AI system spots hidden damage in solar panels with 99.8% accuracy
Researchers deployed a machine-learning system that identifies cracks, discoloration, and other defects in photovoltaic modules by analyzing photos alone. The breakthrough could slash inspection costs for solar farms and manufacturers while catching failures before they tank energy output.
Originaltitel: Intelligent fault detection in photovoltaic modules using attention-based deep learning network
Solcellsanläggningar förlorar drift- och intäktskapacitet genom fel som inte upptäcks i tid. En ny modell baserad på vision transformer-teknik (ViT) möjliggör automatisk och realtidskontroll av solcellsmoduler med 99,84 procent klassificeringsnoggrannhet. Forskare från Luleå tekniska universitet, NTNU och Vellore Institute of Technology tränade ViT-modellen på ett bildbibliotek med sex vanliga modulfel: glasbrott, missfärgning, brännskador, snigelslingor, delaminering och intakta paneler. Modellen överträffade jämförbara metoder i litteraturen. Systemet kan integreras i befintliga inspektionssystem för autonoma provningar. För solcellsägare innebär detta kortare avkastningsförsämring genom snabbare feldetektering och underhållsplanering. För stora portföljägare blir driftövervakningen mer kostnadseffektiv än manuell inspektion, särskilt när anläggningar är geografiskt utspridda.
<p>Photovoltaic (PV) systems experience various faults due to environmental conditions, human errors, and equipment failure during their service life. To necessitate maximum power generation and ensure ideal operating conditions, the development of intelligent fault diagnosis models is essential. In the present study, an attention-based deep learning network, namely, vision transformer (ViT), is adopted to automatically detect the visual faults, such as glass breakage, discoloration, burn marks, snail trail, good panel, and delamination on PV modules. An image dataset has been developed with the true color images of various faulty PV modules. The ViT model was fine-tuned and trained over the custom dataset created. The trained ViT model demonstrates a superior classification accuracy of 99.84% for fault detection and classification in PV modules. The obtained classification results of the model are compared with several other classification results reported in the literature. The ViT model could potentially be integrated into existing inspection systems for autonomous, real-time, efficient, and robust condition monitoring of PV modules.</p>