Classification des défauts des transformateurs électriques par des méthodes d’intelligence artificielle

Abstract

This work focuses on improving the diagnosis of power transformers, which are critical components for the stability and efficiency of electrical networks. Continuous monitoring of this equipment is essential to prevent major failures. This research is primarily based on Dissolved Gas Analysis (DGA), an effective technique for identifying hidden internal faults. In this study, we explored a modern approach based on artificial intelligence and machine learning. Using a dataset of 709 samples, several normalization methods were tested. The simulation results indicate that combining Median IQR normalization with the Bagged Trees model, evaluated through a 5-fold cross-validation, delivers the highest diagnostic accuracy. Furthermore, an ablation study proved that a complete vector of 5 gases remains indispensable for maintaining reliable classification, paving the way toward optimal predictive maintenance for electrical infrastructures.