A Deep Learning-based approach for Fault Detection of Power Lines

Abstract

A transmission network is the most crucial part of modern infrastructure. However, it requires an extensive amount of power line inspection each year to maintain, and with an increased interest in replacing large helicopters with drones for this process, the possibility of including AI is equally compelling. This thesis goes into the second part by taking a deep learning-based approach in the interest of fault detection. A literature review illustrates that earlier research has some to none understanding of the complexity re-quired for inspection. Due to the advancement in object detection and classification, this thesis has identified and implemented an applicable model capable of giving state-of-the-art accuracy in electrical pole and component detection by dividing the process into multiple layers. This thesis takes as well and proposes a new method that presented great result in assuring more reliable fault detection and is a way to improve the quality of images taken by drones. The pole detection layer gave 97.7 mAP, the component detection layer reached 95.6mAP, the fault classifier delivered an accuracy of 93%, and the proposed quality classifier had an accuracy of 93% as well. The presented approach illustrates the possibility of phasing the physical inspection out. The amount of component labeled that must be available for algorithmic training to surpass a human expert is not readily available. Nevertheless, the presented approach is a sufficient tool for assisting the inspector.