Transformer fault diagnosis method based on GAN-MCNN

Abstract

The invention relates to a transformer fault diagnosis method, and particularly discloses a GAN-MCNN-based transformer fault diagnosis method, which comprises the following steps of: 1, carrying out standardization processing on transformer fault diagnosis data of a sampling set, then increasing original sample feature dimensions from 5 dimensions to 25 dimensions based on a ratio method, and enhancing sample features; 2, increasing the number of minority class fault samples by adopting a generative adversarial network algorithm; a third step of designing a transformer fault identification model based on a multi-scale convolutional neural network by adding a multi-scale convolutional layer on a 1DCNN structure; and a fourth step of training and testing the designed model by using the data set after dimension and quantity enhancement, and performing generalization experiment comparison with some intelligent algorithms, and the result shows that the method provided by the invention has the average recognition accuracy of 93.58%, and the generalization performance is good.

Description

Technical Field

[0001] The present invention belongs to the research field of transformer fault diagnosis, and specifically, is a transformer fault diagnosis method based on a generative adversarial network and a multi-scale convolutional neural network model (GAN-MCNN). Background Art

[0002] As an important equipment for power grid operation, once a transformer fails, it will cause a large area of ​​power outage, seriously affecting regional production and life. Therefore, transformer fault diagnosis has become the focus of condition monitoring research and application.

[0003] At present, transformers in power grids are mainly oil-immersed. When the insulation of transformers ages, a small amount of gas will be dissolved in the insulating oil. Its main components are H2, CH4, C2H4, C2H2, and C2H6. By analyzing the composition and concentration of the gas dissolved in the transformer oil, it is possible to determine whether the transformer is operating normally or the fault type. The dissolved gas analysis (DGA) method for analyzing gas components and content is one of the most important transformer state detection and fault diagnosis methods. The modified three-ratio method, Duval triangle method, Rogers ratio method, etc. based on DGA have simple rules and play an important role, but they all have problems such as incomplete state encoding, fuzzy and unclear encoding boundaries, which leads to low fault diagnosis accuracy. With the development of deep learning, various models that directly classify problems based on sample data have been born one after another. Among them, the convolutional neural network has a small number of parameters due to its weight sharing mechanism, and the receptive field mechanism makes it extract features with higher accuracy, so it has been more widely used. Many scholars have also introduced this model in the field of transformer fault diagnosis. However, since most current models use two-dimensional convolution, the samples need to be constructed in two dimensions, which causes serious damage to the original structure of the samples. Summary of the invention

[0004] The purpose of the present invention is to propose a transformer fault diagnosis method based on a generative adversarial network and a multi-scale convolutional neural network model. In order to retain the original structural characteristics of the sample, a one-dimensional convolution model is adopted. At the same time, in order to enhance the feature extraction ability of the model, two multi-scale convolution modules are designed. Finally, a multi-scale convolutional neural network (MCNN) diagnosis model containing a multi-scale convolution module is designed for transformer fault diagnosis. It can effectively eliminate redundant and useless features, increase data, and effectively improve the fault diagnosis accuracy of the transformer.

[0005] The present invention is implemented by the following technical solution: a transformer fault diagnosis method based on a multi-scale convolutional neural network, comprising the following steps:

[0006] The first step is to standardize the collected transformer fault sample data;

[0007] In the second step, the feature dimension of the original sample was increased from 5 to 25 based on the ratio method, thus enhancing the sample features.

[0008] The third step is to build a GAN network, add minority sample experimental data, and divide it into training set, test set and validation set;

[0009] In the fourth step, a transformer fault identification model based on a multi-scale convolutional neural network was designed, the model parameters were adjusted using the validation set, and then the fault diagnosis results were obtained through the test set.

[0010] Furthermore, the specific steps of the first step are as follows:

[0011] Step 1: Collect transformer DGA sample data, m is the number of samples collected, and n is the number of characteristic gas types collected;

[0012] Specifically, the dissolved gases in the oil mainly include H2, CH4, C2H4, C2H2, C2H6, CO, and CO2. Depending on whether the transformer has a fault and what kind of fault exists, the transformer can be divided into normal, high temperature overheating, high energy discharge, medium and low temperature overheating, low energy discharge, partial discharge, and discharge and overheating.

