Search

This article addresses the challenge of limited data availability of dissolved gas analysis (DGA) in fault diagnosis of power transformers. While various classifiers have been employed for fault classification in the prior research, their accuracy depends heavily on balanced and synthetic training data. Although resampling techniques have been implemented to address that, their statistical validity remains unproven. Synthetic data generation is particularly crucial for imbalanced and scarce DGA fault samples. However, there is limited discussion on generating synthetic data for imbalanced datasets, which can negatively impact classifier training and machine learning (ML) models. This article introduces a novel approach for generating synthetic data to overcome the limited availability of fault data in power transformer diagnosis. The effectiveness of the generated data is ensured through ${z}$ -score normalization of the original gas concentration values. In addition, dimension reduction from 5-D to 2-D data is performed to simplify the data generation process. The proposed approach utilizes t-distributed stochastic neighbor embedding (t-SNE) as a superior dimension reduction technique, verified through reconstruction error and explained variance (Ev) metrics. The performance of the proposed generative adversarial network (GAN)-based approach is compared with four standard algorithms using multiple evaluation metrics. The results demonstrate that the GAN outperforms the other algorithms in generating synthetic data that closely resembles actual data. The proposed method achieves an accuracy of up to 98.10%, surpassing existing methods. This confirms the proposed GAN-based approach, in conjunction with t-SNE, effectively processes actual data for power transformer fault diagnosis, thus overcoming the limitations of limited data availability.

The recent development of artificial intelligence (AI) has opened new avenues in processing parts per million (ppm) for fault detection through dissolved gas analysis (DGA). According to the latest IEC and IEEE standards, the existing methods are only applicable on single fault occurrence. The paper focuses on the challenge of detecting multiple faults occurring simultaneously in cases involving many faults using AI. Further, an inadequate training sample for classification and unavailability of balanced per-fault data reduces the model generalization, increases the risk of overfitting and biased learning towards the majority class. The proposed approach involves normalizing raw ppm values using z-score normalization, reducing dimensionality through t-distributed stochastic neighbor embedding (t-SNE), and synthesizing data using a generative adversarial network (GAN). Additionally, the parameters of error-correcting output codes (ECOC) and forest classifiers are optimized using a genetic algorithm (GA), efficiently solving multiple faults. F1 score, area under curve (AUC), and k-fold loss are used to evaluate fitness for improved classifier performance. This method outperforms the Duval method, and the data synthesis represents a new contribution to the field. The proposed method can achieve an overall accuracy of 99.6%, 98.6%, and 97.3% for the 9, 15, and 31 classes, respectively.