Abstract
Unmanned aerial vehicle (UAV) thermal infrared remote sensing can rapidly acquire thermal-response information for large-scale photovoltaic (PV) arrays. Nevertheless, local thermal anomalies occupy small regions within aerial images; their features tend to degrade during successive down-sampling. Moreover, model deployment on edge devices is constrained by limited computational and memory resources. To address these challenges, this paper proposes TP-YOLO, a photovoltaic anomaly-detection approach, and implements it on the Zynq-7020 hardware platform. First, a three-channel thermal-prior input consisting of luminance distribution, local luminance difference, and a luminance gradient is constructed from pseudo-color infrared images, and explicitly characterizes thermal anomalies without requiring absolute-temperature measurements. Next, the detection scales of YOLOv11n are adjusted to P2/P3/P4. Modules including A-C3k2, Conv + Add, and TLCB are adopted to optimize feature extraction and multi-scale fusion, preserving local anomaly features while restraining model complexity. On the hardware side, a PS/PL co-processing architecture is employed. A dedicated convolution IP, ADD IP, and Window IP are designed for convolution, element-wise fusion, and local-window operations, respectively. INT8 quantization and ping-pong buffering are further utilized to realize network deployment. On the PV HSD test set, TP-YOLO achieves an mAP@0.5 of 77.5% and an mAP@0.5:0.95 of 33.7%, with 1.91 M parameters and 6.2 GFLOPs computational cost. In cross-dataset evaluation on the Drone Dataset without fine-tuning, the mAP@0.5 reaches 64.4%. After deploying the INT8 quantized model onto Zynq-7020, the implemented system yields an mAP@0.5 of 76.8%, achieving single-frame inference latency of 162 ms and throughput of 6.17 FPS at 150 MHz, with a power consumption of 3.4 W.