Assessment of current vision-based machine learning modelling efforts in directed energy deposition on defect prediction

Metal-based direct energy deposition processes ideally require feedback sensing of the deposition quality using camera detectors as they provide spatial and temporal state signatures of the process. Image processing algorithms are challenging to develop due to changing process operating conditions. Machine learning models have recently gained in popularity due to their ability to predict process defects using various monitoring technologies, and in particular vision-based cameras. The existing machine learning model architectures in literature are however trained on image-based datasets acquired through in-process recording of a single AM machine, a single camera setup and a limited set of process influencing parameter combinations. Comparison of the existing models is therefore challenging. Furthermore, the generalizability of these models is questionable, since none of the models evaluate model performance on unseen data, acquired from a different DED machine setup. A thorough investigation and comparison of the current state-of-the-art machine learning architectures on a large diverse dataset, acquired through vision camera based monitoring of several different DED machines and camera setups is however currently missing. The objective of this research is to compare the state-of-the-art machine learning architectures which use vision-based cameras to predict defects during metal-based material deposition using DED machines.

Faculty Supervisor:

Mihaela Luminita Vlasea

Student:

Partner:

I-INC Foundation for Business Development

Discipline:

Engineering

Sector:

Professional, scientific and technical services

University:

University of Waterloo

Program: