A method for diagnosing power transformer faults

Abstract

This invention relates to the field of power equipment fault diagnosis technology, specifically to a power transformer fault diagnosis method based on an improved auxiliary classification generative adversarial network (IACGAN). The power transformer fault diagnosis method includes the following steps: IACGAN model construction, data acquisition, dataset construction, partitioning of training and test sets, construction of a fault diagnosis model, and acquisition of fault diagnosis results. The IACGAN proposed in this invention uses a ResNet based on skip connections as the generator, discriminator, and auxiliary classifier of the ACGAN. The discriminator and auxiliary classifier distinguish effective deep features in the generated and real data, verifying the effectiveness of the generator's data generation, avoiding model degradation problems, improving the stability of the algorithm structure, and enabling accurate fault diagnosis.

Description

Technical Field

[0001] This invention relates to the field of power equipment fault diagnosis technology, specifically to a power transformer fault diagnosis method based on an improved auxiliary classification generative adversarial network (IACGAN). Background Technology

[0002] Dissolved gas analysis (DGA) data can reflect the real-time internal operating status of transformers, making it widely used in transformer fault diagnosis. However, existing diagnostic methods largely rely on comprehensive datasets for model training. In reality, transformers typically operate under normal conditions, and DGA fault data is relatively scarce. Furthermore, there is significant imbalance in the data across different fault categories, leading to reduced diagnostic accuracy when using such data directly. Currently, data generation methods are primarily employed to address these imbalances.

[0003] Currently, with the rapid development of artificial intelligence, methods combining DGA data with artificial intelligence are widely used in transformer fault diagnosis, such as Support Vector Machines and Artificial Neural Networks. However, practice has shown that these methods have insufficient learning capabilities. With the rapid development of deep learning, such as Deep Belief Networks (DBN), Stacked Autoencoders (SDAE), Convolutional Neural Networks (CNN), Long Short-Term Memory Networks (LSTM), and Attention mechanisms, these methods have strong learning and generalization capabilities, leading to their increasing application in transformer fault diagnosis. However, most of the original data is taken from relatively balanced and complete fault data; in reality, various types of data exhibit significant imbalance.

[0004] Typically, when modeling a dataset with few faulty samples, the model tends to "prefer" diagnosing the test dataset into the category that predominates in the training data. For example, if the proportion of normal data in the training set is much greater than that of faulty data, and the ratio of normal to faulty data in the test dataset is 9990:10, even if all data is diagnosed as normal, the model will still show an extremely high accuracy of 99.9%. However, such a model is of no use in diagnosing faulty data, because that 0.1% misjudgment can lead to serious consequences.

[0005] Therefore, research on fault diagnosis technology for data imbalance is of practical significance to ensure the stable operation and maintenance of power transformer equipment. This invention's research on power transformer fault diagnosis technology mainly consists of two parts: data generation and fault diagnosis. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the existing technology and provide a power transformer fault diagnosis method based on an improved auxiliary classification generative adversarial network for power transformer data generation and fault diagnosis, which takes into account the imbalance of DGA data.

[0007] The technical solution of this invention is implemented as follows:

[0008] A method for diagnosing faults in power transformers includes the following steps:

[0009] Step 1: IACGAN model construction. By repeatedly stacking one-dimensional convolutional layers and one-dimensional pooling layers, a generator, discriminator and auxiliary classifier of ACGAN auxiliary classification generative adversarial network suitable for power transformer DGA data structure are constructed. At the same time, based on the ResNet principle, skip connections are made every two one-dimensional convolutional layers to form the IACGAN model.

[0010] Step 2: Data acquisition, collecting DGA data samples of power transformers under normal operating conditions and different fault operating conditions;

[0011] Step 3: Dataset construction. The uncoded ratio feature data of the DGA data samples in Step 2 and their corresponding fault type labels are used as the input and output of the IACGAN model for training. The generator of the IACGAN model generates a large number of fault datasets corresponding to specific type labels, thereby constructing the required dataset.

[0012] Step 4: Divide the dataset obtained in Step 3 into training and test sets.

[0013] Step 5: Construct a fault diagnosis model. Use the training set obtained in Step 4 to train the IACGAN model. After training, extract the parameters of the discriminator and auxiliary classifier in the IACGAN model separately to construct the fault diagnosis model.

