Transformer Partial Discharge Data Enhancement and Identification Method

Abstract

The present invention discloses a method for enhancing and identifying partial discharge data of a transformer of the present invention. Taking the PRPD pattern of the partial discharge fault type as the object, after adding labels to the real samples, they are input into the CGAN model. The trained generator model has a stable training process and can generate partial discharge pattern samples of specific types. The partial discharge patterns of various specific discharge types generated by using the improved CGAN model can expand the original discharge data set, increase the diversity of the original data set, so as to better solve the problems of unbalanced and small-sample distribution of the original discharge data set, etc.; to improve the classification effect, the present invention uses an improved CNN as a classifier, inputs the training sample data with expanded and balanced data, trains the improved CNN, and continuously trains and adjusts the network parameters and weights to output the fault classification diagnosis result; solves the problem of low accuracy of partial discharge pattern recognition of transformers under unbalanced small-sample data, and improves the fault recognition accuracy.

Description

Technical Field

[0001] The present invention belongs to the technical field of power equipment fault diagnosis, and particularly relates to a method for transformer partial discharge data enhancement and recognition. Background Art

[0002] Power transformer equipment is a key auxiliary equipment in the power system. Its stable operation is related to the efficient, safe and stable operation of the power system construction. Its partial discharge characteristics can most effectively reflect the degree of deterioration of the insulation state of the local winding of the transformer. With the further rapid development of the future artificial intelligence industry, the theory and technology of pattern recognition have also made great progress. However, since the partial discharge faults occurring in transformers are generally small probability events, with the significant characteristic that the number of normal state samples is much larger than the number of samples with faults that have occurred, it results that in the comparison of on-site measured sample data, the normal fault samples are always much more than the abnormal fault samples, and the actual fault sample ratios between different fault types are always extremely unbalanced. The original data cannot fully train the existing classifiers, making the recognition accuracy of the classifiers not high and there is a problem of misidentifying different fault types. Summary of the Invention

[0003] The purpose of the present invention is to provide a method for transformer partial discharge data enhancement and recognition, which solves the problem of low accuracy of transformer partial discharge pattern recognition under unbalanced small sample data and can improve the fault recognition accuracy.

[0004] The technical solution adopted by the present invention is that the method for transformer partial discharge data enhancement and recognition is specifically implemented according to the following steps:

[0005] Step 1: Extract the PRPD pattern of the partial discharge fault type in the dataset. After graying the pattern and cropping it into a unified size, it is used as the original sample. After adding labels to the original sample;

[0006] Step 2: Construct an improved CGAN model, input the sample into the improved CGAN model to generate local discharge patterns under various specific discharge types, and obtain an expanded discharge dataset;

[0007] Step 3: Use the improved convolutional neural network system as a classifier, input the expanded discharge dataset to train the classifier, and continuously train and adjust the network parameters and weights to obtain an optimized classifier;

[0008] Step 4: Input the partial discharge pattern of the transformer to be recognized into the optimized classifier and output the fault classification diagnosis result.

[0009] The characteristics of the present invention also lie in:

[0010] The process of graying the pattern in Step 1 is:

[0011] f(x, y) = (R(x, y) + G(x, y) + B(x, y)) / 3 (1)

[0012] Wherein, R(x, y), G(x, y), and B(x, y) are the gray values at the (x, y) position on the R, G, and B three channels of the original map respectively, and f(x, y) is the gray value at the (x, y) position after graying the map by the average value method.

[0013] The improved CGAN model in step 2 includes a generator G and a discriminator D. A spatial pooling layer is introduced into the network model of the generator G, then there is:

[0014] The loss function L of the generator G Is expressed as:

[0015]

[0016] The loss function L of the discriminator D Is expressed as:

[0017]

[0018] Wherein, E represents the expectation of the distribution;

[0019] The objective of the conditional generative adversarial network is expressed as:

[0020]

[0021] The specific process of inputting the samples into the improved CGAN model in step 2 to generate partial discharge maps under various specific discharge types is as follows:

[0022] Input all types of labels and random Gaussian noise into the generator G together. Generate fault samples through the generator and add labels to the fault samples;

[0023] Input the fault samples and their corresponding labels and the original samples and their corresponding labels into the discriminator for analysis or determination respectively. Through the confrontation between the generator model and the discriminator model, adjust and optimize the model parameters of the system to obtain the optimal generator model, discriminator model, and the expanded discharge data set.

