Fault Detection in Solar Panels with a Lightweight Differential Convolutional Network and One-Class Classification

Solar energy capacity in the world has been increasing in recent decades, and it still has an uptrend. In order to create a sustainable energy source, the detection of faults in the photovoltaic panels is of utmost importance. Related studies in this emerging area heavily depend on the pre-trained models, which are usually very large. In this study, a lightweight model that has 1.4 million learnable parameters is utilized to detect faulty solar panels from infrared images. The proposed model consists of two differential convolution layers, followed by fully connected layers and a discriminative loss function designed for binary classification. The differential convolution tracks the differences between the pixels, which is an important feature as the faults may occur at different positions and in different shapes. Several discriminative loss functions were considered in the experiments and one-class Softmax approaches were found to be more effective than the others. One-class classification aims to create a margin between the no-fault data and faulty data while avoiding overfit to the training samples. Experiments showed that the proposed model can achieve 93.68% accuracy without any data augmentation. With online augmentation using simple geometric transformations, 94.98% accuracy was observed. Other than the very large models and more advanced augmentation methods, the proposed approach has one of the best accuracies in the related literature. Compared to a Vision Transformer with 42.6 million learnable parameters, the proposed model (1.4 million parameters) reduced the parameter size 96.71%, while the relative performance degradation was only 3.31%.