TY - JOUR
T1 - Attention residual network with multi-scale convolution branch for efficient solar photovoltaic module defect classification
AU - Bamisile, Oluwatoyosi F.
AU - Kun, She
AU - Ukwuoma, Chiagoziem C.
AU - Thomas, Dara
AU - Ejiyi, Chukwuebuka J.
AU - Olagundoye, Omosalewa
AU - Olugbenle, Olatomide
AU - Olotu, Olamide
AU - Bamisile, Olusola
N1 - Publisher Copyright:
© 2026 International Solar Energy Society.
PY - 2026/1/13
Y1 - 2026/1/13
N2 - Solar photovoltaic (PV) systems are increasingly deployed worldwide, intensifying the need for efficient and accurate defect detection methods that ensure long-term performance. Infrared thermography is widely used for PV inspection, yet existing deep learning methods face difficulties detecting small-scale anomalies, handling class imbalance, and maintaining stable performance under real-world thermal variability. This study introduces an Attention Residual Network with Multi-Scale Convolution Branch to capture fine- and coarse-scale features while enhancing robustness and gradient stability. The model was tested on varying solar PV datasets, including the Infrared Solar Modules dataset under binary and multi-class settings and the PV panel defect dataset. The proposed model achieved 0.968 accuracy and 0.981 ROC-AUC (binary) and 0.971 accuracy and 0.993 ROC-AUC (multi-class) for the Infrared Solar Modules dataset, while recording an accuracy of 0.975 and 0.949 kappa (binary), 0.973 accuracy and 0.955 kappa (3 classes) and 0.915 and 0.8950 kappa (6 classes) on the PV panel defect dataset. Ablation studies on the Infrared Solar Modules dataset demonstrated the individual contributions of multi-scale extraction, attention refinement, and residual learning, while Grad-CAM visualisations confirmed the interpretability of defect localisation. The results show that the proposed model offers an accurate, stable, and interpretable approach for infrared-based PV defect classification, supporting scalable deployment in automated inspection systems.
AB - Solar photovoltaic (PV) systems are increasingly deployed worldwide, intensifying the need for efficient and accurate defect detection methods that ensure long-term performance. Infrared thermography is widely used for PV inspection, yet existing deep learning methods face difficulties detecting small-scale anomalies, handling class imbalance, and maintaining stable performance under real-world thermal variability. This study introduces an Attention Residual Network with Multi-Scale Convolution Branch to capture fine- and coarse-scale features while enhancing robustness and gradient stability. The model was tested on varying solar PV datasets, including the Infrared Solar Modules dataset under binary and multi-class settings and the PV panel defect dataset. The proposed model achieved 0.968 accuracy and 0.981 ROC-AUC (binary) and 0.971 accuracy and 0.993 ROC-AUC (multi-class) for the Infrared Solar Modules dataset, while recording an accuracy of 0.975 and 0.949 kappa (binary), 0.973 accuracy and 0.955 kappa (3 classes) and 0.915 and 0.8950 kappa (6 classes) on the PV panel defect dataset. Ablation studies on the Infrared Solar Modules dataset demonstrated the individual contributions of multi-scale extraction, attention refinement, and residual learning, while Grad-CAM visualisations confirmed the interpretability of defect localisation. The results show that the proposed model offers an accurate, stable, and interpretable approach for infrared-based PV defect classification, supporting scalable deployment in automated inspection systems.
KW - Attention mechanism
KW - Infrared thermography
KW - Multi-scale convolution
KW - Photovoltaic fault detection
UR - https://www.scopus.com/pages/publications/105027275179
U2 - 10.1016/j.solener.2026.114323
DO - 10.1016/j.solener.2026.114323
M3 - Article
AN - SCOPUS:105027275179
SN - 0038-092X
VL - 307
JO - Solar Energy
JF - Solar Energy
M1 - 114323
ER -
