Abstract

Oil-filled transformers are critical assets in electrical power systems, both economically and operationally. Their condition is assessed through insulation system, which is greatly affected by various degradation mechanisms. Hence, effective fault diagnosis is essential to prolong their lifespan. Early detection and correction of incipient faults through Dissolved Gas Analysis (DGA) are crucial to prevent irreversible damage. Current measurement systems have significant limitations that impede their use in routine monitoring and underscore the need for new, accessible technologies that are both technically and economically viable to efficiently detect incipient faults. This study evaluates the performance of various Machine Learning (ML) techniques to predict the concentrations of hydrogen (H2), methane (CH4), acetylene (C2H2), ethylene (C2H4), and ethane (C2H6) in oil samples subjected to different types of electrical faults, using data from a novel electronic nose (E-Nose) equipped with eleven MOS-type gas sensors. The evaluated ML techniques include Linear Regression (LR), Multivariate Linear Regression (MLR), Principal Component Regression (PCR), Multilayer Perceptron (MLP), Partial Least Squares Regression (PLS), Support Vector Regression (SVR), and Random Forest Regression (RFR). Experimental results from 218 measurement processes revealed that RFR and MLP models exhibited superior performance, with RFR achieving the highest accuracy for predicting H2, C2H2, and C2H6, while MLP excelled for CH4 and C2H4.A comparison with a commercial DGA system using the Duval Pentagon Method confirmed the effectiveness of these models in diagnosing transformer faults. These findings underscore the potential of combining E-Noses with ML techniques as an innovative and efficient solution for early fault diagnosis.