Comparison of Inline Crack Detection Systems for Multicrystalline Silicon Solar Cells

Abstract

Cracks in silicon solar cells reduce cell efficiency and may lead to critical failures in cell and module production because of breakage. Detecting cracks in an early production stage is, therefore, advisable. Several approaches for crack detection systems have been developed in the past years. In order to compare the system performance, the rating done by the system has to be compared with a reference, i.e., a ground truth defining the crack's presence, position, and/or area. This information can also be used for training of detection algorithms based on supervised learning, e.g., convolutional neural networks. In the present article, we show that finding the ground truth is ambiguous and depends strongly on the operator and the applied metrology. For this end, we apply six crack detection tools to a set of 120 multicrystalline silicon passivated emitter and rear cells that were sorted out in industrial production because of cracks, and compare the human rating of six operators for every tool. The operators’ precision, i.e., the fraction of the contrasts detected as cracks that actually are cracks, ranges from 73% to 100% and their recall, i.e., the fraction of all actual cracks that have been found, from 30% to 92% for all tools and operators. To achieve the highest precision and recall, we suggest rating the images of an optical near-infrared system combined with luminescence imaging by three operators conjointly for the definition of ground truth for the present as well as for larger sample sets.