Detection of defects in photovoltaic cell electroluminescence images using deep learning methods

Photovoltaic (PV) modules, one of the most important components of solar energy systems, are faced with different failures during the production, assembly and operation stages. Some of these failures seriously shorten the life of the modules and cause significant power losses. Therefore, effective detection of PV module defects is of great importance. Electroluminescence (EL) imaging is a non destructive imaging method in which a special-type camera captures the radiation produced by a current-fed solar module. This method can be very effective in detecting defects such as micro cracks, finger interruptions and disconnected cells that are not visible to the naked eye. However, interpretation of EL images is a matter of expertise, and given the large number of PV modules consisting of multiple cells, this process could be time-consuming and costly. In this thesis, in order to produce effective solutions to this problem, a series of deep learning-based classifier models were developed by working with a benchmark dataset in the literature in the first stage and classification was performed with the transfer learning method. In the second stage, data augmentation methods were investigated, and synthetic data were generated to increase the classification performance of the developed models. In the next stage, to see the generalization capacity of the created models and to expand the dataset, an additional EL image dataset was created by processing and labeling the EL images received from the manufacturer. High performance was achieved in the multi-class dataset using the developed customized network architectures. Finally, a fully automatic PV module test software was developed using the obtained models. In this study, both deep feature-based classifiers and a Convolutional Neural Network architecture were created from scratch and as a result, high performance classifier models were obtained. An artificial intelligence-based automatic PV module test software that can work in real time was developed. This decision support system can contribute to the acceleration and automation of quality control processes in production facilities by detecting PV cell defects quickly and effectively. As a result of the studies, in the ELPV data set, 90.42% and 94.52% performance were achieved with deep feature-based methods for 4 classes and 2 classes, respectively. In the part where data augmentation methods were investigated, 94.11% accuracy was attained with the created datasets, and 4.41% accuracy and 17.83% recall improvement were achieved compared to the original dataset. In the studies conducted with the created multi-class custom dataset, defective cells were classified with 98.23% and 98.71% accuracy by the developed CNN and ViT based systems, respectively.