Positive and Unlabelled Learning with Data Approximation for Photovoltaic Defect Detection

This paper explores a Positive and Unlabeled (PU) learning framework enhanced by data approximation via Deep Convolutional Generative Adversarial Networks (DCGAN) for defect detection in photovoltaic cells. Employing Electroluminescence (ELPV) imaging, the proposed methodology significantly reduces the need for labeled samples, utilizing only a small subset to achieve an accuracy of 0.805 and an F1 score of 0.870. This performance is comparable to traditional supervised learning methods that use substantially more data. The results of this work validate the potential of PU learning in scenarios with limited labeled data, paving the way for future studies to optimize and expand this approach.