Task-driven data fusion for additive manufacturing

Hu, Fu 2023. Task-driven data fusion for additive manufacturing. PhD Thesis, Cardiff University. Item availability restricted.

Preview	PDF - Accepted Post-Print Version Download (7MB) | Preview
	PDF (Cardiff University Electronic Publication Form) - Supplemental Material Restricted to Repository staff only Download (483kB)

Abstract

Additive manufacturing (AM) is a critical technology for the next industrial revolution, offering the prospect of mass customization, flexible production, and on-demand manufacturing. However, difficulties in understanding underlying mechanisms and identifying latent factors that influence AM processes build up barriers to in-depth research and hinder its widespread adoption in industries. Recent advancements in data sensing and collection technologies have enabled capturing extensive data from AM production for analytics to improve process reliability and part quality. However, modelling the complex relationships between the manufacturing process and its outcomes is challenging due to the multi-physics nature of AM processes. The critical information of AM production is embedded within multi-source, multi-dimensional, and multi-modal heterogeneous data, leading to difficulties when jointly analysing. Therefore, how to bridge the gap between the multi-physics interactions and their outcomes through heterogeneous data analytics becomes a crucial research challenge. Data fusion strategies and techniques can effectively leverage multi-faceted information. Since AM tasks can have various requirements, the corresponding fusion techniques should be task-specific. Hence, this thesis will focus on how to deal with task-driven data fusion for AM. To address the challenges stated above, a comprehensive task-driven data fusion framework and methodology are proposed to provide systematic guidelines to identify, collect, characterise, and fuse AM data for supporting decision-making activities. In this framework, AM data is classified into three major categories, process-input data, process-generated data, and validation data. The proposed methodology consists of three steps, including the identification of data analytics types, data required for tasks, acquisition, and characterization, and task-driven data fusion techniques. To implement the framework and methodology, critical strategies for multi-source and multi-hierarchy data fusion, and Cloud-edge fusion, are introduced and the detailed approaches are described in the following chapters. One of the major challenges in AM data fusion is that the multi-source data normally has various dimensions, involving nested hierarchies. To fuse this data for analytics, a hybrid deep learning (DL) model called M-CNN-LSTM is developed. In general, two levels of data and information are focused on, layer level and build level. In the proposed hybrid model, the CNN part is used to extract features from layer-wise images of sliced 3D models, and the LSTM is used to process the layer-level data concatenated with convolutional features for time-series modelling. The build-level information is used as input into a separate neural network and merged with the CNN-LSTM for final predictions. An experimental study on an energy consumption prediction task was conducted where the results demonstrated the merits of the proposed approach. In many AM tasks at the initial stage, it is usually time-consuming and costly to acquire sufficient data for training DL-based models. Additionally, these models are hard to make fast inferences during production. Hence, a Cloud-edge fusion paradigm based on transfer learning and knowledge distillation (KD)-enabled incremental learning is proposed to tackle the challenges. The proposed methodology consists of three main steps, including (1) transfer learning for feature extraction, (2) base model building via deep mutual learning (DML) and model ensemble, and (3) multi-stage KD-enabled incremental learning. The 3-step method is developed to transfer knowledge from the ensemble model to the compressed model and learn new knowledge incrementally from newly collected data. After each incremental learning in the Cloud platform, the compressed model will be updated to the edge devices for making inferences on the incoming new data. An experimental study on the AM energy consumption prediction task was carried out for demonstration. Under the proposed task-driven data fusion framework and methodology, case studies focusing on three different AM tasks, (1) mechanical property prediction of additively manufactured lattice structures (LS), (2) porosity defects classification of parts, and (3) investigating the effect of the remelting process on part density, were carried out for demonstration. Experimental results were presented and discussed, revealing the feasibility and effectiveness of the proposed framework and approaches. This research aims to pave the way for leveraging AM data with various sources and modalities to support decision-making for AM tasks using data fusion and advanced data analytics techniques. The feasibility and effectiveness of the developed fusion strategies and methods demonstrate their potential to facilitate the AM industry, making it more adaptable and responsive to the dynamic demands of modern manufacturing.

Item Type:	Thesis (PhD)
Date Type:	Completion
Status:	Unpublished
Schools:	Engineering
Uncontrolled Keywords:	1). Additive Manufacturing 2). Data Fusion 3). Advanced Data Analytics 4). Industrial Data Mining 5). Data-driven Modelling 6). Machine Learning
Date of First Compliant Deposit:	2 April 2024
Last Modified:	02 Apr 2024 13:26
URI:	https://orca.cardiff.ac.uk/id/eprint/167657

Actions (repository staff only)

Edit Item