Deep Learning for a Comprehensive Transformer Fault Detection Through Vibrational Data

Valerio Rucconi, Letizia De Maria, Daniele Bartalesi, Bau Valecillos, Simone Garatti, Sergio Bittanti

Abstract: This paper presents a comprehensive view of the vibrational analysis preformed on a three-phase oil-immersed transformer by expanding and deepening our previous research works. The aim is to show the virtue of deep neural networks in transformer fault detection, as well as to study the vibrational behavior of the transformer tank, a topic that, in literature, has not yet been thoroughly studied. This analysis, which focuses on transformer windings failures, is based on real vibration data recorded by optical sensors located on the transformer tank. The failure of the windings was reproduced by their loosening in a laboratory environment. The measured vibrational spectra were used to develop a classifier capable of detecting winding loosening. Compared to the literature, the robustness of the obtained classifier against possible sensor misplacement was also investigated, which led to an analysis of the tank locations most relevant to this type of analysis. This analysis proved that neural networks were able to detect fault with a high accuracy, robustly to possible misplacements in the positioning of the sensor, and that the no-load condition performed better the load condition.