Computational method for automation in the detection of defects in photovoltaic cells using electroluminescence images

Abstract

The generation of photovoltaic (PV) energy has been widely adopted on a global scale due to its clean nature and affordable cost. A crucial issue in PV generation systems is the integrity of PVs cells, which often operate under challenging environmental conditions such as wind, rain, salinity, and dust. Thus, there is a growing demand for technologies aimed at improving the efficiency and reliability of these systems, especially through the automation of the defect detection process in PV cells. In this context, a computational model has been proposed to distinguish defective monocrystalline silicon PV cells based on texture attributes extracted from electroluminescent (EL) images. Additionally, the model incorporates a Custom Convolutional Neural Network (CNN), which features a reduced number of parameters compared to other conventional neural network topologies. The model’s performance was evaluated through cross-validation using a widely recognized dataset of PV cell images in the literature. The parameter reduction implemented in the Custom CNN did not compromise the model’s performance, demonstrating consistent results with those reported in the literature. Therefore, the proposed model is a viable solution to the problem of defect detection in PV cells, capable of correctly identifying 94% of defective cells and 91% of non-defective ones.