Application of Deep Learning models to characterize manufacturing defects in additive manufactured components

Abstract

This work is part of a large package of work investigating the structural integrity of additively manufactured components in nuclear applications. Components manufactured using additive manufacturing (AM) often contain high degree of manufacturing defects which may lead to pre-mature failure in service conditions, for example, during fatigue loading. The aim of this work is to automate the detection and analysis of various types of manufacturing defects in AM using image segmentation techniques such that, robust predictive models can be built to correlate the defects with material deformation. The types of defects considered in this work are lack of fusion, micro-crack, spherical porosity, and unmelted particles.Four rectangular blocks of 316L stainless steel were fabricated using Selective Laser Melting (SLM) and Direct Lased Deposition (DED) techniques. The blocks were cut longitudinally, and the surfaces were metallographically polished and imaged using Scanning Electron Microscope (SEM). Deep neural network models were constructed based on three architectures: Fast Fully Convolutional Network (FCN), SegNet and UNet. The models were then trained using a set of micrographs utilizing various sets of hyperparameters. Model testing and validation were conducted using new and unseen images. Performance of the models were assessed based on training run times, testing run times and accuracy of detecting the correct locations, types, and sizes of the defects in the validation datasets.