Convolution Neural Network (CNN) Architectures Analysis for Photovoltaic (PV) Module Defect Images Classification

Abstract

Photovoltaic (PV) module is the medium to convert solar energy to electrical energy. The existence of defects in the PV module will affect the system's efficiency to generate electricity. In this work, Convolutional Neural Network (CNN) architecture is proposed in PV module defect image classification due to its capability to extract patterns and consist of several unique layers to classify images accurately. Thus, the objective of this paper is to analyze PV defect image classification using CNN architectures namely Residual Neural Network (ResNet) and Visual Geometry Group (VGG), and to identify which architectures give the best performance. In this paper, two types of ResNet architectures which are ResNet-18 and ResNet-50, and two types of VGG architectures which are VGG-16 and VGG-19 were applied. For validation and verification purposes, other performance metrics such as F1-score, sensitivity, specificity, and precision are evaluated. The result shows that VGG-19 is outperformed in terms of accuracy with a value of 92.58%, sensitivity (73.14%), and specificity (95.10%) compared to ResNet-18, ResNet-50, and VGG-16 architecture.