Model Optimization for Multi-Class Real Time UAV Thermal Anomaly Detection in Solar PV Systems

This study investigates an intelligent UAV-based thermal inspection framework for photovoltaic (PV) modules, aiming to improve fault detection accuracy while ensuring suitability for real-time aerial deployment. Utilizing deep learning techniques, the system evaluates and compares multiple Convolutional Neural Network (CNN) architectures and YOLO classifiers for multi-class thermal anomaly detection. A comprehensive dataset of 20,000 aerial infrared images, categorized into nine distinct fault classes, was used to assess each model’s classification performance and computational efficiency. CNNs such as EfficientNet and ResNet50 achieved strong classification results, with EfficientNet reaching an accuracy of 84%. However, their high inference latency and large model sizes pose challenges for onboard processing in UAV-based systems, particularly under power and memory constraints. Alternatively, the YOLO family of models demonstrated a compelling balance of speed, compactness, and accuracy. Among them, YOLOv8n offered exceptional runtime performance with an inference time of just 8 ms and a model size of 2.94 MB, making it ideal for lightweight deployment. At the same time, YOLOv8l achieved a Top-1 accuracy of 83.02% after hyperparameter tuning. The study highlights how YOLO classifiers can outperform traditional CNNs in practical UAV deployment scenarios, maintaining high accuracy while significantly reducing computational costs. The findings underscore YOLO's viability as a scalable and efficient solution for real-time PV anomaly detection, supporting the main goal of autonomous and cost-effective solar farm maintenance.