| Author: | Wang, Ruoxin |
| Title: | Deep learning based surface characterization for additive manufacturing |
| Advisors: | Cheung, C. F. Benny (ISE) |
| Degree: | Ph.D. |
| Year: | 2023 |
| Subject: | Three-dimensional printing Deep learning (Machine learning) Hong Kong Polytechnic University -- Dissertations |
| Department: | Department of Industrial and Systems Engineering |
| Pages: | xviii, 165 pages : color illustrations |
| Language: | English |
| Abstract: | Additive manufacturing (AM) has been extensively applied in various industries, such as aerospace, biomedicine, energy, and optics, especially for those products with complex structures and high value. Due to its ability to transfer models directly into products, AM is more effective and low-cost compared with traditional processing methods. Selective laser melting (SLM) is a common AM method that is based on powder-bed fusion to process metals. SLM involves complicated physics during processes such as powder melting, laser beam jetting, material evaporation, and solidification, resulting in the formation of many defects such as cracks, porosities, incomplete fusion holes, and so on. These defects have adverse effects on the mechanical and physical properties of the workpiece, such as stiffness, surface quality, and dimensional accuracy. As a result, the characterization of surface defects is of great importance, which enables us to obtain detailed information on surface quality. Based on this information, product selection or process analysis can be done to further conduct process optimization and online monitoring. In the AM process, many process parameters have a determinative effect on the final surface quality, and linking defect features to process parameters helps to analyze processes and improve surface quality. Surface characterization is responsible for exploring a representation to render a surface, which helps to extract more detailed and deeper features of surface. Since defects on the surface of 3D-printed components are inevitable and have various morphologies and distributions, extracting the features of defects is a challenging task. Although there are some filters developed to extract features from images, manual feature selection is a tedious and time-consuming task. Thanks to the development of deep learning which has powerful feature extraction and selection abilities, it has been widely applied in many industrial fields, such as face recognition, automobiles, and security. This research aims to explore deep learning-based methods and their application to the surface characterization of 3D-printed components. Although there are some studies that have made some effort to classify the defects or detect their existence, their specific location or distribution and other surface features have not yet been fully investigated. To fill this gap, a supervised convolutional neural network (CNN)-based method for surface characterization is developed that counts, classifies, and localizes surface defects. To save time and cost, the distribution and count of defects are first estimated through a proposed model named a detail-aware dilated convolutional neural network (DDCNN). DDCNN can provide information related to location and surface quality. To detect and characterize surface defects more comprehensively and accurately by combining distribution features, a CenterNet with a count loss (CenterNet-CL) model is built to perform count, localization, and classification characterization simultaneously. The proposed model achieves better counting and detection accuracy on the defect dataset compared with other state-of-the-art models. Since labeling is a tedious and time-consuming manual task, we further develop an unsupervised CNN-based method, named unsupervised segmentation with self-attention (USSA), to segment defects without labeled data while obtaining exact defect morphology. The proposed method is composed of three modules, which are responsible for extracting features, capturing global information, and assigning cluster labels, respectively. The model shows good segmentation performance for three materials. To conduct process analysis and optimization based on defect characteristics, exploring the relations among process parameters, materials, defect features, and surface quality is necessary. A hypergraph, a kind of graph that can describe high-order correlations in networks, has received much research attention due to the high complexity of real-world networks. In this research, a graph convolutional network-based process analysis and optimization method is proposed. In the method, a novel heterogeneous hypergraph embedding learning method named exp2vec is developed to obtain a low-dimensional representation of a graph. These embeddings are fed into a modified generative model named a variational graph auto-encoder with correction (VGAE-Corr) to reconstruct the graph for link prediction. The model shows superiority in regard to link prediction performance. |
| Rights: | All rights reserved |
| Access: | open access |
Copyright Undertaking
As a bona fide Library user, I declare that:
- I will abide by the rules and legal ordinances governing copyright regarding the use of the Database.
- I will use the Database for the purpose of my research or private study only and not for circulation or further reproduction or any other purpose.
- I agree to indemnify and hold the University harmless from and against any loss, damage, cost, liability or expenses arising from copyright infringement or unauthorized usage.
By downloading any item(s) listed above, you acknowledge that you have read and understood the copyright undertaking as stated above, and agree to be bound by all of its terms.
Please use this identifier to cite or link to this item:
https://theses.lib.polyu.edu.hk/handle/200/12722