Power Transformer DGA Data Augmentation Using Conditional Tabular Generative Adversarial Network

Dissolved gas analysis (DGA) serves as an effective method for diagnosing transformer faults, but the sample data obtained from the measurement of equipment failure are often insufficient, low in quality due to the presence of noise and are unevenly distributed. These characteristics of the DGA data, especially the imbalance in the fault classes make it a major challenge significantly impacting the accuracy of fault diagnosis. In this paper, conditional tabular generative adversarial network (CTGAN) is proposed to achieve an even probability distribution of the fault classes through generating realistic synthetic data to augment the minority fault classes. To evaluate the performance of the proposed technique descriptive statistics and evenness indices were used to compare the sample homogeneity of the generated synthetic data to the real DGA data. The results showed fidelity of the synthetic data generated by CTGAN in relation to the original DGA data and an increase in the probability of the minority fault classes.