[0024] The improved convolutional neural network system in step 3 includes an input layer, two convolutional layers, two pooling layers, two fully connected layer systems, a T-ReLU activation function layer, and an output layer. The number of convolutional kernels in the convolutional layer is 2×2. The first convolutional layer consists of 16 3×3 convolutional kernels, and the second convolutional layer consists of 32 3×3 convolutional kernels.

[0025] The function expression adopted by the T-ReLU activation function layer is:[[]]

[0026]

[0027] Among them, α is an adjustable parameter.

[0028] The beneficial effects of the present invention are as follows:

[0029] The method for enhancing and identifying partial discharge data of the transformer of the present invention enhances the samples by adding random Gaussian noise, introduces a spatial pooling layer on the basis of the original CGAN model, and increases the multi-scale feature learning ability in the network model to ensure the rapid generation of more high-quality image detail information; combined with the improved convolutional neural network T-ReLU activation function, the classification accuracy of the network model function can be improved. Brief Description of the Drawings

[0030] Figure 1 is the partial discharge data enhancement and identification framework based on CGAN-CNN in the present invention;

[0031] Figure 2 is the schematic diagram of the basic structure of the GAN model in the present invention;

[0032] Figure 3 is the schematic diagram of the basic structure of the CGAN in the present invention;

[0033] Figure 4 is the schematic diagram of the partial discharge data enhancement process based on the improved CGAN in the present invention;

[0034] Figure 5 is the schematic diagram of the partial discharge type identification based on CNN in the present invention;

[0035] Figure 6 is the grayscale image of tip discharge in the embodiment;

[0036] Figure 7 is the grayscale image of surface discharge in the embodiment;

[0037] Figure 8 is the grayscale image of air gap discharge in the embodiment;

[0038] Figure 9 is the grayscale image of floating discharge in the embodiment. Detailed Embodiments

[0039] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0040] To solve the problem of low accuracy in the recognition of transformer partial discharge patterns under unbalanced small-sample data, a method for partial discharge data enhancement and discharge type recognition based on an improved conditional generative adversarial network (CGAN) and convolutional neural network (CNN) is proposed. First, the method of the present invention takes the PRPD pattern of the partial discharge fault type as the object, adds labels to the real samples and inputs them into the CGAN model. The trained generator model has a stable training process and can generate partial discharge pattern samples of specific types. Secondly, the partial discharge patterns of various specific discharge types generated by the CGAN can expand the original discharge data set, increase the diversity of the original data set, and better solve the problems of unbalance and small-sample distribution of the original discharge data set, etc.; finally, to improve the classification effect, the present invention uses CNN as the classifier, inputs the training sample data with expanded and balanced data to train the CNN, and continuously trains and adjusts the network parameters and weights to output the fault classification diagnosis result.

[0041] The method for enhancing and recognizing transformer partial discharge data of the present invention uses the model framework as Figure 1 shown, and is specifically implemented according to the following steps:

[0042] Step 1: Extract the PRPD pattern of the partial discharge fault type from the data set. After graying the pattern, crop it into a unified size as the original sample, and add labels to the original sample;

[0043] In order to eliminate the irrelevant information in the original pattern that may interfere with the pattern features, the present invention uses the average value method to gray the pattern, and the formula is shown in Equation (1):

[0044] f(x,y)=(R(x,y)+G(x,y)+B(x,y)) / 3 (1)

[0045] In the formula, R(x,y), G(x,y), and B(x,y) are the gray values at the (x,y) position on the R, G, and B channels of the original pattern respectively, and f(x,y) is the gray value at the (x,y) position after graying the pattern using the average value method.

[0046] Since the patterns may be of different sizes during the collection process, resulting in inability to be trained in the same model, the present invention uniformly transforms the image size to 300×150.

