Hybrid Dempster–Shafer and Random Forest Machine Learning Approach for Fault Diagnosis and Condition Assessment of Power Transformers

Abstract

Dissolved gas analysis is a widely used tool for the early detection of faults in power transformers. However, conventional interpretation methods, such as Rogers, Doernenburg, Key Gas method, and the Duval graphical methods, may yield different diagnoses for the same sample because of their distinct criteria, gas ratios, and decision regions. To address this limitation, this study proposes a hybrid methodology based on weak labeling, Dempster–Shafer evidence fusion, and Random Forest classification. Initially, the outputs of multiple dissolved gas analysis methods are fused to generate consensus pseudo-labels at the global and detailed classification levels. These pseudo-labels were subsequently used to train Random Forest models for fault diagnosis. The methodology was evaluated through internal and external validation, achieving accuracies of 95.97% and 88.71% for external global and detailed classification, respectively. In addition, the final diagnosis incorporates complementary information related to thermal subtypes, paper degradation, oil oxidation, and the possible influence of the on-load tap changer. The results show that the proposed methodology improves diagnostic consistency and achieves favorable performance compared with the evaluated reference strategies, providing a data-efficient alternative for scenarios with limited labeled data.