[0014] Step 6: Obtain the fault diagnosis results. Input the test set obtained in Step 4 into the fault diagnosis model. The test data in the test set is first input into the discriminator of the IACGAN model to extract discriminative features. Then, the discriminative features are input into the auxiliary classifier of the IACGAN model to obtain the specific fault category.

[0015] Preferably, the specific steps for using the training set in step 5 to train the IACGAN model are as follows:

[0016] 5.1 Input the white noise and fault type labels into the generator of the IACGAN model to obtain the generated data;

[0017] 5.2 Input the generated data and real data into the discriminator of the IACGAN model, combine the fault type labels, and calculate L using the ACGAN objective function. s L s This refers to the ACGAN objective function used to measure the likelihood that the data source is real data.

[0018] 5.3 Input the generated data and real data into the discriminator of the IACGAN model, extract the discriminative features of the discriminator and input them into the auxiliary classifier of the IACGAN model, combine the fault type labels and calculate L using the ACGAN objective function. c Among them, L c This refers to the objective function of ACGAN, which measures the likelihood of data belonging to a specific category.

[0019] 5.4 Minimize L s +L c The gradient descent method is used to update the parameters of the generator, discriminator, and auxiliary classifier of the IACGAN model.

[0020] 5.5 Repeat steps 5.1-5.4 until training is complete.

[0021] Preferably, in step 4, the dataset obtained in step 3 is randomly divided in a 7:3 ratio to obtain the required training set and test set respectively.

[0022] The positive effects of this invention are: 1. The solution proposed in this application consists of feature extraction and fault diagnosis technologies, which are simple to operate and easy to implement; 2. The proposed IACGAN is an improvement on the ACGAN algorithm, and for the first time uses ResNet based on skip connections as the generator, discriminator, and auxiliary classifier of ACGAN. The discriminator and auxiliary classifier distinguish the effective deep features in the generated data and the real data, verifying the effectiveness of the data generated by the generator, avoiding the model degradation problem, improving the stability of the algorithm structure, and enabling accurate fault diagnosis; 3. Compared with the traditional oversampling SMOTE and ACGAN algorithms, the solution proposed in this application based on the IACGAN method achieves better performance improvement. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the GAN structure.

[0024] Figure 2 This is a schematic diagram of the ACGAN structure.

[0025] Figure 3 This is a schematic diagram comparing the structures of direct connections and ResNet skip connections.

[0026] Figure 4This is a schematic diagram of the IACGAN model network structure.

[0027] Figure 5 This is a flowchart of the fault diagnosis process of the present invention.

[0028] Figure 6 This is the loss function curve of the IACGAN model network.

[0029] Figure 7 Visualization results of the test set discrimination features before and after the generation of data for fault diagnosis model 1.

[0030] Figure 8 Visualization results of the test set discrimination features before and after the generation of data for fault diagnosis model 2. Detailed Implementation

[0031] Step 1: IACGAN Model Construction

[0032] By repeatedly stacking one-dimensional convolutional layers and one-dimensional pooling layers, a generator, discriminator, and auxiliary classifier for the auxiliary classification generative adversarial network (ACGAN) suitable for the DGA data structure of power transformers are constructed. At the same time, based on the ResNet principle, skip connections are made every two one-dimensional convolutional layers to form the IACGAN model.

[0033] (1.1) The principle of GAN

[0034] A GAN consists of a generator network G and a discriminator network D, and its network structure is as follows: Figure 1 As shown.

[0035] The objective function of GAN is:

[0036]

[0037] In the formula: min represents the discriminator minimizing the expected value of the objective function when the generated data G(z) from the input z enters the discriminator, causing the discriminator to mistakenly believe that the data is real data; max represents the discriminator maximizing the expected value of the objective function, enabling it to identify the sample as fake data generated by the generator network to the greatest extent. Through the game between the generator and the discriminator, a Nash equilibrium is eventually reached, at which point the discriminator can no longer determine whether the input sample is real data or generated data, and the distribution of generated data at this point is closest to the distribution of real data.

[0038] (1.2) The principle of ACGAN

[0039] Because GANs were proposed relatively early, they have many drawbacks, such as training difficulties and pattern collapse. Furthermore, due to the lack of utilization of data category information, the generated data does not know its specific category, and it is impossible to start trying to generate data with a fixed category.

