residual block is added to the feature extraction, and a skip structure is used to transmit the input signal directly from any shallow layer to the deep layer to form an identity mapping. The input sample x in the traditional network is directly learned to the output result H(x), where the difference between H(x) and x is the residual, which solves the network degradation problem to a certain extent.

[0012] The calculation formula of the residual block is:

[0013] F(X)=H(X)

[0014] F(X)=H(X)-X

[0015] Preferably, the hybrid attention mechanism includes:

[0016] Channel attention module CAM and spatial attention module SAM.

[0017] In CAM, the input spectrogram passes through two parallel MaxPool layers and AvgPool layers, and the feature map is changed from C*H*W to C*1*1. After passing through the Share MLP module, the number of channels is first compressed to 1 / r times the original number of channels, and then expanded to the original number of channels. The ReLU activation function is used to obtain two activated results, and the two output results are added element by element. The sigmoid activation function is used to obtain the output result of the channel attention, and then the output result is multiplied by the original image to return to the size of C*H*W.

[0018] The formula for CAM is:

[0019]

[0020] In SAM, the output results of channel attention are subjected to maximum pooling and average pooling to obtain two 1*H*W feature maps, which are then concatenated through a Concat operation and converted into a 1-channel feature map through a 7*7 convolution. The sigmoid function is then used to activate the feature map of spatial attention, and finally the output result is multiplied by the original image to return to the size of C*H*W.

[0021] The formula for SAM is:

[0022]

[0023] By adding additional regularization weight coefficients to calculate the weight of CBAM, the contribution of elements on the same channel is different. The calculation formula is:

[0024]

[0025] in is the correction weight coefficient of each spatial position, and the calculation formula is:

[0026]

[0027] Preferably, a decision tree network boosting is used as the classifier:

[0028] The random forest classifier randomly samples the feature samples (with replacement) to form a new sampling set, and establishes a CART decision tree network for learning and training, which is repeated β times. The CART decision algorithm selects features by comparing the Gini coefficient with the threshold. The Gini coefficient Gini (D) represents the purity of the set D. The smaller the value, the higher the sample purity. For the sample set D, there are w categories, and the number of the jth category is C. The expression of the Gini coefficient of sample D is:

[0029]

[0030] Then, pattern recognition is achieved by combining the conclusions of each CART decision tree with strategy analysis.

[0031] Preferably, since the CART decision tree network cannot directly obtain feature information from the spectrogram, it is necessary to perform feature cutting operation on the spectrogram. First, the compressed spectrogram (θa, θb) is cut into N ordered feature blocks. t=1,2,…,N, where Pc is the length and width of the feature block, and the step length between adjacent feature blocks is S. The expression of the number of feature blocks N is as follows:

[0032]

[0033] Feature Block After passing through the linear mapping layer, it is stretched and linearly mapped into a D-dimensional block embedding vector (PatchEmbeddings) t=1,2,...,N, through stretching mapping, the complex two-dimensional convolution is simplified to a one-dimensional linear mapping calculation, which can greatly reduce the difficulty of calculation and increase the calculation speed. Let the block embedding vector set be D P , in D P Add word vector before As the initialization vector, the new block embedding vector set is recorded as D′ P ∈R (N+1)*D .

[0034] Preferably, since the collected feature values ​​are scattered, in order not to affect the accuracy of diagnosis, the collected data needs to be Z-score standardized to eliminate diagnostic errors caused by large differences in data magnitude. Therefore, each type of feature is standardized.

[0035] Preferably, the standardization processing formula is:

[0036]

[0037] In the formula is the joint feature after normalization; μ j is the characteristic mean; σ j is the characteristic standard deviation x i,j are the original features. After the standardized transformation, the average of each type of fault feature is 0, the standard deviation is 1, the feature concentration degree and discrete trend are consistent, and the purpose of eliminating the difference between features is achieved.

[0038] Compared with the prior art, the present invention has the following beneficial effects:

[0039] This application improves the transformer fault diagnosis accuracy and speed by constructing a transformer fault diagnosis method based on the VoaRT algorithm, reduces the diagnostic error caused by sensor failure, enhances the stability of the system, reduces the possibility of misjudgment, and significantly improves the accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 It is a schematic diagram of the process of the present invention;

[0041] Figure 2 The overall process diagram of the present invention is

[0042] Figure 3 This is the improved hybrid attention mechanism CABM model diagram of the present invention. DETAILED DESCRIPTION

[0043] Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in the field without making any creative work shall fall within the scope of protection of the present invention.

