Laser scan-based anomaly detection in metal additive manufacturing using multimodal ensemble learning

Le résumé fourni par la source

Metal additive manufacturing (AM) is gaining traction in product development as a general-purpose technology that is used to manufacture rapid prototypes, rapid tooling, and end-use products. However, concerns about reliably attaining stringent product specifications hinder its broader uptake. This study designs and evaluates a novel process for detecting laser scan-based defective regions very close to the nominal process caused by uncertainty in the powder bed fusion. A tailor-made test specimen (n=442) was designed for the experiment. The build parameters within the individual layers of each test specimen were manipulated to seed laser scan-based defective regions representing basic primitives, i.e., circle, triangle, square, and pentagon. The novelty also stems from the design that the laser power and/or laser scan speed fluctuate on the fly without having the laser switch ON/OFF within a single vector scan. This scan-based design represents a closer imitation of real process defects representative of primitive shapes as a function of laser scans rather than geometry of the parts. Volumetric energy density was varied to ±30% with an increment of 5% from the nominal. The size of the defective primitives ranged from 5 µm to 2000 µm. A late-fusion multimodal ensemble approach was developed and deployed based on datasets from both a near-infrared Scientific Complementary Metal-Oxide-Semiconductor (sCMOS)-based optical tomography (OT) system and a dual-photodiode-based meltpool monitoring (MPM) system. The results revealed the probability of detection using multimodal ensemble learning. The outcome aids designers and researchers in developing products in line with stringent specifications put forth by precision industries.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.