Rede neural convolucional híbrida otimizada por algoritmos genéticos para classificação em imagens eletroluminescentes de células fotovoltaicas

• Ano de defesa: 2024
• Autor(a) principal: Rocha, Alan Marques da
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Palavras-chave em Português: Detecção automática de defeitos; Rede neural convolucional híbrida; Célula fotovoltaica; Ajuste fino de hiperparâmetros; Algoritmos genéticos evolutivos; Automatic defect detection; Hybrid Convolutional Neural Network; Photovoltaic cell; Fine-tuning of hyperparameters; Evolutionary Genetic Algorithms; CNPQ::ENGENHARIAS
• Link de acesso: http://repositorio.ufc.br/handle/riufc/78461

The global demand for electricity has been increasing exponentially in recent years. In light of this, investments in renewable energy sources have become increasingly necessary, with photovoltaic (PV) solar energy standing out as a source that is easy to install and cost-effective. The growth of the installed capacity of PV generation sources brings a demand for sophisticated and precise methods to detect defects in the cells that make up such systems. This work proposes a method for classifying monocrystalline silicon (Si-m) and polycrystalline silicon (Si-p) PV cells into functional and non-functional categories, using a Hybrid Convolutional Neural Network (HCNN) based on ResNet50 and VGG16 architectures, pre-trained with the ImageNet database to extract image features, where the best hyperparameters for each network were obtained through Evolutionary Genetic Algorithms (EGA). The classification process of the HCNN was conducted using a Support Vector Machine (SVM). Four classification experiments were performed. Initially, elementary algorithms such as SVM, Naïve Bayes (NB), k-Nearest Neighbors (k-NN), and Random Forest (RF) were tested. Subsequently, experiments were conducted with the ResNet50, VGG16, and InceptionV3 architectures. The HCNN models ResNet50+SVM and VGG16+SVM were tested with the original dataset containing 2.624 samples and an augmented dataset containing 13.120 images. The fine-tuning using EGA without data augmentation resulted in the VGG16+SVM topology achieving an accuracy of 95.21% and a Kappa index of 78.23%. Finally, as the main result of this work, the same HCNN topology surpassed its previous performance, achieving an accuracy of 99.67% and a Kappa index of 80.17% with data augmentation. The ResNet50+SVM model also showed robust results with data augmentation, achieving an accuracy of 98.17% and a Kappa index of 85.26%. These results highlight the effectiveness of the proposed techniques, positioning the HCNN optimized by EGA and data augmentation as a promising solution for the automatic detection of defects in Si-m and Si-p PV cells.