Condition Assessment of Overhead Conductors in Aerial Images Using Domain Specific Knowledge and Machine Learning

Abstract

Condition-based Risk Management (CBRM) is a methodology that incorporates current asset condition along with common engineering knowledge and practical experience to predict and manage future asset failure risk. CBRM is adopted widely in the electricity industry. To obtain asset condition, regular asset inspections are carried out. In this thesis, the aim is to automate the visual inspection of overhead electrical conductors from aerial images taken by an Unmanned Aerial Vehicle (UAV). Its output serves as the current asset condition for CBRM.

To achieve this goal, two main subtasks are involved. One is the identification/detection of the overhead conductors in aerial images, another is the condition assessment/recognition of the conductors. For each subtask, domain specific knowledge is exploited and leveraged to our advantage, and to make the automation more resilient to external influences. Conductor identification is automated by utilizing the fact that the target conductor possesses a unique strand winding characteristic. This differentiates it from other conductor types or false lines appearing in the inspection images. A high detection rate of at least 98.2% is achieved. Condition assessment of conductors is carried out mainly from three distinctive aspects: shape, texture, and colour. Image feature extraction is manually designed in accordance with the condition assessment criteria supplied by industry experts.

Throughout the research multiple real inspection image datasets were gathered by our industrial sponsor from their 11 kV electrical distribution network, including one large-scale UAV flight. The effectiveness of the automated visual inspection is validated using these datasets. The development of this automation encompasses great challenges in a real-world scenario, for instance, distractions of background elements, loss of image focus, colour calibration, and screening of conductor’s region of interest images. These challenges were addressed and studied.

The conventional multi-staged domain-driven approach in image processing is also put in comparison with the data-driven end-to-end supervised deep learning. The former approach preprocesses an input image as much as possible aiming to extract features that are thought to be critical to condition assessment. While the latter disregards this preprocessing almost completely. This comparison was done primarily using a laboratory dataset to exclude those issues encountered in a real-world scenario. It was found that the end-to-end model performs better, but only when each classification category is sufficiently represented. The domain-driven approach, on the other hand, is derived based on background subject knowledge, which enables accurate and reliable inferences without sufficiently large data observations. But its performance is generally worse, as a compromise.

This thesis contributes to the automated asset inspection literature, in particular, visual inspection of overhead conductors in an electrical distribution network. Notably, it predominantly explores a domain-driven approach. Parts of the research output has been put in practical use by our sponsoring utility company, and commercialisation is planned in the near future.