Development of a Convolutional Neural Network Architecture for Production Based Photovoltaic Fault Detection

This paper presents a framework for the detection of cracked Photovoltaic (PV) cells within the production environment through the creation of a Convolutional Neural Network (CNN). The paper demonstrates how a simple CNN using certain data augmentation techniques and specific regularization i.e., batch normalisation is efficient for PV based crack detection, achieving a recall rate of 99.2% and F1-score of 97.4%. We validate each methodological iteration via KFold cross validation providing granular metrical details allowing us to further optimize the model recall metric to suit the needs of our end deployment environment. We understand and appreciate the lack of highly quality PV image data of defect and therefore through our research we propose specific data augmentations for scaling and injecting variance into PV datasets. The augmentations do not provide any artificial propping of the model performance but are rather cases that may be found on the production line for PV cell manufacturing.