Enhancing Composite Micrograph Analysis with Semantic Segmentation

Abstract

The realization of future hydrogen-powered aircraft requires the development of liquid hydrogen storage tanks, possibly made of fiber composite materials, that must meet stringent leak tightness requirements under cryogenic temperature conditions. A typical challenge in the production of fiber composite components is that the material and the component are simultaneously formed by the interaction of fiber architecture, auxiliary materials, and manufacturing parameters in the production process. This results in a high risk of undesired inhomogeneities. Visual inspection of micrographs is standard practice to assess material and process parameters, but it involves high manual effort and leads to hardly quantifiable results. In this paper, we discuss the use of semantic segmentation to inspect micrographs with a special focus on crack formation. We present our data labeling process and machine learning models the best of which achieves a mean intersection over union of 0.9071. We also provide two post-model tools to assist domain experts in local fiber volume ratio analysis and in crack detection with a crack pixel recall of more than 95 %. The proven segmentation accuracy allows us to state that our semantic segmentation model greatly simplifies, accelerates, and quantifies the analysis of fiber composite micrographs.