Work

ResViT-Solar: A Hybrid Residual-Transformer for Photovoltaic Panel Fault and Cleanliness Detection

Year: 2026

Type: article

Abstract: Solar photovoltaics (PV) play a central role in the global clean-energy transition, necessitating reliable and cost-effective inspection methods to ensure yield and safety at scale. This paper introduces ResViT-Solar, a compact hybrid architecture that integrates a frozen, ImageNet-pretrained ResNet50 stem for translation-equivariant local feature extraction with a lightweight Vision Transformer (ViT) head for global, content-adaptive reasoning. High-level convolutional feature maps are projected into a concise token sequence and refined through two pre-norm self-attention layers equipped with GELU-MLP blocks, facilitating globally consistent classifications using only image-level labels. The model was trained on a Kaggle dataset comprising six classes: Bird-drop, Clean, Dusty, Electrical-damage, Physical-damage, and Snow-covered. Experimental evaluations reveal superior performance, achieving 94% accuracy and outperforming baselines such as ResNet50 (89%), InceptionV3 (90%), and ViT (91%). Grad-CAM visualizations confirm the model’s attention to semantically relevant regions. Furthermore, validation on two external binary datasets, ELPV (93% accuracy) and Solar-Cell Cleanliness (95%), highlights its strong generalization capabilities. By harmonizing local saliency with long-range contextual understanding within constrained computational resources, ResViT-Solar advances accurate, explainable, and practical PV module monitoring for real-world operation and maintenance. The design emphasizes deployability through a shallow attention stack and concise tokenization, reducing training cost and memory footprint, while accommodating edge-oriented constraints by freezing the convolutional backbone and limiting transformer complexity.

Source: IEEE Access

DOI: 10.1109/access.2026.3661638

Authors: Nazrul Amin, Yong-Woon Kim, Chulung Kang, Yung-Cheol Byun

Institution: Jeju National University

Language: English

Topic: Industrial Vision Systems and Defect Detection

Subfield: Industrial and Manufacturing Engineering

Field: Engineering

Domain: Physical Sciences

SDGs: Affordable and clean energy, Industry, innovation and infrastructure

Open Access status: gold

Estimated APC paid: $2,075