[0047] Step 2: Construct an improved CGAN model, input the samples into the improved CGAN model to generate partial discharge patterns of various specific discharge types, and obtain an expanded discharge data set;

[0048] A generative adversarial network (GAN) generally consists of the following two networks. One is the generator, whose main purpose is to capture the distribution of adversarial training data, and the other is the discriminator, which mainly provides a probability for estimating that the sample data of adversarial training comes from the data of the adversarial training generation network rather than the data generated by the model generator. The goal of the adversarial training generation process of the generator is to maximize the probability of the discriminator making mistakes. This training framework corresponds to an equivalent two-player game with a minimax probability. The generator attempts to generate an image close to the real one. After the image is transmitted to the discriminator, the task of the discriminator is to evaluate the authenticity of the image by comparing it with the real data. The network parameters between the generator and the discriminator need to be continuously updated through the game to automatically optimize their own generation and classification capabilities. The basic structure of the GAN is as shown in the appendix Figure 2 as shown.

[0049] Traditional unsupervised GANs have no control over the patterns of the generated data and it is difficult to control the generation process of the model. To improve the uncontrollability in the unsupervised GAN network, the present invention adopts a conditional generative adversarial network (CGAN), which guides the data generation process to generate specific samples by placing the model on additional information. CGAN is an extension of the GAN framework and is exactly the same during the training process. The difference between CGAN and GAN lies in the input quantity. The input of the GAN generator only contains random noise, while CGAN also adds label information to the input, solving the uncontrollability problem of GAN.

[0050] To solve the problem of lack of partial discharge fault samples and the imbalance between different types, the present invention adopts a conditional generative network to generate partial discharge fault samples of specific types. The conditional generative adversarial network (CGAN) changes the GAN by adding label information as an additional parameter to the generator. As Figure 1 shown, the CGAN model is a type of supervised learning, and the label information serves as an additional input to the generator and the discriminator. If either the generator or the discriminator can obtain any additional auxiliary information as a discrimination condition, the generative adversarial network model can be infinitely extended to another discrimination condition model. c can be said to be an additional auxiliary information that can be obtained, such as a class label.

[0051] To generate high-quality data samples, the present invention adopts an improved CGAN model including a generator G and a discriminator D. A spatial pooling layer is introduced into the network model of the generator G, increasing the multi-scale feature learning ability of the images in the network and generating more image detail information. There is:

[0052] The loss function L of the generator G is expressed as:

[0053]

[0054] The loss function L of the discriminator D It is expressed as:

[0055]

[0056] In the formula, E represents the expectation of the distribution;

[0057] The goal of the conditional generative adversarial network is expressed as:

[0058]

[0059] The improved CGAN can specifically generate high-quality samples of specific types of partial discharge faults. By adding a spatial pooling layer to the model, it overcomes the problem that the quality of samples generated by traditional GAN ​​is difficult to ensure.

[0060] In order to solve the problem of insufficient accuracy, precision and reliability in the identification of partial discharge fault types in power transformer partial discharge patterns due to the small number of fault samples in the data set and the imbalance in the number of maps of different fault types, an improved CGAN technology is used to expand the currently collected partial discharge maps. The present invention adopts the CGAN model structure as follows Figure 3 As shown in the figure, it consists of 1 generator and 1 discriminator. The discriminator contains 1 input layer, 2 convolutional layers, 2 pooling layers, 2 fully connected layers and 1 image output connection layer; the generator only contains 1 input layer, 2 convolutional layers, 2 pooling layers, 1 fully connected layer and 1 output layer. The optimized discriminator used in the generator and the discriminator is Adam. The specific process of inputting samples into the improved CGAN model to generate partial discharge maps under various specific discharge types is as follows:

[0061] like Figure 4 As shown, all kinds of labels and random Gaussian noise are input into the generator G, and fault samples are generated by the generator, and labels are added to the fault samples;

[0062] The fault samples and their corresponding labels and the original samples and their corresponding labels are respectively input into the discriminator for analysis or judgment. By confronting the generator model and the discriminator model, the model parameters of the system are adjusted and optimized to obtain the optimal generator model and discriminator model as well as the expanded discharge data set.

