Development and Evaluation of an Edge AI System for Real-Time UAV-Based Photovoltaic Thermal Inspection

Defect detection in photovoltaic installations increasingly relies on UAV thermal imagery followed by offline post-processing, which delays critical maintenance decisions. Real-time on-board inspection resolves this latency but is strictly bounded by the edge hardware's computational and power capacities.

To address this, we developed a complete edge AI system for real-time UAV-based photovoltaic inspection integrating an NVIDIA Jetson Orin NX module on a DJI M300 RTK drone. Its pipeline includes preprocessing of thermal imagery, YOLOv12 inference which was fine-tuned on a dataset of distinct thermal anomalies (e.g., hotspots, bypassed diodes) and Non-Maximum Suppression (NMS) post-processing. While YOLOv12 achieves real-time speeds on desktop GPUs, its raw execution on edge devices introduces critical latency bottlenecks. To maximize throughput within the hardware limits, TensorRT Post-Training Quantization (PTQ) was systematically evaluated across FP32, FP16, and INT8 formats.

The results indicate the system's energy viability, recording a maximum power consumption of 14.4 W and a temperature of 42.8 °C. The INT8 format (42.73 FPS) underperformed compared to FP16 (44.19 FPS) in the inference. This is caused by the quantization/dequantization (Q/DQ) node execution. The lack of native INT8 kernels for YOLOv12's advanced linear attention mechanisms forces a fallback from Tensor Cores to the CPU. Furthermore, INT8 quantization is followed by a drop in the mean average precision (mAP) across the dataset compared to FP16 (2.06%).

Results validate the FP16 format as the optimal deployment configuration for this edge AI system. Achieving an end-to-end processing latency of 24.80 ms, the system successfully processes sequential thermal image frames at 30 Hz with zero frame drops, enabling reliable and immediate in-flight diagnostics.