Electroluminescence (EL) imaging is a useful method for photovoltaic PV diagnostic inspection. EL images provide high spatial resolution, which makes it possible for detecting and quantifying different types of faults inflicted the PV modules. A dataset of 982 EL images of polycrystalline module with different defects are analyzed and several image pre-processing steps were applied before using them for analysis. For classification, the dataset are manually labelled into three categories (micro-crack mode B, C and no failures) and five categories (micro-crack mode A, B and C, finger interruptions and no failures) respectively for automatic fault detection. To obtain a strong performance, we investigate and compare two supervised machine-learning approaches i.e. k-Nearest Neighbors and Random Forest Classification for the above-labelled dataset. The goal of the paper is to check whether the addition of extracted statistical parameters from the histogram used as a feature descriptor provides any advantage over the classical machine learning method that uses the histogram of the pixel value as an input. Both approaches are trained on yielded 40000 cells extracted from high-resolution EL intensity images. A training and testing framework will be generated using different levels of training to testing ratio to address the goal and importance of the derived statistical parametric block as an input and to derive the statistical parameters sensitivity with respect to different faulty class labels.
|Tidsskrift||Progress in Photovoltaics: Research and Applications|
|Status||Under udarbejdelse - 2020|
Parikh, H. R., Buratti, Y., Spataru, S. V., A. dos Reis Benatto, G., Poulsen, P., Mantel, C., Séra, D., & Hameiri, Z. (2020). Automatic Fault detection through Supervised Machine Learning Techniques using histogram and extracted statistical parameters. Manuskript under forberedelse.