Non-Destructive Inspection of Solar Modules using Machine Vision
Solar power is central to the shift away from conventional energy, already making up over a fifth of Germany’s renewable energy mix in 2021, which makes it a key technology in the fight against climate change. But solar modules degrade over time from mechanical and chemical stress, and while manufacturers routinely inspect modules during production using imaging, regular on-site inspection of installed modules hasn’t become standard practice yet. This thesis works toward making that on-site inspection efficient, using electroluminescence (EL) and photoluminescence (PL) measurements.
Before any module or cell-level analysis can happen, the raw measurements need preprocessing. The first part of the thesis tackles detecting and segmenting individual modules and cells, proposing two methods: one using thresholding and morphological operations that reliably detects 95% of modules in a diverse test set, and a second that uses intensity profiles to precisely locate module and cell boundaries, enabling rectification and cell-level segmentation. This second method improves detection rate by 2.5x and runs 40x faster than the prior state of the art.
The second part addresses automatically detecting and segmenting cracks in EL images. Since pixel-level annotations are usually needed for this kind of segmentation but are expensive to produce, the thesis instead trains only on simple image-level labels and derives coarse segmentation masks through weakly-supervised learning, making the method easy to apply to new data with little extra effort.
The third part tackles a related problem: tiny defects are hard to spot in low-resolution EL images. The thesis introduces a customized multi-frame super-resolution method that combines several low-resolution images into one higher-resolution image, and shows both quantitatively that this improves automated defect segmentation, and through an expert study that it also helps human inspectors, outperforming the prior state of the art for this specific task.
The fourth part looks at a module’s core purpose, producing power, and estimates module power directly from a single EL or PL measurement using deep learning, achieving a mean absolute error of just 3.2% on EL measurements, with the same approach transferring to PL measurements via transfer learning. A small architectural tweak also enables computing localized power loss per cell or per defect, simplifying statistical analysis of how different defect types affect power output across large datasets, notably while still training on only module-level labels, showing that weakly-supervised learning extends effectively to regression problems too.
Finally, the thesis releases an open-source toolbox bringing together the algorithms developed throughout, along with convenient tools for working with EL and PL images and their metadata, interfaces to common data science libraries like NumPy, Pandas, and PyTorch, and a design built for easy extension with new algorithms.