Multi-stage deep learning networks for automated assessment of electricity transmission infrastructure using fly-by images

Abstract

The modern electric power transmission system is a geographically extensive network which can span hundreds of kilometres, crossing harsh terrain and making manual inspection of its components costly. This paper proposes a novel condition assessment methodology for transmission overhead lines that is significantly more cost effective compared to traditional foot-patrol visual inspections. The proposed methodology utilizes a multi-stage ensemble deep learning network to automatically classify tower conditions based on high resolution aerial images. While aerial inspections allow relatively quick inspection of transmission routes, they are not usually used for condition assessment due to the high capturing altitude of images which would require time consuming manual processing to identify defects on hundreds to thousands of images. The proposed methodology automatically isolates transmission poles, disaggregates components, detects defects and determines the health index of concrete structures and insulators. The method involves pushing images through layers of automatic detection, region of interest (RoI) extraction and patching. A number of recent object detectors were tested on real-world data to evaluate their performance and an ensemble model is composed to improve the reliability of the detection. Results indicate that the multi-layer approach with output-based ensemble modelling, can effectively detect critical defects, although incipient fault conditions remain uncertain.