Edge AI Implementation and Model Optimisation for Real-Time Remote Inspection of Power Line Using UAV

Kang, Yan Sheng (2026) Edge AI Implementation and Model Optimisation for Real-Time Remote Inspection of Power Line Using UAV. Final Year Project (Bachelor), Tunku Abdul Rahman University of Management and Technology.

Text KANG YAN SHENG_Full Text.pdf Restricted to Registered users only Download (84MB)

Abstract

Automated inspection of power line insulators using unmanned aerial vehicles (UAVs) o!ers a safer and more cost-e!ective alternative to traditional manual methods, but deploying deep learning models for this task on embedded hardware is challenging because existing single-stage approaches train one model to jointly localise insulator strings and classify disc-level defects, driving input resolution upward and limiting throughput on energy-constrained edge devices. This thesis proposes and evaluates a two-stage YOLO pipeline with binary Shannon entropy gating, in which a YOLO11n Stage 1 model localises insulator strings at 640-pixel resolution, a binary Shannon entropy gate accepts high-confidence detections as defect-free while forwarding uncertain crops to a YOLO11n Stage 2 model for disc-level defect classification, and both models are deployed as FP16 ONNX engines with TensorRT acceleration on the NVIDIA Jetson Orin NX (16 GB). Evaluated against a PyTorch single-model baseline and an ONNX single-model baseline on the Insulator Defect Image Dataset (IDID), the proposed pipeline at threshold