Evaluation of CNN and Pre-trained Models in Image-Based PV Panel Defect Identification
International Symposium for Production Research, ISPR 2025, İstanbul, Türkiye, 9 - 11 Ekim 2025, ss.113-121, (Tam Metin Bildiri)
- Yayın Türü: Bildiri / Tam Metin Bildiri
- Doi Numarası: 10.1007/978-3-032-22784-3_9
- Basıldığı Şehir: İstanbul
- Basıldığı Ülke: Türkiye
- Sayfa Sayıları: ss.113-121
- Anahtar Kelimeler: convolutional neural network, defect detection, image classification, Photovoltaic panel
- İstanbul Üniversitesi-Cerrahpaşa Adresli: Evet
Özet
The rapid expansion of photovoltaic (PV) energy systems has made efficient and accurate defect detection essential for ensuring optimal performance and extending system lifespan. Manual inspection methods are often labor-intensive, time-consuming, and prone to human error, creating a strong need for automated solutions. This study presents a comparative evaluation of a custom-designed Vanilla Convolutional Neural Network (CNN) and four widely used pre-trained architectures ResNet50, VGG16, DenseNet201, and MobileNetV2 for classifying PV panel defects from image data. A labeled dataset containing six defect categories, including dust, bird droppings, electrical faults, physical damage, and snow coverage, was used to train and test the models. Each architecture was assessed in terms of classification accuracy, generalization capability, and computational efficiency. Results indicate that while pre-trained models generally achieve higher accuracy due to their deep feature extraction capabilities, the Vanilla CNN offers competitive performance with significantly lower computational requirements. These findings provide practical insights into balancing model complexity and deployment feasibility, supporting sustainable PV system maintenance.