Zero-Shot Classification of Additive Manufacturing Anomalies

Additive Manufacturing (AM) process involves building part layer by layer. Some of the AM processes ( Laser and Electron beam based) generate a melt pool during printing process. This melt pool can be captured periodically during AM process using special optical arrangements. These images capture high intensity melted zone, heat affected zone, splattered molten metal particles and overall shape of the melt pool. These images carry similar characteristics for good AM processes within a range. When there is an anomaly the above said characteristics of the melt pool changes, for example a low intensity melted zone signifies low energy condition which can lead to defects like balling etc. Hence the captured image at this condition appears significantly different from other images. The common defects which can be detected by analyzing melt pool images are porosity, spatter, lack of fusion, cracks, balling and keyhole instability. There are many machine learning methods available to quantify this (supervised and unsupervised). The proposed approach does not require a trained machine learning (ML) model from scratch but rather utilize a pretrained self-supervised vision transformer (ViT) model DINO v2. The melt pool images acquired during the additive manufacturing (AM) process are processed by DINO v2 ViT model. These future vectors or embeddings for all printing instances are extracted and stored for downstream processing. Zero shot classification (accept or reject) of an AM printing instance is done by comparing the production printing instance image with a baseline printing instance image at the same spatial coordinates using Euclidean distance metric. This baseline Euclidean distance metric is established from multiple printed instances of the baseline coupons and by calculating root mean square deviation (RMSD) at each spatial coordinate of the print instances. The deviations of the production print instances are calculated by calculating RMSD of the Euclidean distances of embeddings of the melt pool images with that of baseline image embeddings at the same spatial coordinates. The deviations observed can be correlated to actual defects by analyzing AM printed parts layer by layer at each spatial coordinates using destructive or non-destructive testing techniques. The scope of this paper is only on establishing the deviation calculation methodology.