Crack Detection in Photovoltaic Modules Using Convolutional Neural Network Optimized with Komodo Mlipir Algorithm

Abstract

The growing deployment of photovoltaic (PV) systems underscores the need for robust fault detection to maintain efficiency and reliability. This study introduces an automatic crack classification framework using electroluminescence (EL) images of solar panels. A Convolutional Neural Network (CNN) is employed for feature extraction, while the Komodo Mlipir Algorithm (KMA), a metaheuristic optimizer, is integrated to optimize classifier weights and improve classification of defect severity levels. Evaluation on the ELPV Dataset comprising monocrystalline and polycrystalline PV modules categorized into four classes (Defective Confident, Defective Non-confident, Functional Non-confident, Functional Confident) demonstrates that the baseline CNN achieves 76.41% validation accuracy and an F1-score of 61.09%, but suffers from overfitting and poor minority-class performance. By contrast, the proposed CNN-KMA model achieves 82.04% validation accuracy and a 71.35% F1score, yielding more balanced performance across categories. Grad-CAM visualization enhances interpretability by highlighting critical defect regions, while a Gradio-based user interface enables real-time inspection. These findings indicate that the CNN-KMA framework delivers improved robustness, interpretability, and practical value compared to a standalone CNN, thereby supporting predictive maintenance and extending PV system lifespan.

Access to Document

Other files and links

• Link to publication in Scopus

Fingerprint