Fault Diagnosis of PV Modules based on Convolution Neural Network and Out-of-distribution Detection

As one of the most clean energy sources, solar energy is playing an increasingly important role in energy generation, thus driving the rapid development of photovoltaic (PV) power plants. In one PV power plant, there may be hundreds of thousands PV modules. Fault diagnosis of such a huge number of PV modules is critical and challenging. Recently Unmanned Aerial Vehicles (UAVs) equipped with infrared cameras are taken to execute this fault diagnosis by taking photos and analyzing these photos for faults. To improve the fault diagnosis accuracy and efficiency of PV modules, this paper proposes an automatic PV module fault diagnosis algorithm based on deep learning for infrared images, which is made up of two steps, including localization and classification. In the localization step, we first design a lightweight convolution neural network (CNN) to detect edges of PV modules in infrared images; then a region extraction method is proposed to segment PV modules. In the classification step, a lightweight classifier is used to detect faulty PV modules. Moreover, we introduce an Out-Of-Distribution (OOD) detection algorithm to identify and eliminate false PV modules, which actually come from the background and are wrongly segmented in the localization step. The effectiveness of the proposed PV module fault diagnosis algorithm has been successfully verified on real images.