Transfer Learning-Based Fault Detection in Solar Panels Using Pretrained DenseNet121 DenseNet169 and DenseNet201Architectures

Abstract

Physical and electrical residue, dust, and other foreign contaminants accumulated on the surfaces of solar panels negatively impact the efficiency of solar modules and the amount of energy directly produced. At the same time, solar energy is a natural resource that is becoming increasingly important globally for sustainable energy production. Therefore, early and accurate detection of faults that may occur in solar panels is crucial for the continuity of energy production. Furthermore, timely monitoring and cleaning of solar panel surfaces with the right techniques is also ciritical for increasing the efficiency of these modules. Traditional observational or sensor-based methods for monitoring, cleaning, troubleshooting, and maintaining solar panels exhibit limited performance due to their high cost and vulnerability to human error. Thus, this study proposed an innovative transfer learning-based autonomous deep learning (DL) approach to detect faults from solar panel images. The proposed system utilized a publicly available solar system image dataset consisting of six classes. After completing the data cleaning and preprocessing steps, feature extraction was performed using three different pre-trained DenseNet121, DenseNet169, and DenseNet201 transfer learning architectures. Six different artificial intelligence (AI) based classification algorithms were executed to perform predictions on the resulting feature maps. The performance of the proposed innovative DL-based system was evaluated using quality metrics such as accuracy, precision, recall, AUC, and F₁-score. The experimental results demonstrate that the highest accuracy rate of 88.14% was achieved with the DenseNet201+Logistic Regression (LR) hybrid model. Other results obtained in this proposed study were explained in detail and compared using tables and graphs. The findings demonstrate that AI-assisted DL and transfer learning-based approaches offer effective, fast, and low-cost solutions for solar panel monitoring and maintenance processes.