YOLO-based deep learning architectures for photovoltaic fault localisation: A comprehensive review of architectures, datasets, and edge-optimised deployment strategies

The rapid growth of photovoltaic (PV) installations has increased the demand for accurate, efficient, and deployment-ready fault localisation systems. Deep learning-based object detection methods, particularly the You Only Look Once (YOLO) family, have become the dominant approach because of their high localisation accuracy and efficient inference. This review systematically analyses 130 peer-reviewed studies published between 2016 and 2025 using a PRISMA-guided methodology. The reviewed literature is organised into 12 technical categories covering baseline and lightweight YOLO models, transformer- and attention-enhanced architectures, feature-fusion methods, hybrid approaches, transfer learning, GAN-based augmentation, and deployment-oriented optimisation strategies. In addition, 17 benchmark datasets comprising electroluminescence (EL), infrared (IR), RGB, UAV-acquired, and synthetic imagery are reviewed, together with commonly used evaluation metrics and deployment settings. To improve the transparency of published results, the review provides comparative analyses of datasets, IoU thresholds, input resolutions, hardware platforms, and experimental protocols. Experimental benchmarking of YOLOv11n on a Raspberry Pi 5 further demonstrates that ONNX-based deployment improves inference efficiency, achieving a best-case latency of 153.01 ms and 6.54 FPS. Finally, the review discusses current challenges and future research directions for developing robust, lightweight, and edge-deployable photovoltaic fault localisation systems.