A structured pipeline with an adaptive Bayesian model selector for power transformer fault diagnosis using machine learning and voltage-current technique

Abstract

. Power transformers play a key role in the generation, transmission, and distribution of electrical energy. The stability and continuity of electrical service depend on their efficient operation. During operation, transformers are subjected to several types of faults, which if not corrected, can lead to premature failure. This study presents a structured pipeline for power transformer fault diagnosis based on the Voltage-Current Technique and machine learning, including a model selector that adaptively assigns input data to the optimal classifier according to the environmental noise levels. Three types of faults are simulated: turn-to-turn short circuit, buckling stress, and axial displacement. The diagnosis pipeline integrates transformer modelling, data collection, feature engineering, data transformation, classification using multiple machine learning algorithms, and a model selector. The proposed model provides accurate results in both clean and highly noisy environments, with accuracies ranging from 70% in very noisy conditions (7 dB) to 89% in clean conditions.