Machine vision-based defect detection of solar cells/modules

in electroluminescence images

ABSTRACT

Finger interrupt, micro-crack and breakage are severer defects in the multicrystalline solar cell and cannot be observed by the naked eyes or the conventional CCD camera. The Electroluminescence (EL) imaging technique can be used instead to highlight these defects. In this study, a machine vision scheme is proposed to detect the defects of solar cells and solar modules in EL images. The EL image of a multicrystalline solar cell presents a heterogeneously textured pattern, which makes the defect detection task extremely difficult.

    The first topic in this research focuses on defect detection of the multicrystalline solar cells in EL images. Since the finger interrupt and crack are line- or bar-shaped, the Fourier transform is used to eliminate suspected defects and results in a defect-free surface in the reconstructed image. By subtracting the reconstructed image from the original image, the defects can be distinctly enhanced in the difference image. Then, the defect is effectively segmented by a simple statistical control limit. The second topic of this research aims at defect inspection in the solar module, which is formed by a matrix of solar cells through series and parallel combinations. The Independent component analysis (ICA) is used to generate the basis images from defect-free solar cells. Each test image is reconstructed by a linear combination of the basis images. The accumulated gray-level difference between the test image and the reconstructed image is effectively used as a discrimination to detect the presence of defect in the solar cell subimages.

    In the experimental results of solar cell EL images, the Fourier transform reconstruction scheme can effectively detect fingers interrupt, micro-crack, and breakage. The average computation time is 0.29 seconds for a solar cell image of size 550×550 pixels. The experimental results of solar module inspection show that the ICA image reconstruction method can provide up to 98.7% of correct classification. The average computation time is 1.08 seconds for a solar module image (containing 36 solar cells) of size 1250×1250 pixels.

Keywords: Machine vision, Defect detection, Electroluminescence, Fourier transform, Independent component analysis, Solar cell, Solar module.