Transformer-based InspecNet for improved UAV surveillance of electrical infrastructure

Abstract

Surveillance is crucial for maintaining critical infrastructure integrity and disaster risk reduction. Unmanned Aerial Vehicles have emerged as vital tools for aerial inspections, offering flexibility, efficiency, and cost-effectiveness. A significant challenge in UAV surveillance is the precise detection of damaged electrical components, particularly in complex environments where numerous objects are in close proximity that may cause hazards. Traditional methods often fall short under these demanding conditions, leading to notable monitoring deficiencies. To address these challenges, we introduce a novel detection method utilizing the Transformer architecture, named InspecNet. This approach leverages the architecture’s proficiency in understanding contextual information, which significantly enhances the accuracy of identifying key damaged components: damaged ceramic insulators, burned ceramic insulators, and loose U-bolts. These components are particularly challenging to detect due to their subtle and variable damage signatures. Through extensive data augmentation, we have created a new and diverse sample set to train our model and improve its detection capabilities. Our experimental evaluation, conducted with an extended set of UAV image data, demonstrates a detection accuracy increase of 20 % over conventional methods, achieving a precision of 95.7 %, recall of 93.1 %, and a mean average precision (mAP) of 92.9 %. These results underscore InspecNet potential to deliver accurate and reliable infrastructure monitoring, setting a new standard in automated UAV surveillance technology to reduce hazards.