Transformer Fault Diagnosis: A Shallow Learning Approach for DGA-Based Incipient Fault Detection

Abstract

Power transformers are exposed to electrical, thermal, and mechanical stresses during operation, leading to the degradation of insulation and the generation of dissolved gases. Utilities use IEEE and IEC standards use dissolved gas analysis (DGA) to detect incipient faults in oil-filled in-service transformers. Traditional gas ratio-based DGA methods, at times inconclusive diagnoses, limiting their effectiveness in scheduling preventive maintenance. This study presents the application of a shallow learning Backpropagation Neural Network (BP-NN) for assessing the condition of normal ageing and classification of incipient faults in oil-immersed power transformers. The model is trained using the concentrations (ppm) of five key gases—H₂, CH₄, C₂H₂, C₂H₄, and C₂H₆—as input features. The classified condition of a transformer is normal ageing and five fault type, namely partial discharge, low-energy and high-energy discharges, and thermal faults across two varying temperature ranges. The data set used for the classification of incipient faults within transformers is that where the fault type is confirmed by physical inspection. The 256 samples used in this work are from published sources, including the IEC TC10 database. The results achieved by the BP-NN demonstrate its capability to accurately classify normal ageing and diagnose five types of faults. For evaluating the performance of the trained NN, the IEEE/IEC method of classification, the benchmark used is the actual fault type. The shallow network of pattern recognition successfully identified the presence of normal ageing and the five fault types. The performance of the test set is 94.73%. The results highlight the potential of BP-NNs for enhanced transformer condition monitoring and early fault detection. As more high-quality labelled data become available, the diagnostic accuracy and robustness of the model are expected to improve further.