Automatic Online Partial Discharge Diagnosis via Deep Learning

Abstract

Partial discharge (PD) may be one of the most common defects widely existing in power systems. If such conditions are left unattended, they can eventually develop into breakdowns., causing significant interruption of services, huge financial lost and serious safety problems. Owing to the large number of electrical components existing across electric equipment on every level within power system, timely detection and proper diagnosis of PD defects are usually quite challenging. The development of artificial intelligence and data analytics has provided new opportunities for early detection and diagnosis of such defects, which can lead to maintenance scheduling optimization and improved system reliability. This paper proposes to build an automated non-invasive online PD diagnosis system that is designed to provide effective classifications of different PD conditions based on deep learning technique. The system incorporates with the use of non-invasive PD monitoring via ultrasonic sensing., deep learning models and advanced feature extraction techniques. Through a comprehensive laboratory case study, it is shown that our proposal is significantly better than traditional PD diagnosis with expert knowledge. More specifically, the proposed deep neural network models with advance feature extraction can provide the best overall PD diagnosis performance, while the proposed convolutional neural network structure can also give comparable supreme results without complex feature extraction.