Hier finden Sie wissenschaftliche Publikationen aus den Fraunhofer-Instituten.

Efficient Deployment of Deep Neural Networks for Quality Inspection of Solar Cells using Smart Labeling

: Kunze, P.; Greulich, J.; Rein, S.; Ramspeck, K.; Hemsendorf, M.; Vetter, A.; Demant, M.

Volltext urn:nbn:de:0011-n-6181703 (1.7 MByte PDF)
MD5 Fingerprint: 5f17ab3ef7a743bbdba18490078b887f
Erstellt am: 11.12.2020

Pearsall, Nicola (ed.):
37th European Photovoltaic Solar Energy Conference and Exhibition, EU PVSEC 2020 : 07-11 September 2020, Online Conference
München: WIP, 2020
ISBN: 3-936338-73-6
European Photovoltaic Solar Energy Conference and Exhibition (EU PVSEC) <37, 2020, Online>
Konferenzbeitrag, Elektronische Publikation
Fraunhofer ISE ()
Photovoltaik; deep learning; defect detection; defects; luminescence imaging; solar cell; Silicium-Photovoltaik; Charakterisierung von Prozess- und Silicium-Materialien; Messtechnik und Produktionskontrolle

Luminescence images of solar cells show material- and process-related defects in solar cells, which are relevant for monitoring, optimization and processing. Convolutional neural networks (CNNs) allow the reliable segmentation of these defects in images of the solar cells. Nevertheless, the training of CNNs requires a large amount of empirical data, in which the defects have to be labeled expensively by experts. We introduce a method allowing efficient training by using Smart Labels. We show how this technique can be used for process monitoring to detect systematic errors. This approach differs from previous methods, which rely on human heuristics in the form of feature engineering or learning-based methods with human-annotated defects. However, this previous approach has some limitations and risks. These include label mistakes due to overlapping defect structures, poorly reproducible annotations and varying label quality. Furthermore, existing algorithms have to be adapted to new cell lines or a new labeling process is required. We overcome these challenges by avoiding the use of human labels and instead perform the CNN training on the basis of spatially resolved reference measurements, which allows us to calculate spatially resolved labels in less than a second. This purely data-driven approach allows a fast training to quantify defects with physical relevance regarding dark saturation current density (0) and series resistance ( ). The trained CNN achieves a precision of 88% and a recall of 91% for 0 defects while for defects it attains a precision of 78% and a recall of 86%. The accelerated training process allows a fast deployment of deep learning models in the solar cell line.