Twin Network Approach for Multifault Classification in Power Transformers Using Enriched Features and Zero Synthetic Data

This study introduces a new deep learning-based methodology for classifying single-label and multilabel faults in power transformers. The proposed approach leverages a deep twin network architecture to effectively analyze dissolved gases (DG) without relying on conventionally used synthetic data techniques. Initially, DGA data in ppm format is log normalized, followed by predicting transformer age using a deep Bidirectional Long Short-Term Memory (BiLSTM) network. The method has another novelty of generating a large number of features from the DG data through feature engineering. Feature selection is then performed using Neighborhood Component Analysis (NCA), followed by semi-supervised dimension reduction using t-SNE. To address multi-labeled scenarios, combined fault labels are created and classified using a deep twin network. The performance of the proposed method is evaluated through testing and comparison using the Silhouette Index (SI). Results demonstrate that the proposed methodology achieves high accuracy in classifying multiple faults, showcasing better cluster separation than other state-of-the-art methods.