High speed in-process defect detection in metal additive manufacturing

Chen, Siwen (2022) High speed in-process defect detection in metal additive manufacturing. MPhil thesis, University of Nottingham.

PDF (Thesis - as examined) - Requires a PDF viewer such as GSview, Xpdf or Adobe Acrobat Reader
Available under Licence Creative Commons Attribution.
Download (2MB) | Preview


Additive manufacturing (AM) is defined as the process of joining materials to make objects from 3D model data, usually layer upon layer, as opposed to subtractive manufacturing technologies. This fabricating technique is also famously known as ‘3D printing’. Although its entire manufacturing chain is becoming more mature by improved pre-defined design, more accurate heat input and motion system and cleaner in-chamber atmosphere, there are still a number of influential factors that can have a negative impact on the manufacturing process that introduce ‘defects’, which will greatly lessen the density of the parts or even result in failure. For this reason, it is critical to be able to discover them effectively during the manufacturing process. This thesis aims to develop a methodology for the measurement and characterisation of surface texture of AM parts. Typically, optical metrology instruments including focus variation (FV) microscopy and fringe projection (FP) have been used to measure the surface texture of AM samples due to their suitability and reliability in the field of metrology. The thesis also develops optimum filtration methodology to characterise the AM surface by comparing different filters. In the recent decades, machine learning (ML) is presenting a high robustness and applicability in defect detection in comparison to the traditional digital image processing technique. In this thesis, several ML techniques have been investigated into in terms of their suitability for the research based on the processed data secured from the optical measuring instrument. A detailed defect review that collects the information in terms of the defects in LPBF process based on the related research of the global researchers is given. It provides the details about different types of defects and discusses the potential correlation between process parameters and generated defects. ML and AM are both research fields that have developed rapidly in recent decades. In particular, the combination of the two can effectively achieve the purpose of AM parameter optimisation, process control and defect detection. A review of the adaptability of ML to different types of data and its application in feature extraction to achieve in-line or offline defect detection is given. Specifically, it demonstrates how to select proper ML technique given various types of data and how to choose appropriate ML model depending on different forms of defect detection (defect classification and defect segmentation). For data acquisition, the parameters including the magnification of objective lens and illumination source of the optical instrument are optimised to provide accurate and reliable data. Then the surface is pre-processed and filtered with the discovered optimal filtration method. The applicability of different types of machine learning methods for defect detection is also investigated. Results show that principal component analysis may not be a suitable tool for classifying defects if using exclusively whereas convolutional neural network and U-Net (full convolutional network) have shown good performance in correctly classifying defects and segmenting defects from the measured surface. For future work, more measurement instruments which can potentially achieve efficient and accurate metrology can be considered being developed and used, and the variety of samples needs to be increased to provide more types of surface topographies. In addition, how to improve the applicability of PCA in defect classification for AM parts can be studied on and more values of hyperparameters and number of parameters of neural networks can be used to further improve the suitability of the model for the training data.

Item Type: Thesis (University of Nottingham only) (MPhil)
Supervisors: Lawes, Simon
Leach, Richard
Keywords: Additive manufacturing; machine learning; focus variation; fringe projection; PCA; CNN; Moblie-Net; U-Net;
Subjects: Q Science > Q Science (General)
T Technology > TS Manufactures
Faculties/Schools: UK Campuses > Faculty of Engineering
Item ID: 67197
Depositing User: Chen, Siwen
Date Deposited: 31 Jul 2022 04:40
Last Modified: 31 Jul 2022 04:40
URI: https://eprints.nottingham.ac.uk/id/eprint/67197

Actions (Archive Staff Only)

Edit View Edit View