Ensemble Deep-Learning Model for Phase-Resolved Partial Discharge Diagnosis in Hydrogenerators

Abstract

Condition monitoring of hydrogenerators is extremely essential for the optimisation of the power supply system’s maintenance, especially in a context where energy consumption will increase more and more. Industry investigations have consistently shown that the electrical insulation used in stator windings is one of the most potential components for breakdown. One of the major causes of this breakdown was shown to be partial discharges (PDs). PDs are designated as defects in electrical insulation, and their presence may cause future insulation failures in equipment. Online monitoring of PD activity is essential to anticipate degradation and avoid a sudden shutdown with a high economic impact. Thus, automating the recognition of PD sources by an intelligent system is a major issue in a Condition-Based Maintenance (CBM) process for hydrogenators. First, a method based on the U-Net model is used to isolate the main sources of a Phase-Resolved PD signal (PRPD). Afterwards, a set of deep learning models (DLM) is then applied to automatically recognize the PD sources. Finally, a decision-making method is then proposed to find the optimal output category considering the posterior and the prior probability estimated by various individual models. Experimental results on real signals from hydrogenerators show an accuracy of 87% on the detection of gap sources, and respectively 91%, 95%, 96% and 94% for the Internal, Corona, Slot and Delamination PD. The proposed decision-making method shows a significant advantage of the final decision over the performances of each individual DL model.