[0040] ACGAN, by incorporating category information, can generate data of a specified category. Furthermore, the addition of an auxiliary classifier allows ACGAN to both distinguish between true and false data and perform data classification, attracting widespread attention from scholars. Its structure is as follows: Figure 2 As shown.

[0041] The objective function of ACGAN is:

[0042] L c =E[logP(S=real|X real )]+

[0043] E[logP(S=fake|X fake )]

[0044] L s =E[logP(C=c|X real )]+

[0045] E[logP(C=c|X fake )]

[0046] In the formula: L s Used to measure the likelihood that the data source is real data; L c This is used to measure the probability that data belongs to a specific category. Throughout the training process, the discriminator D and the auxiliary classifier Q strive to distinguish between real samples and generated samples and to accurately classify the data, while the generator G strives to make the generated data as close as possible to real data and to be correctly classified.

[0047] (1.3) The principle of ResNet

[0048] Practice shows that deeper neural networks exhibit more complex feature extraction patterns, but network degradation occurs as network depth increases. ResNet primarily addresses this by proposing a skip connection structure to replace the direct connections between layers in traditional CNNs. Its structure is as follows: Figure 3 As shown, Figure 3 The left side (a) shows direct connections, and the right side (b) shows the skip connection structure of ResNet. From... Figure 3 It can be seen that even if the weight layer does not learn any features, i.e., F(x) = 0, its output will not be 0, thus avoiding the degradation problem.

[0049] (1.4) Structure of IACGAN

[0050] Based on the above principles, a network model suitable for power transformer data generation and fault diagnosis is designed by combining ACGAN and ResNet. The network structure of the IACGAN model is as follows: Figure 4As shown in the figure, Conv1D is a one-dimensional convolutional layer, AvgPool1D is a one-dimensional pooling layer, Dense is a fully connected layer, and Feature refers to the deep features extracted by the discriminator.

[0051] like Figure 4 As shown, firstly, by repeatedly stacking Conv1D and AvgPool1D, a generator, discriminator, and auxiliary classifier suitable for the DGA data structure of power transformers are constructed in ACGAN. Simultaneously, based on the ResNet concept, skip connections are made every two Conv1D layers to prevent degradation caused by insufficient sample data and a deep network structure. Secondly, the uncoded ratio feature data of the DGA data reflecting the transformer's operating status and its corresponding fault category labels are used as input and output for network training. The generator generates a large number of fault datasets corresponding to specific category labels. Then, the discriminator and auxiliary classifier learn more effective deep features from real and generated data, which are better used to distinguish between real and false samples while achieving accurate fault diagnosis. Ultimately, accurate diagnosis is achieved in power transformer fault diagnosis under conditions of insufficient and imbalanced fault datasets. This collectively constitutes what is referred to as IACGAN. The specific fault diagnosis process is shown in steps 2-6.

[0052] Step 2: Collect DGA data samples of power transformers under normal operating conditions and different fault operating conditions, as shown in the attached document. Figure 5 Data acquisition and feature data extraction in the process;

[0053] (2.1) Definition of data characteristics

[0054] When a power transformer experiences overheating or discharge faults, the transformer oil decomposes and produces gases. These gases mainly include H2, CH4, C2H4, C2H6, C2H2, CO, CO2, O2, and N2. Because CO, CO2, O2, and N2 are highly dispersed and difficult to obtain, they are usually not included in the characteristic gases for fault diagnosis. The characteristic quantities are selected from the content of the five gases and their 22 ratios shown in Table 1 below.

[0055] Table 1 Transformer Fault Characteristics

[0056]

[0057] Step 3: Dataset Construction. The uncoded ratio feature data obtained after standardizing the DGA data samples from Step 2 (after feature data extraction), and the fault type labels corresponding to each DGA data sample after feature data extraction, are used as the input and output of the IACGAN model for training. The IACGAN model generator produces a large number of fault datasets corresponding to specific type labels, thus constructing the required dataset; see attached. Figure 5Data standardization and training to generate datasets;

[0058] Step 4: Split the dataset into training and test sets. Randomly divide the dataset in a 7:3 ratio to obtain the required training and test sets; see attached. Figure 5 The training and test sets are divided in the process;

