SolarFaultNet: CNN-based photovoltaic fault classification via channel-wise attention mechanisms

The growing deployment of solar energy systems requires precise fault detection in photovoltaic (PV) modules. Conventional detection methods—including current–voltage curve analysis and electroluminescence imaging—while effective for specific applications, exhibit limitations in cost-effectiveness and deployment scale. Infrared (IR) thermography provides an efficient alternative, but challenges persist in analyzing thermal images under variable conditions. This paper presents SolarFaultNet, a deep learning framework implementing channel-wise attention mechanisms specifically designed for PV fault classification. Our approach incorporates systematic data augmentation techniques (rotation, scaling, flipping) to enhance model generalization, resulting in 99% classification accuracy on thermal/infrared imagery. Quantitative comparison with established convolutional neural network architectures demonstrates SolarFaultNet’s superior performance metrics in distinguishing critical faults such as micro-cracks, hotspots, and bypass diode failures. The model achieves 13% improvement in classification performance on previously unseen test samples through implemented augmentation methods. These results validate the effectiveness of channel-wise attention for feature selection and fault discrimination in PV monitoring applications.