Automatic Fault Classification in Photovoltaic Modules Using Denoising Diffusion Probabilistic Model, Generative Adversarial Networks, and Convolutional Neural Networks

Current techniques for fault analysis in photovoltaic (PV) systems plants involve either electrical performance measurements or image processing, as well as line infrared thermography for visual inspection. Deep convolutional neural networks (CNNs) are machine learning algorithms that perform tasks involving images, such as image classification and object recognition. However, to train a model effectively to recognize different patterns, it is crucial to have a sufficiently balanced dataset. Unfortunately, this is not always feasible owing to the limited availability of publicly accessible datasets for PV thermographic data and the unequal distribution of different faults in real-world systems. In this study, three data augmentation techniques—geometric transformations (GTs), generative adversarial networks (GANs), and the denoising diffusion probabilistic model (DDPM)—were combined with a CNN to classify faults in PV modules through thermographic images and identify the type of fault in 11 different classes (i.e., soiling, shadowing, and diode). Through the cross-validation method, the main results found with the Wasserstein GAN (WGAN) and DDPM networks combined with the CNN for anomaly classification achieved testing accuracies of 86.98% and 89.83%, respectively. These results demonstrate the effectiveness of both networks for accurately classifying anomalies in the dataset. The results corroborate the use of the diffusion model as a PV data augmentation technique when compared with other methods such as GANs and GTs.