[0013] Step 2: Perform z-score standardization on the collected fault sample data to obtain

[0014] Furthermore, the specific steps of the second step are as follows: the sample characteristics are constructed using the uncoded ratio method, specifically combining 5 gas capacity and gas ratio characteristics, enhancing the sample characteristics from 5 dimensions to 25 dimensions, and constructing uncoded ratio characteristic parameters of dissolved gas in oil.

[0015] Furthermore, the specific steps of the third step are as follows:

[0016] Step 1: Construct the generator G and the discriminator D;

[0017] Step 2: Build a GAN network and increase the amount of experimental data;

[0018] Step 3: Use five-fold cross validation to divide all fault diagnosis data into training data set and test data set.

[0019] Furthermore, the specific steps of the fourth step are as follows:

[0020] Step 1: Construct a 1DCNN model by cross-combining convolutional layers, pooling layers, and fully connected layers;

[0021] Step 2: Design a multi-scale convolution module;

[0022] Step 3: Set the corresponding initial parameters and build a multi-scale 1DCNN convolution model;

[0023] Step 4: Input the training data set, calculate the loss between the input and the true value and adjust the model parameters;

[0024] Step 5: Input the test data set and get the classification results.

[0025] The beneficial effects of the present invention are as follows: in order to retain the original structural characteristics of the sample, the present invention adopts a one-dimensional convolution model, and in order to enhance the feature extraction capability of the model, two multi-scale convolution modules are designed. Finally, a multi-scale convolutional neural network (MCNN) diagnosis model containing a multi-scale convolution module is designed for transformer fault diagnosis. The model can effectively eliminate redundant and useless features, increase data, and effectively improve the transformer fault diagnosis accuracy. The results show that the average recognition accuracy of the method proposed by the present invention is 93.58%, and the generalization performance is good. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 It is a flowchart of the method involved in the present invention;

[0027] Figure 2 This is a schematic diagram of the working of the Generative Adversarial Neural Network;

[0028] Figure 3 is the recognition accuracy of the multi-scale neural network with different convolution kernels of the embodiment;

[0029] Figure 4 is a multi-scale convolution module of an embodiment;

[0030] Figure 5 is a schematic diagram of the multi-scale convolutional neural network structure of an embodiment;

[0031] Figure 6 is the curve of loss function changing with the number of iterations

[0032] Figure 7 is a curve showing the variation of the accuracy rate in the model training phase with the number of iterations in the embodiment;

[0033] Figure 8 4 is a comparison chart of the diagnosis results of each classification model in the embodiment. DETAILED DESCRIPTION

[0034] In order to make the purpose, technical solution and advantages of the present invention clearer, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. The specific embodiments described here are only used to explain the present invention and are not used to limit the present invention.

[0035] 563 DGA data samples were collected. By using the oversampling algorithm, 1034 transformer operating status data were finally obtained.

[0036] The first step is to standardize the collected transformer fault sample data. The specific steps are as follows:

[0037] S101: Collect transformer DGA sample data

[0038] S102: Perform z-score standardization on the collected fault sample data to obtain

[0039] The specific formula is as follows:

[0040]

[0041] Where: x * is the characteristic quantity obtained after standardization; x is the original data; x is the mean of the original data, which is specifically expressed as σ is the standard deviation of the original data, which is specifically expressed as

[0042] In the second step, the uncoded ratio method is used to construct sample features. The specific steps are as follows;

[0043] Combining the 5 gas capacity and gas ratio features, the uncoded ratio method is used to construct sample features, and the sample features are enhanced from 5 dimensions to 25 dimensions. The specific steps are as follows;

[0044] The specific sample characteristics are as follows: H2, CH4, C2H2, C2H4, C2H6, CH4 / H2, C2H2 / CH4, C2H2 / H2, C2H2 / CH4, C2H2 / C2H6, C2H6 / C2H2, C2H4 / H2, C2H6 / C2H4, C2H6 / H2, C2H6 / CH4, H2 / (T1+H2), CH4 / T1, C2H2 / T1, C2H4 / T1, C2H6 / T1, H2 / T1, CH4 / T2, C2H2 / T2, C2H2 / T2, T2 / T1.

