Machine learning for defect detection for PBFAm using high resolution layerwise imaging coupled with post-build CT scans

This paper develops a methodology based on machine learning to detect defects during Powder Bed Fusion Additive Manufacturing (PBFAM) processes using data from high resolution images. The methodology is validated experimentally using both a support vector machine (SVM) and a neural network (NN) for binary classification. High resolution images are collected each layer of the build, and the ground truth labels necessary for supervised machine learning are obtained from a 3D computed tomography (CT) scan. CT data is processed using image processing tools-extended to 3D-in order to extract xyz position of voids within the component. Anomaly locations are subsequently transferred from the CT domain into the image domain using an affine transformation. Multi-dimensional features are extracted from the images using data surrounding both anomaly and nominal locations. Using cross-validation strategies for machine learning and testing, accuracies of close to 90% could be achieved when using a neural network for in-situ anomaly detection.