Photovoltaic systems play a central role in the worldwide transition to renewable energy sources, yet their efficiency and safety are frequently compromised by defects introduced during manufacturing or arising during their operation. Irregularities ranging from structural discontinuities to electrical anomalies can lead to considerable power losses and the development of localized hotspots. Electroluminescence (EL) imaging offers high-resolution, non-invasive visualization of cell-level defects, but manual inspection remains labor-intensive, subjective, and unscalable for large solar farms or remote installations. Recent advances in lightweight deep learning models, particularly the You Only Look Once (YOLO) family, enable automated, real-time detection. Nevertheless, achieving both strong detection performance and minimal computational demand on resource-constrained edge devices remains difficult, particularly when working with imbalanced, long-tailed distributions such as those found in the PVEL-AD dataset.
This feasibility study assesses two compact variants from the YOLO series, YOLOv8n, which serves as the established reference, and YOLOv26n, the most recent edge-focused model featuring a native end-to-end design without non-maximum suppression (NMS). Both models are trained and evaluated using a representative subset of the PVEL-AD dataset, which comprise 4500 electroluminescence (EL) images (3600 for training and 900 for validation) containing 7799 annotated defect instances distributed across eight key defect types: black_core, crack, finger, thick_line, star_crack, horizontal_dislocation, vertical_dislocation, and short_circuit. The dataset displays a pronounced long-tail distribution, where finger and crack defects constitute more than half of all instances. Training was performed with conservative settings to simulate resource-constrained experimentation: 512×512 input resolution, 50 epochs, early stopping (patience=10), minimal augmentation. And the models were trained and inferenced on a Tesla P100 GPU using the Ultralytics framework.
The evaluation adopts standard object detection metrics: mean Average Precision at IoU thresholds of 0.5 (mAP@0.5) and across 0.5 to 0.95 (mAP@0.5:0.95), along with precision, recall, and inference latency (encompassing preprocessing, inference, and postprocessing time in milliseconds per image, as well as frames per second). All measurements are derived from the 900-image validation subset, which includes 1586 defect instances.
On this validation set, YOLOv26n attains an overall mAP@0.5 of 0.897 and mAP@0.5:0.95 of 0.654, accompanied by a precision of 0.832 and recall of 0.891. In comparison, YOLOv8n yields modestly higher accuracy, with mAP@0.5 reaching 0.926, mAP@0.5:0.95 at 0.671, precision at 0.915, and recall at 0.889. Both architectures demonstrate excellent performance on larger or more structurally distinct defect categories (AP@0.5 exceeding 0.96 for black_core, short_circuit, and both dislocation types). Detection remains more moderate, however, on particularly difficult classes such as crack (0.687–0.737) and star_crack (0.768–0.771).
Inference speed markedly favors YOLOv26n, which records an average total latency of approximately 2.5 ms per image (corresponding to roughly 397 FPS), in contrast to approximately 3.3 ms (~302 FPS) for YOLOv8n. This advantage stems largely from the model's NMS-free architecture, which eliminates the computational overhead associated with traditional post-processing steps.
This preliminary evaluation demonstrates that lightweight YOLO models, particularly YOLOv26n, offer a compelling trade-off between detection performance and inference speed for PV defect monitoring using EL imagery. Despite training on a subset with minimal augmentation, the models achieve competitive mAP@0.5 values (0.897–0.926) and sub-10 ms latency, making them suitable for real-time deployment on edge devices in industrial or remote solar environments. The strong performance on dominant defect types and reasonable results on rarer classes highlight their practical potential. Future work will extend this study to the full PVEL-AD dataset, incorporate additional architectures, stronger augmentation, and quantization for ultra-low-power settings.