[0045] Where T1 is the total gas content and T2 is the total hydrocarbon in the gas.

[0046] Finally, the sample feature matrix is ​​constructed, that is,

[0047] The third step is to use the adversarial neural network to increase the number of minority fault samples in the training set. The specific steps are as follows:

[0048] S301: construct a generator and a discriminator and form an adversarial neural network;

[0049] S302: increasing the number of samples through an adversarial neural network according to the imbalance ratio of the samples;

[0050] S303: Divide the data set into a training set and a test set in a ratio of 4:1.

[0051] The fourth step is to construct a 1DCNN model by cross-combining convolutional layers, pooling layers, and fully connected layers. The specific steps are as follows;

[0052] S401: construct convolutional layer;

[0053]

[0054] Where: is the activation value of the jth neuron in the l+1th layer, is the activation value of the i-th neuron in the l-th layer, is the weight between the jth neuron in the l+1th layer and the ith neuron in the lth layer, is the bias of all neurons in layer l to the jth neuron in layer l+1.

[0055] S402: Constructing a convergence layer;

[0056]

[0057] Where: represents the value of the tth neuron in the ith feature vector of the lth layer; w is the width of the pooling area; Represents the value corresponding to the neuron in the l+1th layer.

[0058] S403: construct a fully connected layer;

[0059]

[0060] Where: z i is the activation value of the i-th neuron in the output layer; c is the number of categories; p(z i ) is the probability output of the output layer neurons.

[0061] S404: Cross-combining the convolutional layer, the pooling layer and the fully connected layer to form a 1DCNN model;

[0062] The fifth step is to design a multi-scale convolution module and build an MDCNN model. The specific steps are as follows;

[0063] S501: We selected seven convolution kernels of different scales (1×1, 3×1, 5×1, 7×1, 9×1, 11×1, and 15×1), and connected the different convolution kernel combinations as a multi-scale convolution module to 1DCNN. The overall accuracy of the model was measured as follows: Figure 2 , and take the convolution kernel combination with the highest accuracy to build multi-scale convolution modules a and b as follows Figure 3 ;

[0064] S502: Constructed a multi-scale convolutional neural network framework such as Figure 4 It contains 11 layers in total, of which the first two layers are continuous multi-scale convolution layers, followed by normal convolution, pooling and fully connected layers. In order to reduce the distribution difference of input data after each parameter update and to speed up the convergence of model training, batch normalization is set after the output of each layer in the middle layer;

[0065] S503: Input the training data set, calculate the loss between the input and the true value, and adjust the model parameters;

[0066] S504: Input the test data set and obtain the classification result.

[0067] Example:

[0068] First, 563 DGA data samples were collected, including 149 normal states, 103 high-temperature overheating, 60 medium-temperature overheating, 48 low-temperature overheating, 102 high-energy discharges, 63 low-energy discharges, and 38 partial discharges. Then, based on the ratio method, the original sample feature dimension was increased from 5 dimensions to 25 dimensions, enhancing the sample characteristics. Secondly, the number of samples was increased by adversarial neural networks, and finally 1034 transformer operating status data were obtained, including 162 normal states, 146 high-temperature overheating, 138 medium-temperature overheating, 154 low-temperature overheating, 142 high-energy discharges, 154 low-energy discharges, and 138 partial discharges.

[0069] In this embodiment, the standardized data is used to enhance the sample characteristics by the ratio method. The component information of the samples before and after the processing is shown in Table 1.

[0070] Table 1 Component information of reconstructed samples

[0071]

[0072] Then, the experimental data is increased through adversarial neural network. The distribution of sample quantity before and after processing is shown in Table 2.

[0073] Table 2 Comparison of experimental data before and after enhancement

[0074]

[0075] The processed experimental data are divided into a test set and a training set through five-fold cross validation, as shown in Table 3.

[0076] Table 3 Test set and training set

[0077]

[0078] This embodiment uses the Pytorch deep learning model framework to build the multi-scale 1DCNN model proposed in the present invention, and trains and tests the model under the GPU of NVIDIA GeForce GTX 1050@4GB. First, the training sample is input into the model. Then, the number of iterations and batch size are set to 55 and 15 respectively, the initial learning rate is set to 0.0002, and the weights and biases of each layer of neurons in the model are initialized with Gaussian random numbers. Finally, cross entropy is used as the loss function and the Adam optimizer is used to update the parameters.

