Solar energy plays a key role in the ongoing transition from conventional to renewable energies, making up already 20.2% of the renewable energies produced in Germany in 2021. As such, they are an indispensable technology to fight global warming. For efficient operation, solar modules require regular on-site inspection since they are known to degrade over time through mechanical and chemical stress. While manufacturers regularly apply automated inspection using imaging techniques during fabrication, regular on-site inspection of individual modules has not yet become a standard procedure. This thesis contributes to the efficient on-site inspection of individual solar modules using electroluminescence (EL) or photoluminescence (PL) measurements. For processing of EL or PL measurements on a module or cell level, preprocessing of the raw measurements is a prerequisite. The first part of this thesis deals with the detection and segmentation of module instances and cells. We propose two different methods, where the first one focuses on the detection of module instances. Using a series of thresholding and morphological operations, it is able to reliably detect 95 % of the modules in a diverse testing dataset. The second method uses intensity profiles to precisely detect the module boundaries and cells. This allows for a rectification and, if required, segmentation into individual cell instances. We have shown that our method improves over the state of the art in terms of the detection rate (2.5× higher) and computational performance (40× faster). The second part of this thesis deals with the automated detection and segmentation of cracks on EL images. A key challenge with defect segmentation is that it is usually required to have a comprehensive dataset with pixel-wise annotations available. Our method only takes image-level labels for training and derives coarse segmentation masks by weakly-supervised learning. As such, it can be applied to new types of data with little effort only. The identification of tiny defects on low-resolution EL images can be challenging. In the third part of this thesis, we present a customized multi-frame super-resolution (MFSR) method to compute an image with increased spatial resolution out of a series of low-resolution images. We quantitatively show that a successive defect segmentation benefits from the increased spatial resolution. With an additional expert study, we show that MFSR is beneficial, also for the manual inspection and that our proposed method improves upon the state-of-the-art for the specific task. The key property of a solar module is the ability to produce power. In the fourth part of this thesis, we analyze the determination of the module power from a single EL or PL measurement. Using a deep learning (DL)-based approach, we are able to estimate this property with a mean absolute error of 3.2 % for EL measurements and show that the same method can be applied to PL measurements as well by transfer-learning. Furthermore, we show that a slight modification of the architecture of the DL model enables the computation of a localized power loss. That localized power loss allows to compute the power loss per cell or per defect. As such, it simplifies statistical analyses of different defect types across large datasets. Remarkably, it is trained with module-level labels only and shows that weakly-supervised learning can be effectively applied to regression problems as well. The fifth part of this thesis presents an open-source toolbox for analyzing solar modules. It includes many of the algorithms developed as part of this thesis. In addition, it features a convenient set of tooling to work with EL or PL images along with meta data and provides interfaces to popular data science libraries like NumPy, Pandas or PyTorch as well. Furthermore, it is designed with extensibility in mind, such that new algorithms can be easily integrated
Mathis Hoffmann (Thu,) studied this question.