Dissolved Gas Analysis Using Knowledge-Filtered Oversampling-Based Diverse Stack Learning

Imbalanced distribution of fault classes poses a significant challenge in diagnosing faults in transformers using dissolved gas analysis (DGA). Oversampling and ensemble algorithms have been proposed to deal this issue, while their improvement either limited in specific or less-imbalanced datasets. To overcome this issue, a knowledge-filtered oversampling-based diverse stacking (KODS) method is proposed, which incorporates domain knowledge and expert rules into the artificial intelligence (AI) models to improve the performance of the diagnosis results. First, this method uses DGA knowledge to ensure the diversity and accurate oversampled new instances. Then, diverse classifier stacking is proposed to improve fault classification accuracy by leveraging the complementary strengths of diverse base classifiers. The amalgamation of these two components generates a robust and reliable fault classification model for transformer DGA analysis. The comparison among different algorithms proves the efficiency of the proposed method.