Photovoltaic module fault detection method using infrared images based on a lightweight hybrid network

With the rapid expansion of photovoltaic (PV) power plants, efficiently identifying faults in PV modules is crucial for ensuring their safe operation and sustainable development. However, many existing deep learning methods struggle to jointly model local and global features and often involve high computational complexity. To address this challenge, an infrared image fault detection method for PV modules based on a lightweight hybrid network is proposed in this study. First, a convolutional neural network (CNN)-Vision Transformer (ViT) hybrid block is designed to combine depthwise separable convolution and ViT to achieve joint modeling of local features and global context. Second, a hierarchical feature enhancement architecture is constructed based on the MobileNetV3-Small backbone, in which CNN-ViT hybrid blocks and the spatial attention mechanism (SAM) are embedded at different stages to enable the model to focus more effectively on fault-relevant regions. Finally, the proposed lightweight hybrid network is incorporated into the U-Net framework as the encoder to mitigate interference from complex backgrounds, while the same lightweight hybrid network is employed for fault detection on the segmented images. Experimental results demonstrate that the improved U-Net model maintains segmentation accuracy while reducing the parameter count to 14.92% of the standard U-Net, and the fault detection model achieves an F1-score of 98.57% using only 1.84 M parameters. Overall, the proposed method achieves competitive performance while maintaining low computational complexity.