Accelerated and reliable diagnostics of photovoltaic modules with neighborhood attention CNN

Abstract This paper presents a neighborhood attention convolutional neural network (NACNN) to enhance the reliability and speed of diagnostics in photovoltaic (PV) modules. By integrating a neighborhood attention mechanism into the convolutional layers, our model can dynamically emphasize local regions of interest within infrared (IR) images captured by drones, enabling more precise detection of defects. Unlike global self-attention, the neighborhood attention mechanism focuses on local context, allowing the model to selectively attend to defect-prone areas while maintaining computational efficiency, closely mimicking the behavior of expert human inspectors. Experimental results on the Infrared Solar Modules dataset show that NACNN significantly outperforms conventional CNNs in both accuracy and processing speed. Additionally, Grad-CAM visualizations confirm that NACNN reliably localizes defect regions, improving the interpretability and trustworthiness of predictions. The proposed architecture offers a scalable, fast, and robust solution for real-time maintenance of large-scale PV systems.