[0044] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0045] The present invention provides a technical solution: a transformer fault diagnosis method based on the VoaRT algorithm, comprising the following steps:

[0046] Step 1: Collect partial discharge information during transformer operation;

[0047] Step 2: Construct the partial discharge information phase analysis spectrum PDPR;

[0048] Step 3: Introduce the residual block to reduce the feature dimension without losing key information;

[0049] Step 4: Cut the spectrogram into multiple feature vector blocks and input them into the convolutional attention network CBAM to obtain weighted features;

[0050] Step 5: Take the feature vector as the input vector, randomly sample it, and use the CART decision tree for learning and training, and classify it by combining the strategy;

[0051] Step 6: Input the discharge information phase analysis spectrum PDPR to obtain the diagnosis result;

[0052] In order to improve the quality of the spectrum signal and improve the signal characteristics, the spectrum signal needs to be preprocessed in advance. The spectrum phase shift, spectrum clipping, spectrum grayscale, and spectrum compression processing are performed in sequence to achieve spectrum preprocessing.

[0053] Perform spectrum phase shift. Partial discharge has strong randomness within the 360° phase cycle, and the number of discharges and pulse sizes are different. The PRPD spectra of some types with fewer numbers are phase shifted by 120°, 240° and 360° respectively to expand the type samples to achieve a balanced number; perform spectrum trimming, retain the areas closely related to the partial discharge information, and discard the annotation information such as discharge type, amplitude, peak value, etc.

[0054] Grayscale the spectrum. Use the grayscale formula to grayscale the spectrum according to the average value, operate pixel by pixel, and select the average value of the three components of R, G, and B as the grayscale value of the pixel each time, so as to achieve the purpose of concentrating information and strengthening features; finally, compress the spectrum, continue to process the spectrum using the same proportion of compression, reduce data, and reduce network calculations. The grayscale formula is:

[0055]

[0056] During feature extraction, a feature extractor composed of convolution blocks and residual blocks is used to extract the features of various partial discharge spectra, and the hybrid attention mechanism CBAM optimization is used to obtain the best features; the spectrum data set is 55×55×256 after passing through a set of basic blocks, and then after passing through three sets of basic blocks, a 7×7×2048 partial discharge feature data set is output, and the self-attention mechanism is combined to select the optimal partial discharge feature. The extracted partial discharge feature data set is transferred to the average pooling layer for pooling processing, and the output is 1×1×2048. The flattening layer is used to convert the multi-dimensional features into one dimension, and the output 2048 is transferred to the fully connected layer, and the one-dimensional optimal feature data is output to the random forest classifier.

[0057] A residual block is added to the feature extraction, and a skip structure is used to transmit the input signal directly from any shallow layer to the deep layer to form an identity mapping. The input sample x in the traditional network is directly learned to the output result H(x), where the difference between H(x) and x is the residual, which solves the network degradation problem to a certain extent.

[0058] The calculation formula of the residual block is:

[0059] F(X)=H(X)

[0060] F(X)=H(X)-X

[0061] The hybrid attention mechanism includes the channel attention module CAM and the spatial attention module SAM.

[0062] In CAM, the input spectrogram passes through two parallel MaxPool layers and AvgPool layers, and the feature map is changed from C*H*W to C*1*1. After passing through the Share MLP module, the number of channels is first compressed to 1 / r times the original number of channels, and then expanded to the original number of channels. The ReLU activation function is used to obtain two activated results, and the two output results are added element by element. The sigmoid activation function is used to obtain the output result of the channel attention, and then the output result is multiplied by the original image to return to the size of C*H*W.

[0063] The formula for CAM is:

[0064]

[0065] In SAM, the output results of channel attention are subjected to maximum pooling and average pooling to obtain two 1*H*W feature maps, which are then concatenated through a Concat operation and converted into a 1-channel feature map through a 7*7 convolution. The sigmoid function is then used to activate the feature map of spatial attention, and finally the output result is multiplied by the original image to return to the size of C*H*W.

[0066] The formula for SAM is:

[0067]

[0068] By adding additional regularization weight coefficients to calculate the weight of CBAM, the contribution of elements on the same channel is different. The calculation formula is:

[0069]

[0070] in is the correction weight coefficient of each spatial position, and the calculation formula is:

[0071]

[0072] Use random forest as a classifier, randomly sample feature samples (with replacement) to form a new sampling set, and establish a CART decision tree network for learning and training, and repeat this cycle β times. The CART decision algorithm selects features by comparing the Gini coefficient with the threshold. The Gini coefficient Gini (D) represents the purity of set D. The smaller the value, the higher the sample purity. For sample set D, there are w categories, and the number of the jth category is C. The expression of the Gini coefficient of sample D is:

[0073]

[0074] Then, pattern recognition is achieved by combining the conclusions of each CART decision tree with strategy analysis.