[0063] Step 3: Using the improved convolutional neural network system as a classifier, inputting the expanded discharge data set to train the classifier, and adjusting the network parameters and weights through continuous training to obtain an optimized classifier;

[0064] As a current multi-layer neural network recognition algorithm that can be widely and effectively applied to image feature recognition problems, the convolutional neural network system (CNN) is a feedforward neural network. One of the main technical performance characteristics is that it has the characteristics of local perception, high sharing of network weights, and can support multiple convolutional kernels. A typical convolutional neural network structure system usually only contains at least the following 5 layers, namely the image input layer, the convolutional layer, the pooling layer, the fully connected layer system, and the output layer. The convolutional connection layer system is mainly composed of multiple image feature plane systems, and each image feature bread system contains multiple neurons. During the feedforward transmission of the image signal, the convolutional connection layer performs image convolution operations on the entire image signal input local area through the convolutional kernel, and extracts all the image feature sets on the input local area image. The biggest role of the design of the convolutional layer is of course to quickly extract some useful features in the image information. The formula is as follows:

[0065]

[0066] Y p = f(Z p ) (7)

[0067] In the formula *, W p is the p-th convolutional kernel; X is the input feature map; b is the bias vector; f is the activation function; Y p is the p-th output feature obtained.

[0068] The pooling layer effectively samples the convolutional layer to the greatest extent, reduces the feature dimension, and reduces the computational amount.

[0069]

[0070] In the formula, x i is the activation value input by each neuron within the sampling area range, and the sampling result of Y is equal to subsampling the m×n areas containing the p-th input activation feature.

[0071] In the present invention, as Figure 5 shown, the improved convolutional neural network system includes an input layer, two convolutional layers, two pooling layers, two fully connected layer systems, a T-ReLU activation function layer, and an output layer. The number of convolutional kernels in the convolutional layer is 2×2. The first convolutional layer is composed of 16 3×3 convolutional kernels, and the second convolutional layer is composed of 32 3×3 convolutional kernels.

[0072] The function expression adopted by the T-ReLU activation function layer is:

[0073]

[0074] Among them, α is an adjustable parameter.

[0075] Step 4: Input the local discharge pattern of the transformer to be recognized into the optimized classifier, and output the fault classification and diagnosis result.

[0076] Embodiment

[0077] 230 partial discharge fault patterns are adopted, including 80 patterns of tip discharge, 60 patterns of floating discharge, 50 patterns of surface discharge, and 40 patterns of air gap discharge; after grayscale processing, the grayscale images of typical tip discharge patterns are as Figure 6 shown, the grayscale images of typical surface discharge patterns are as Figure 7 shown, the grayscale images of typical air gap discharge patterns are as Figure 8 shown, the grayscale images of typical floating discharge patterns are as Figure 9 shown, and the image size is uniformly transformed into 300×150.

[0078] The method of the present invention is compared and analyzed with the existing ROS and SMOTE methods. The multi-scale structural similarity (MS-SSIM) index of the system multi-level structure is used as a reference index to quantitatively analyze the three methods respectively. For the local discharge pattern samples under each local discharge type condition, 50 pairs of real and generated samples are randomly and quantitatively selected respectively, and the arithmetic mean of their MS-SSIM indexes is calculated and shown in Table 1.

[0079] Table 1

[0080]

[0081] The dataset after expanding the collected partial discharge PRPD spectrograms by using the improved CGAN model is shown in Table 2.

[0082] Table 2

[0083]

[0084] The improved convolutional neural network in the present invention is trained. Among them, T-ReLU outputs 4 kinds of recognition probability vectors. The local discharge pattern samples are input into a pre-trained and established convolutional neural network model to obtain a discharge type that can be used for prediction.

[0085] In order to further verify the recognition test effect of the method of the present invention, in the partial discharge pattern sample library, about 400 groups were randomly selected as the recognition training test objects respectively, and the ELM, BPNN algorithms and the method proposed in the method of the present invention were trained. Then, 100 groups of data randomly selected from them were respectively used as the pattern sample library for the pattern recognition test method to recognize and test the recognition results of the above three algorithms, and the confusion matrix of the test results was drawn according to the calculation results. The final recognition test results of BPNN classification are shown in Table 3:

[0086] Table 3

[0087]

[0088] The final recognition test results of ELM classification are shown in Table 4:

[0089] Table 4

[0090]

[0091] The final recognition test results of the improved CGAN-CNN classification are shown in Table 5:

