Efficient Lightweight Networks for Solar Panel Fault Classification Using EL and RGB Imagery

Accurately detecting faults in photovoltaic (PV) modules is essential for ensuring the reliability and efficiency of solar energy systems. Existing image-based deep learning methods often face challenges such as prolonged training times, real-time requirements, and limited accuracy when dealing with environmental factors such as dust, snow, or bird droppings. A new module called fast Fourier transform (FFT)-power coefficient (PC) is proposed to address these limitations, which improves electroluminescence (EL) imagery accuracy by the implicit enhancement effect of PCs. Second, a new slant fast orthogonal transformation module (SFOT) is proposed tailored for red, green, blue (RGB) PV images with slanted edges. Third, the spatial power coefficient module (S-PC) is designed to enhance features and edges in the spatial domain. The proposed modules combined the strengths of mobile visual transformer (ViT) architecture, and two new networks, MobileFFT and MobileSFOT, were designed. It is shown in this article that traditional transformer and attention blocks can be effectively replaced by achieving approximately 4× fewer GFLOPs (1.26), $\sim 2.5\times $ fewer parameters (1.45 M), reduced latency in seconds, required ~20% less GPU memory than the standard attention block, and 33% less energy consumption (EN), enabling real-time performance on central processing unit (CPU)-based devices. Tested on benchmark datasets of RGB and EL images, the method outperforms recent state-of-the-art (SOTA) models across several key metrics. Specifically, it achieves an accuracy of 0.874 and an ${F}1$ score of 0.858 for RGB images. This advancement makes the proposed approach highly suitable for monitoring applications in solar energy systems.