[0059] Step 5: Construct a fault diagnosis model. Use the training set to train IACGAN. After training, extract the parameters of the discriminator and auxiliary classifier from IACGAN separately to construct the fault diagnosis model. See Appendix. Figure 5 The fault diagnosis model for the right-middle region is established by inputting the training set into the generator in IACGAN for training to generate data. The generated data is then sequentially fed into the discriminator and auxiliary classifier for training. The data generated by the auxiliary classifier is then fed into the generator for training, achieving alternating training. After training, the parameters of the discriminator and auxiliary classifier in IACGAN are extracted separately, and the fault diagnosis model is constructed based on this, as shown in the attached figure. Figure 5 (As shown in the lower middle area); The specific steps for using the training set to train the IACGAN model are as follows:

[0060] 5.1 Input the white noise and fault type labels into the generator of the IACGAN model to obtain the generated data;

[0061] 5.2 Input the generated data and real data into the discriminator of the IACGAN model, combine the fault type labels, and calculate L using the ACGAN objective function. s L s This refers to the ACGAN objective function used to measure the likelihood that the data source is real data.

[0062] 5.3 Input the generated data and real data into the discriminator of the IACGAN model, extract the discriminative features of the discriminator and input them into the auxiliary classifier of the IACGAN model, combine the fault type labels and calculate L using the ACGAN objective function. c Among them, L c This refers to the objective function of ACGAN, which measures the likelihood of data belonging to a specific category.

[0063] 5.4 Minimize L s +L c The gradient descent method is used to update the parameters of the generator, discriminator, and auxiliary classifier of the IACGAN model.

[0064] 5.5 Repeat steps 5.1-5.4 until training is complete.

[0065] Step 6: Obtain the fault diagnosis results, see appendix. Figure 5The flowchart in the lower middle area shows that the test set obtained in step 4 is input into the fault diagnosis model constructed in step 5. The test data in the test set is first input into the discriminator of the IACGAN model to extract the discriminant features. Then, the discriminant features are input into the auxiliary classifier of the IACGAN model to obtain the specific fault category, and finally the fault diagnosis result can be obtained.

[0066] Case Analysis

[0067] I. Data Analysis

[0068] According to IEC60599, the operating status of transformers is divided into 7 categories as shown in Table 2 below, with fault numbers F1-F7.

[0069] Table 2 Transformer Fault Types

[0070] Fault Number Fault type Fault Number Fault type <![CDATA[F1]]> normal <![CDATA[F5]]> Partial discharge <![CDATA[F2]]> Low temperature overheating <![CDATA[F6]]> Low-energy discharge <![CDATA[F3]]> medium temperature overheating <![CDATA[F7]]> High-energy discharge <![CDATA[F4]]> High temperature overheating

[0071] Seven types of operating state sample data were collected from a power company and some authoritative literature. The feature dataset was constructed using the method in Table 1. After manually removing some abnormal samples through outlier detection, 1138 sets of labeled real data were obtained. The training set and test set were randomly divided in a 7:3 ratio. The final sample distribution is shown in Table 3. As can be seen from Table 3, there are significant differences in the number of samples in each state. Some fault types have fewer data, which is consistent with the aforementioned insufficiency and imbalance.

[0072] Table 3. Composition of Fault Samples

[0073] Dataset <![CDATA[F1]]> <![CDATA[F2]]> <![CDATA[F3]]> <![CDATA[F4]]> <![CDATA[F5]]> <![CDATA[F6]]> <![CDATA[F7]]> total train 32 86 55 18 32 109 465 797 test 12 47 15 6 23 43 195 341 total 44 133 70 24 55 152 660 1138

[0074] Based on the training data in Table 3 above (i.e., data that does not include oversampling or generated data), a ResNet model with the same 'discriminator + auxiliary classifier' structure was directly constructed and used as the baseline model. The diagnostic results for the corresponding test set are shown in Table 4 below. Since the data is imbalanced, traditional accuracy metrics have limited effectiveness; therefore, precision, recall, and f1 score were used as evaluation metrics.