[0092] Table 5

[0093]

[0094] It can be clearly seen from Table 3, Table 4 and Table 5 that the recognition accuracy rate of the CGAN-CNN method reaches the highest, and its precision rate and recall rate are also very high, that is, the recognition precision rate and the recognition recall rate both reach a very high level. Compared with these two other discriminant methods, the CGAN-CNN technology has a very good recognition effect for the recognition of the discharge pattern after partial discharge of power transformers. Since the single-level test method may be more accidental, the method of the present invention is used to test the ELM, BPNN and improved CGAN-CNN three recognition methods 20 times respectively, and the average recognition accuracy rates of the three methods are compared. The result analysis is shown in Table 6 below.

[0095] Table 6

[0096]

[0097] It can be seen from Table 6 that the recognition method proposed in the present invention has achieved very good results, and the average prediction accuracy reaches 91.5%, which is higher than the recognition accuracy rates of other traditional classification methods.

[0098] Through the above method, the method for enhancing and identifying partial discharge data of the transformer of the present invention takes the PRPD pattern of the partial discharge fault type as the object, adds labels to the real samples and inputs them into the CGAN model. The trained generator model has a stable training process and can generate partial discharge pattern samples of specific types. Secondly, the partial discharge patterns under various specific discharge types generated by the improved CGAN model can expand the original discharge data set and increase the diversity of the original data set to better solve the problems of unbalanced and small-sample distribution of the original discharge data set, etc.; finally, to improve the classification effect, the present invention uses the improved CNN as the classifier, inputs the training sample data with expanded and balanced data to train the improved CNN, and continuously trains and adjusts the network parameters and weights to output the fault classification diagnosis result; solves the problem of low accuracy of partial discharge pattern recognition of transformers under unbalanced small-sample data and can improve the fault recognition accuracy.

Claims

1. A method for enhancing and identifying partial discharge data of a transformer, characterized in that The implementation is specifically carried out according to the following steps: Step 1: Extract the PRPD pattern of the partial discharge fault type from the dataset. After graying the pattern, crop it into a unified size as the original sample. After adding labels to the original sample; Step 2: Construct an improved CGAN model. Input the sample into the improved CGAN model to generate partial discharge patterns under various specific discharge types, and obtain an expanded discharge dataset; The improved CGAN model includes a generator G and a discriminator D . By introducing a spatial pooling layer into the network model of the generator G, we have: Loss function of the generator It is expressed as: （2） Loss function of the discriminator It is expressed as: （3） In the formula, E represents the expectation of the distribution; The objective of the conditional generative adversarial network is expressed as: （5）； Step 3: Use the improved convolutional neural network system as a classifier. Input the expanded discharge dataset to train the classifier, and continuously train to adjust the network parameters and weights to obtain an optimized classifier; The improved convolutional neural network system includes an input layer, two convolutional layers, two pooling layers, two fully connected layer systems, a T-ReLU activation function layer, and an output layer. The number of convolutional kernels in the convolutional layer is 2×2. The first convolutional layer consists of 16 3×3 convolutional kernels, and the second convolutional layer consists of 32 3×3 convolutional kernels; The function expression of the T-ReLU activation function layer is: wherein, is an adjustable parameter; Step 4: Input the partial discharge pattern of the transformer to be identified into the optimized classifier, and output the fault classification diagnosis result.

2. The method for enhancing and identifying partial discharge data of a transformer according to claim 1, wherein, The process of graying the pattern in Step 1 is: （1） In the formula, , , are the gray values of the original spectrum at the , , positions on the three channels , and is the gray value at the position after graying the spectrum by the average value method.

3. The method for enhancing and identifying partial discharge data of a transformer according to claim 1, characterized in that, The specific process of inputting the sample into the improved CGAN model to generate partial discharge patterns under various specific discharge types in Step 2 is: Input all types of labels and random Gaussian noise into the generator G. The generator generates fault samples and adds labels to the fault samples; Input the fault samples and their corresponding labels, and the original samples and their corresponding labels into the discriminator for analysis or determination. Through the confrontation between the generator model and the discriminator model, adjust and optimize the model parameters of the system to obtain the optimal generator model, discriminator model, and an expanded discharge dataset.

Images