[0075] Since the CART decision tree network cannot directly obtain feature information from the spectrogram, it is necessary to perform feature cutting on the spectrogram. First, the compressed spectrogram (θa, θb) is cut into N ordered feature blocks. t=1,2,…,N, where Pc is the length and width of the feature block, and the step length between adjacent feature blocks is S. The expression of the number of feature blocks N is as follows:

[0076]

[0077] Feature Block After passing through the linear mapping layer, it is stretched and linearly mapped into a D-dimensional block embedding vector (PatchEmbeddings) t=1,2,...,N, through stretching mapping, the complex two-dimensional convolution is simplified to a one-dimensional linear mapping calculation, which can greatly reduce the difficulty of calculation and increase the calculation speed. Let the block embedding vector set be D P , in D P Add word vector before As the initialization vector, the new block embedding vector set is recorded as D′ P ∈R (N+1)*D .

[0078] In order to further extract the position information in the MGK spectrum, the feature blocks need to be sorted according to the sequence order generated by the feature blocks. Position Embeddings is performed. Suppose the position encoding vector of the feature block with the block sequence order t is P t ∈R D ,t=0,1...,N, the expression is as follows:

[0079]

[0080] Where i is the i-th dimension of the position encoding vector, i = 1, 2, ..., D, and the set of position vectors is P ∈ R (N+1)*D :

[0081] P=[p0,p1,...,p N ]

[0082] The position vector encoding vector P and the block embedding vector D p Merge to generate a new feature vector Z0∈R (N+1)*D , as the input vector for the subsequent steps.

[0083] Since the collected feature values ​​are scattered, in order not to affect the accuracy of diagnosis, the collected data needs to be Z-score standardized to eliminate the diagnostic errors caused by the large difference in the order of magnitude of the data. Therefore, each type of feature is standardized.

[0084] The standardization formula is:

[0085]

[0086] In the formula is the joint feature after normalization; μ j is the characteristic mean; σ j is the characteristic standard deviation x i,j are the original features. After the standardized transformation, the average of each type of fault feature is 0, the standard deviation is 1, the feature concentration degree and discrete trend are consistent, and the purpose of eliminating the difference between features is achieved.

[0087] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or includes elements inherent to such process, method, article or device.

[0088] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A transformer fault diagnosis method based on VoaRT algorithm, characterized in that: The following steps are involved: Step 1: Collect partial discharge information during transformer operation; Step 2: Construct the partial discharge information phase analysis spectrum PDPR; Step 3: Introduce the residual block to reduce the feature dimension without losing key information; Step 4: Cut the spectrogram into multiple feature vector blocks and input them into the convolutional attention network CBAM to obtain weighted features; Step 5: Take the feature vector as the input vector, randomly sample it, and use the CART decision tree for learning and training, and classify it by combining the strategy; Step 6: Input the discharge information phase analysis spectrum PDPR to obtain the diagnosis result.

2. The transformer fault diagnosis method based on the VoaRT algorithm according to claim 1 is characterized in that: In order to improve the quality of the spectrum signal and improve the signal characteristics, the spectrum signal needs to be preprocessed in advance. The spectrum phase shift, spectrum clipping, spectrum grayscale, and spectrum compression processing are performed in sequence to achieve spectrum preprocessing. Perform spectrum phase shift. Partial discharge has strong randomness within the 360° phase cycle, and the number of discharges and pulse sizes are different. The PRPD spectra of some types with fewer numbers are phase shifted by 120°, 240° and 360° respectively to expand the type samples to achieve a balanced number; perform spectrum trimming, retain the areas closely related to the partial discharge information, and discard the annotation information such as discharge type, amplitude, peak value, etc. Grayscale the spectrum. Use the grayscale formula to grayscale the spectrum according to the average value, operate pixel by pixel, and select the average value of the three components of R, G, and B as the grayscale value of the pixel each time, so as to achieve the purpose of concentrating information and strengthening features; finally, compress the spectrum, continue to process the spectrum using the same proportion of compression, reduce data, and reduce network calculations. The grayscale formula is:

3. The transformer fault diagnosis method based on VoaRT algorithm according to claim 1 is characterized in that: Self-Attention Feature Extraction: The feature extractor composed of convolutional blocks and residual blocks extracts the features of various partial discharge spectra, and the hybrid attention mechanism CBAM optimization is used to obtain the best features; the spectrum data set passes through a set of basic blocks to obtain data of 55×55×256, and then passes through three sets of basic blocks to output a 7×7×2048 partial discharge feature data set, and combines the self-attention mechanism to select the optimal partial discharge feature. The extracted partial discharge feature data set is transferred to the average pooling layer for pooling processing, and the output is 1×1×2048. The flattening layer is used to convert the multi-dimensional features into one dimension, and the output 2048 is transferred to the fully connected layer, and the one-dimensional optimal feature data is output to the random forest classifier. A residual block is added to the feature extraction, and a skip structure is used to transmit the input signal directly from any shallow layer to the deep layer to form an identity mapping. The input sample x in the traditional network is directly learned to the output result H(x), where the difference between H(x) and x is the residual, which solves the network degradation problem to a certain extent. The calculation formula of the residual block is: F(X)=H(X) F(X)=H(X)-X 4. The transformer fault diagnosis method based on VoaRT algorithm according to claim 3 is characterized in that: The hybrid attention mechanism includes: Channel attention module CAM and spatial attention module SAM. In CAM, the input spectrogram passes through two parallel MaxPool layers and AvgPool layers, and the feature map is changed from C*H*W to C*1*1. After passing through the Share MLP module, the number of channels is first compressed to 1 / r times the original number of channels, and then expanded to the original number of channels. The ReLU activation function is used to obtain two activated results, and the two output results are added element by element. The sigmoid activation function is used to obtain the output result of the channel attention, and then the output result is multiplied by the original image to return to the size of C*H*W. The formula for CAM is: In SAM, the output results of channel attention are subjected to maximum pooling and average pooling to obtain two 1*H*W feature maps, which are then concatenated through a Concat operation and converted into a 1-channel feature map through a 7*7 convolution. The sigmoid function is then used to activate the feature map of spatial attention, and finally the output result is multiplied by the original image to return to the size of C*H*W. The formula for SAM is: By adding additional regularization weight coefficients to calculate the weight of CBAM, the contribution of elements on the same channel is different. The calculation formula is: in is the correction weight coefficient of each spatial position, and the calculation formula is:

5. The transformer fault diagnosis method based on VoaRT algorithm according to claim 1 is characterized in that: Use decision tree network boosting as a classifier: The random forest classifier randomly samples the feature samples (with replacement) to form a new sampling set, and establishes a CART decision tree network for learning and training, which is repeated β times. The CART decision algorithm selects features by comparing the Gini coefficient with the threshold. The Gini coefficient Gini (D) represents the purity of the set D. The smaller the value, the higher the sample purity. For the sample set D, there are w categories, and the number of the jth category is C. The expression of the Gini coefficient of sample D is: Then, pattern recognition is achieved by combining the conclusions of each CART decision tree with strategy analysis.

6. The transformer fault diagnosis method based on VoaRT algorithm according to claim 1 is characterized in that: Since the CART decision tree network cannot directly obtain feature information from the spectrogram, it is necessary to perform feature cutting on the spectrogram. First, the compressed spectrogram (θa, θb) is cut into N ordered feature blocks. Where Pc is the length and width of the feature block, and the step length between adjacent feature blocks is S. The expression of the number of feature blocks N is as follows: Feature Block After passing through the linear mapping layer, it is stretched and linearly mapped into a D-dimensional block embedding vector (PatchEmbeddings) By stretching the mapping, the complex two-dimensional convolution is simplified to a one-dimensional linear mapping calculation, which can greatly reduce the difficulty of calculation and increase the speed of calculation. Let the block embedding vector set be D P , in D P Add word vector before As the initialization vector, the new block embedding vector set is recorded as D′ P ∈R (N+1)*D .

7. The transformer fault diagnosis method based on VoaRT algorithm according to claim 1 is characterized in that: Since the collected feature values ​​are scattered, in order not to affect the accuracy of diagnosis, the collected data needs to be Z-score standardized to eliminate the diagnostic errors caused by the large difference in the order of magnitude of the data. Therefore, each type of feature is standardized.

8. The transformer fault diagnosis method based on VoaRT algorithm according to claim 7 is characterized in that: The standardization formula is: In the formula is the joint feature after normalization; μ j is the characteristic mean; σ j is the characteristic standard deviation x i,j The original features are obtained, achieving the purpose of eliminating the differences between features.

Images