A lightweight deep learning model for fault detection of PV modules using thermal images

• HDUNet, a new CNN architecture combining UNet, DenseNet, HRNet for fault detection. • Achieves state-of-the-art accuracy with a 92.25 % parameter reduction. • Suitable for low-power hardware requiring only 0.5GFLOPs for inference. • Supports rapid retraining with continuous field data. • Python implementation released in a public GitHub repository. Thermal imaging is an emerging and valuable tool in evaluating and inspecting photovoltaic modules, allowing the identification of anomalies invisible to the human eye, such as bypass diode failures, internal cell defects, and hot or cold spots. In this context, artificial intelligence and computer vision facilitate the analysis and interpretation of images. Nevertheless, the large number of parameters associated with this approach significantly affects the training and inference time of models, limiting its practical applicability for real-time fault detection. Therefore, we proposed a CNN architecture called HDUnet, which combines and optimizes key design principles from UNet, DenseNet, and HRNet for object detection, segmentation, feature extraction, and pattern recognition. The architecture comprises 1 million parameters and ensures continuous data feedback from field observations, reducing training time and requiring approximately 0.3 GFLOPs per inference. For binary fault classification tasks, HDUNet achieves performance comparable to the top-performing models with 13.9 million parameters and improves accuracy by 3.77 % compared to the best-reported lightweight model with 1.5 million parameters. In contrast, for the 12-class fault detection task, HDUNet approaches state-of-the-art precision while outperforming the best lightweight model by 6.5 %. These findings highlight the model’s capability to operate as a lightweight model with a remarkable trade-off between fault detection classification accuracy and computational efficiency.