Comparisons of Unsupervised and Supervised Machine Learning Algorithms for Prediction of Damage in Composites From 3D Micro-CT Images

Abstract The objective of this work was to conduct statistically meaningful comparisons of the unsupervised and supervised machine learning (ML) methods for damage segmentation in composites from 3D Micro Computed Tomography (micro-CT) data and explore synergy between unsupervised and supervised ML. This synergy enables one to combine strong mathematical rigor of the unsupervised ML methods with flexibility and accessibility of the supervised ML algorithms. The unsupervised ML method relied on the statistical distances in conjunction with grayscale threshold intensity segmentation to isolate damage present in high resolution image data. The deep learning models used in this work were based on the U-Net and FC-DenseNet architectures. Both unsupervised and supervised ML methods were applied to the analysis of low velocity impact damage in the carbon fiber reinforced polymer (CFRP) composites. The performance of the methods was assessed using metrics from the statistical classification theory.