Detecção automática de defeitos em células fotovoltaicas através de redes neurais convolucionais

Abstract

The climate crisis observed in recent years, coupled with the depletion of fossil fuels, has made the adoption of alternative, sustainable and renewable energy sources imperative. In this scenario, solar energy has been employed as an energy source to meet the global demand for electric power. The observed increase in photovoltaic (PV) solar electricity generation brings demands for methods for inspecting and maintaining PV systems, as their components are subject to various types of defects such as microcracks, shading, open circuits, short circuits, overheating, among others. For the detection of these defects, some techniques have been developed such as the analysis of the I-V curve and inspection of thermal images or Electroluminescence (EL) of PV cells. The latter has been addressed in several works that seek to automate defect detection through computer vision techniques and machine learning. The present work follows this line of research, seeking to automate the detection of defects in monocrystalline PV cells through EL images. A three-step computational model is proposed here: pre-processing, feature extraction, and defect diagnosis. In pre-processing, a Bilateral Gaussian filter is applied to the images, to then obtain, during the feature extraction phase, a texture descriptor of the images by the Local Binary Pattern (LBP). Finally, the descriptions of the texture of the resulting images feed a Customized Convolutional Neural Network (CCNN), trained and tested in the task of classifying cells as defective or non-defective, in the defect diagnosis phase. The CCNN is the main contribution of the present work, being composed of a topology with a reduced number of parameters, when compared to other traditional networks. The performance obtained in cross-validation tests proved that the proposed convolutional network is competitive in relation to other approaches present in the literature. It was concluded that the model meets the needs of the problem of defect detection in PV cells, with accuracies of 94% and 86% in the classification of defective and non-defective cells, respectively.