Identify Problems of Electrical Insulators using Multi-Task Learning

Abstract

Electrical insulators are important parts of electrical systems that often have some popular issues and require maintenance such as broken and dirty. Automatic detection using computer vision models for these problems could help to enhance the safety and continuity of the electrical operation in a proactive and timely manner with lower operational and maintenance costs. In this paper, we develop a multi-task model adopting Efficient-net as our base architecture following three branches for simultaneously three learning tasks such as identifying the type, cleanness, and broken status of the insulators. To train this multi-task model, we cleaned the dataset and performed the data augmentation of the practical dataset comprising 1500 images of electrical insulators provided by CPC Vietnam collected from pole-mounted surveillance cameras and drone survey flights. Throughout the extensive evaluation, the proposed muti-task models outperformed the single-task model by around 15-20% and demonstrated a robust design for identifying multiple problems of electrical equipment.