Development of a Methodology to Automatically Identify Active PD Sources from Phase Resolved Partial Discharge Patterns

Hydro-Québec is working on the development of a prognostic approach for hydrogenerators based on the modelling of failure mechanisms in the form of graph. The failure mechanism graph consists of several cause-and-effect chains formed by the combination of sequential physical degradation states that ultimately lead to a failure. The first step of the approach is to automatically identify the active failure mechanisms based on the analysis of observable symptoms extracted from diagnostic tools. For the stator winding of hydrogenerators, more than 100 failure mechanisms have been consigned and most of them involve the presence of partial discharge (PD) activity. PD measurements as a diagnostic tool for hydrogenerators have been performed over the last 30 years at Hydro-Québec. During these years, a large PD database has been built up, including the results from two measurement techniques: 2D Partial Discharge Analyzer (PDA) as well as Phase Resolved Partial Discharge (PRPD) patterns. These two measurement techniques provide complementary results that will be combined in the prognostic approach. In this context, the objective of this paper is to present the development of a two-phase methodology aiming to automatically identify the active PD sources from PRPD patterns using machine learning techniques. The first phase of the proposed methodology is the preprocessing where filters have been applied to adjust the phase shift of each PRPD pattern. Then, two U-Net learning models combined with a Convolutional Neural Network (CNN) learning model are used to identify the presence of gap PD activity and to extract the main active PD sources. The next and final phase of the methodology is the PD source recognition which is based on the aggregation of results from six CNN learning models. These CNN models were trained on different features extracted from the preprocessed PRPD patterns resulting from the first phase of the methodology. Results showed that using the U-net learning model is very effective in separating PD sources and that the precision of the PD source recognition is significantly improved by combining the six CNN models instead of using each individual CNN learning model.

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.