Partial Discharge (PD) analysis plays an important role in assessing the insulation condition of electrical generators. Conventional analysis of Phase-Resolved Partial Discharge (PRPD) patterns relies heavily on expert interpretation, which is time-consuming and leads to inconsistent results. This research presents an automated PRPD pattern classification system using a deep learning-based artificial intelligence (AI) approach on the CiRA CORE platform. Two Convolutional Neural Network (CNN) models, YOLOv4-tiny and YOLOv7-tiny, were trained on 100 PRPD images with a 90:10 training-testing split. The models achieved mean Average Precision (mAP) at IoU 0.5 of 81.77% and 84.97% for YOLOv4-tiny and YOLOv7-tiny, respectively. Further evaluation used 80 PRPD test images benchmarked against three human experts. The AI system outperformed the experts in both accuracy and speed. YOLOv7-tiny achieved 85.8% accuracy in 10 minutes, while YOLOv4-tiny achieved 79.5% in 12 minutes. Human experts averaged 50% accuracy and required 50.66 minutes. The proposed system reduces analysis time and reliance on expert judgement while improving diagnostic accuracy, supporting a more efficient predictive maintenance strategy for the power generation industry.

Keywords

Artificial Intelligence, CiRA CORE, Deep Learning, Generator, Partial Discharge