Enhancing the robustness of solar photovoltaic fault classification based on depthwise separable group equivariant CNNs

Abstract

Solar photovoltaic (PV) power generation has experienced significant growth and thermal infrared (IR) imaging via unmanned aerial vehicles (UAVs) has become an efficient method for inspecting large-scale PV plants. However, variations in UAV flight paths and weather conditions cause orientation changes, luminance variations, and increased noise, challenging the robustness of fault classification models. This study introduces p4(m) depthwise separable group equivariant convolution module to address these challenges. The proposed models offer advantages in terms of model size, parameter count, fault classification performance, and robustness for solar PV panel images. Without data augmentation, the proposed model achieves 84.0% accuracy for the 12-Class task and 75.0% for the 11-Class task on the Infrared Solar Module dataset. Compared to data augmentation-based methods, the proposed model shows 1.7% higher accuracy in the 12-Class task and 4.2% in the 11-Class task. Additionally, the proposed model achieves a 7.3% improvement over non-augmented ensemble models in the 12-Class task, while maintaining model size and parameters below 20% of baseline models. Robustness evaluation reveals significant accuracy improvements under real-world image transformations: 9.25% for rotational changes, 8.66% for luminance variations, and 28.37% for noise interference. These results demonstrate the model’s effectiveness in handling challenging conditions while maintaining computational efficiency.