[0075] Table 4. Fault Diagnosis Results Based on Original Data

[0076] Precision Recall f1-score <![CDATA[F1]]> 0.5714 0.6667 0.6154 <![CDATA[F2]]> 0.8824 0.6383 0.7407 <![CDATA[F3]]> 1 0.1333 0.2353 <![CDATA[F4]]> 1 0.1667 0.2857 <![CDATA[F5]]> 0 0 0 <![CDATA[F6]]> 0.6324 1 0.7748 <![CDATA[F7]]> 0.8784 1 0.9353 average value 0.7092 0.5150 0.5125

[0077] As shown in Table 4, except for F7, which has a relatively superior diagnostic performance, the diagnostic performance of the other fault types is poor. Furthermore, because the training samples used for the other fault types are limited, the models have not learned the necessary features, resulting in the worst diagnostic performance. Therefore, it is necessary to oversample or generate new data for fault data with limited sample sizes.

[0078] II. Validation of the IACGAN Method

[0079] Adopting such Figure 4 The network structure is shown in the figure, with parameters labeled for each layer. In addition, the generator's intermediate layers use Leaky-ReLU activation functions, the output layer uses Tanh activation functions, and the learning rate is 0.0005; the discriminator's intermediate layers use Leaky-ReLU activation functions, the output layer uses Sigmoid activation functions, and the learning rate is 0.0001; the discriminator's intermediate layers use Leaky-ReLU activation functions, the output layer uses Softmax activation functions, and the learning rate is 0.0005. IACGAN uses the RMSProp optimizer with a batch size of 64 and 200 training iterations. The training process employs an optimization strategy based on Wasserstein distance and gradient penalty, with a penalty coefficient of 10. After 200 iterations of training, the loss curves of each IACGAN network are shown below. Figure 6 As shown, the horizontal axis represents the number of training iterations, the vertical axis represents the loss value, the dashed line represents the discriminator loss, the curve that intersects the discriminator loss curve in an X-shape is the generator loss curve, and the curve that first decreases exponentially and then approaches 0 is the auxiliary classifier loss curve.

[0080] from Figure 6 It can be seen that the generator loss is high and the discriminator loss is low at the beginning of the training stage. This is because the network has not yet truly learned the distribution model of the power transformer feature data, and the discriminator can easily identify the authenticity of the data. After 100 training iterations, the losses of the generator and the discriminator tend to stabilize. At this point, the two reach a Nash equilibrium through game theory. The discriminator can no longer easily identify whether the data generated by the generator is fake data. At the same time, the loss of the auxiliary classifier approaches 0, indicating that the network can also effectively classify and diagnose faults between generated data and real data.

[0081] To demonstrate the impact of using IACGAN for data generation on fault diagnosis results.

[0082] Based on the principle that if the generated data improves the fault diagnosis effect, the inter-class discrimination of the test set will increase, the test set is input into the following two fault diagnosis models. Features from the layer before the classification layer, i.e., discriminative features, are extracted, and principal component analysis (PCA) is used for dimensionality reduction and visualization. The results are as follows: Figure 7 and 8 As shown, the two fault diagnosis models are set up as follows:

[0083] Model 1: Directly train a fault diagnosis model with the same structure as 'discriminator + auxiliary classifier'.

[0084] Model 2: The fault diagnosis model constructed by extracting the 'discriminator + auxiliary classifier' from the trained IACGAN.

[0085] Model 1 and Model 2 have completely identical structures. By extracting the discriminant features of the test sets of the two models, we can use them to analyze the effects before and after the data generation.

[0086] from Figure 7 and 8 Visualization of the discriminative features on the test set shows that... Figure 7 The clusters formed after dimensionality reduction of the discriminant features of F2 and F7 are more concentrated, while the clusters formed after dimensionality reduction of the discriminant features of F4 and F5 in the test set are relatively scattered due to the smaller number of training samples, making it difficult to distinguish them correctly from other classes. This corresponds to the diagnostic results in Table 4. Meanwhile, Figure 8 The 'clusters' formed by dimensionality reduction of discriminative features of different categories are relatively... Figure 7 The significant improvement reflects the effectiveness of the IACGAN method proposed in this invention. This is because the generator continuously generates data during the training process to train the subsequent discriminator and auxiliary classifier, thereby improving the imbalance of the original power transformer fault feature dataset.

[0087] III. Comparative Analysis

[0088] The proposed solution in this application is based on IACGAN for data generation and fault diagnosis, integrating these two processes. Existing methods generally follow these steps: 1) generating new data samples from the training set using oversampling or data generation methods; 2) combining the training set with the generated data to form a new training set for fault diagnosis modeling. Therefore, the IACGAN method proposed in this application cannot be directly compared with traditional methods. For this reason, the following methods are selected for comparative analysis:

[0089] Method 1: SMOTE oversampling. This method is compared with traditional oversampling methods to verify its effectiveness over traditional methods.