[0079] Table 4 Multi-scale 1DCNN parameter settings

[0080]

[0081]

[0082] Since the model used has a total of 11 layers and many parameters, it has a certain degree of complexity. In order to train the best parameter combination and prevent the model from overfitting, the model is trained and tested using 5-fold cross validation, and the model with the smallest average test error among the five models is selected as the final model. In order to visualize the model training process, the curves of the accuracy rate and loss function during the model training phase changing with the number of iterations are extracted, as shown in Figure 2. Figure 5 , 6 shown.

[0083] Then the test set data is input into the trained multi-scale convolutional neural network to obtain the following test results:

[0084] Table 5 Model test results

[0085]

[0086] In order to demonstrate the effectiveness of the ratio method feature reconstruction, the original samples, the uncoded ratio method feature samples and the reconstructed feature samples were used as data sets to train and test the model respectively. By comparing the F1 score values ​​of the three, it was finally determined that the sample feature reconstruction enhanced the model performance.

[0087] Table 6 F1 scores of three MCNN models

[0088]

[0089] In order to verify the superiority of this method, the SSA-SVM model, capsule network and CNN model were compared and analyzed respectively. The diagnostic accuracy of each model is shown in the following table. Figure 7 shown.

[0090] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, it should be understood by those skilled in the art that the technical solutions described in the aforementioned embodiments can still be modified, or some or all of the technical features therein can be replaced by equivalents. Therefore, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope defined by the claims of the present invention.

Claims

1. A transformer fault diagnosis method based on GAN-MCNN, characterized in that: The following steps are involved: The first step is to standardize the collected transformer fault sample data; In the second step, the feature dimension of the original sample was increased from 5 to 25 based on the ratio method, thus enhancing the sample features. In the third step, the generative adversarial network algorithm is used to increase the number of minority class fault samples and divide them into training set, test set and validation set; In the fourth step, a transformer fault identification model based on a multi-scale convolutional neural network was designed, the model parameters were adjusted using the validation set, and then the fault diagnosis results were obtained through the test set.

2. A transformer fault diagnosis method based on GAN-MCNN according to claim 1, characterized in that: The specific steps of the first step are as follows: Step 1: Collect transformer DGA sample data, m is the number of samples collected, and n is the number of characteristic gas types collected; Step 2: Perform z-score standardization on the collected fault sample data to obtain 3. A transformer fault diagnosis method based on GAN-MCNN according to claim 2, characterized in that: In step 1, the dissolved gases in the oil mainly include H2, CH4, C2H4, C2H2, C2H6, CO, and CO2. Depending on whether the transformer fails and what kind of fault exists, the transformer can be divided into normal, high-temperature overheating, high-energy discharge, medium- and low-temperature overheating, low-energy discharge, partial discharge, and discharge and overheating.

4. A transformer fault diagnosis method based on GAN-MCNN according to claim 3, characterized in that: The specific steps of the second step are as follows: the sample characteristics are constructed using the uncoded ratio method, specifically combining 5 gas capacity and gas ratio characteristics, enhancing the sample characteristics from 5 dimensions to 25 dimensions, and constructing uncoded ratio characteristic parameters of dissolved gas in oil.

5. A transformer fault diagnosis method based on GAN-MCNN according to claim 1, characterized in that: The specific steps of the third step are as follows: Step 1: Construct the generator G and the discriminator D; Step 2: Build a GAN network to increase the number of minority class samples; Step 3: Use five-fold cross validation to divide all fault diagnosis data into training data set and test data set.

6. A transformer fault diagnosis method based on GAN-MCNN according to claim 1, characterized in that: The specific steps of the fourth step are as follows: Step 1: Construct a 1DCNN model by cross-combining convolutional layers, pooling layers, and fully connected layers; Step 2: Design a multi-scale convolution module; Step 3: Set the corresponding parameters and build a multi-scale 1DCNN convolution model; Step 4: Input the training data set, calculate the loss between the input and the true value and adjust the model parameters; Step 5: Input the test data set and get the classification results.

Images