Photovoltaic Cell Defect Classification from Luminescence Images: Embedding and Clustering with Unsupervised Machine Learning

The significant challenge of photovoltaic (PV) module end-of-life (EoL) management is receiving increasing attention from the broader PV community. However, currently there is a lack of adequate large-scale characterisation techniques and associated analytical methods for the quality assessment of fielded modules in utilityscale PV plants. This study aims to develop a methodology for (a) classifying defects and degradation using luminescence images of PV modules (obtained on-site or indoors), and (b) automated decision-making regarding the preferred EoL route. We achieve this through the development of a novel linear unsupervised machine learning (ML) methodology that provides robust and consistent cell defect/degradation classification capability. This approach involves three key stages: (a) a suite of algorithms for image feature extraction, (b) feature vector embedding with principal component analysis (PCA), and (c) k-means clustering for classification of individual cells by defect/degradation type and severity. For validation, module quality assessment is performed using the aggregate celllevel quality (identified defects or degradation). This automated quality assessment is then compared to a manual assessment performed by a representative domain expert evaluating the same images, and also against reference current-voltage measurements of each module, as well as a reference area-statistics-based method. The proposed MLbased assessment methodology achieves strong alignment with that of the expert assessment, including both the expected module performance and risk, and outperforms the reference method. Importantly, this automated quality assessment is able to perform accurately and consistently at a speed that is orders of magnitude faster than possible by manual assessment, enabling the low-cost utilisation of contactless imaging techniques for utility-scale PV system EoL monitoring and assessment.