[0090] Method 2: ACGAN data generation. This involves direct connections between layers to verify the effectiveness of our proposed method compared to the unimproved method.

[0091] Method 3: Extract the generator from the trained IACGAN separately to generate samples.

[0092] The data generated by methods 1, 2, and 3 above were added to the original training set, and ResNet networks with the same parameters were used for training. After training, the data were input into the test set to obtain the fault diagnosis results. The average precision (AP), average recall (AR), and average f1 score (A-f1) of each method are recorded in Table 5.

[0093] Table 5 Comparison of fault diagnosis results under different methods

[0094] method AP AR A-f1 Raw data 0.7092 0.5150 0.5125 1 0.8325 0.7027 0.7231 2 0.8148 0.7815 0.7830 3 0.8991 0.8351 0.8587

[0095] As shown in Table 5, compared with directly using the original data for fault diagnosis modeling, the results of the fault diagnosis test set after oversampling or data generation are all improved. Specifically, after oversampling using SMOTE in Method 1, the corresponding AP, AR, and A-f1 are improved by 12.33%, 18.77%, and 21.06% respectively compared with the original data; after data generation using ACGAN in Method 2, the corresponding indicators are improved by 10.56%, 26.65%, and 27.05% respectively; while the corresponding indicators of the proposed solution are improved by 18.99%, 32.01%, and 34.62% respectively. At the same time, compared with Methods 1 and 2, the proposed method has a higher performance improvement. Therefore, the proposed solution is superior to the traditional oversampling method and the original ACGAN.

Claims

1. A method for diagnosing faults in power transformers, characterized in that, Includes the following steps: Step 1: IACGAN model construction. By repeatedly stacking one-dimensional convolutional layers and one-dimensional pooling layers, a generator, discriminator and auxiliary classifier of ACGAN auxiliary classification generative adversarial network suitable for power transformer DGA data structure are constructed. At the same time, based on the ResNet principle, skip connections are made every two one-dimensional convolutional layers to form the IACGAN model. Step 2: Data acquisition, collecting DGA data samples of power transformers under normal operating conditions and different fault operating conditions; Step 3: Dataset construction. The uncoded ratio feature data of the DGA data samples in Step 2 and their corresponding fault type labels are used as the input and output of the IACGAN model for training. The generator of the IACGAN model generates a large number of fault datasets corresponding to specific type labels, thereby constructing the required dataset. Step 4: Divide the dataset obtained in Step 3 into training and test sets. Step 5: Construct a fault diagnosis model. Use the training set obtained in Step 4 to train the IACGAN model. After training, extract the parameters of the discriminator and auxiliary classifier in the IACGAN model separately to construct the fault diagnosis model. Step 6: Obtain the fault diagnosis results. Input the test set obtained in Step 4 into the fault diagnosis model. The test data in the test set is first input into the discriminator of the IACGAN model to extract discriminative features. Then, the discriminative features are input into the auxiliary classifier of the IACGAN model to obtain the specific fault category.

2. The power transformer fault diagnosis method according to claim 1, characterized in that: The specific steps for using the training set in step 5 to train the IACGAN model are as follows: 5.1 Input the white noise and fault type labels into the generator of the IACGAN model to obtain the generated data; 5.2 Input the generated data and real data into the discriminator of the IACGAN model, combine the fault type labels, and calculate L using the ACGAN objective function. s L s This refers to the ACGAN objective function used to measure the likelihood that the data source is real data. 5.3 Input the generated data and real data into the discriminator of the IACGAN model, extract the discriminative features of the discriminator and input them into the auxiliary classifier of the IACGAN model, combine the fault type labels and calculate L using the ACGAN objective function. c Among them, L c This refers to the objective function of ACGAN, which measures the likelihood of data belonging to a specific category. 5.4 Minimize L s +L c The gradient descent method is used to update the parameters of the generator, discriminator, and auxiliary classifier of the IACGAN model. 5.5 Repeat steps 5.1-5.4 until training is complete.

3. The power transformer fault diagnosis method according to claim 1, characterized in that: In step 4, the dataset obtained in step 3 is randomly divided in a 7:3 ratio to obtain the required training set and test set.

Images