SPFNet2: A lightweight solar panel fault detection framework using parallel U-Net and MobileNetV3Large

Identifying and determining defects in solar panels is necessary to ensure the efficiency and reliability of renewable energy systems. This research introduces SPFNet2, a novel lightweight deep learning model that integrates advanced segmentation and classification techniques to address traditional fault detection limitations. The model combines MobileNetV3Large with a custom convolutional pathway and an attention mechanism, enhancing accuracy in identifying and isolating faults such as physical damage, dust accumulation, and bird droppings. The model effectively addresses dataset imbalances by leveraging advanced preprocessing methods, including Binary masks and Gaussian noise. Segmentation is achieved using a dual-path architecture with up-sampling layers, recording outstanding metrics, including a Dice coefficient of 0.97 and an intersection over union (IoU) of 0.95. SPFNet2 demonstrated exceptional performance, achieving an overall accuracy of over 96%, while training and validation accuracies reached 99.82% and 97.65%, respectively, outperforming existing models. Furthermore, the proposed model attained F1-score and Precision–Recall values of 99% and 98%, respectively, highlighting its superior effectiveness.