Li, Yixin
2024.
Leveraging deep learning for energy
consumption prediction in selective laser
sintering.
PhD Thesis,
Cardiff University.
![]() Item availability restricted. |
Preview |
PDF (Thesis)
- Accepted Post-Print Version
Download (6MB) | Preview |
![]() |
PDF (Cardiff University Electronic Publication Form)
- Supplemental Material
Restricted to Repository staff only Download (1MB) |
Abstract
Additive Manufacturing (AM), commonly known as 3D Printing (3DP), is a layer-by-layer manufacturing technique for fabricating objects based on digital models. This technology is widely applied across industries due to its highly customisable and flexible design capabilities. The considerable energy consumption that potentially limits widespread applications is particularly important when seeking suitable and cost-effective manufacturing methods. Energy management and optimisation before or during the process are challenging due to the high demand for data analytics in the dynamic working environment in AM systems. Therefore, Deep Learning (DL) has increasingly been recognised by academia and enterprises to manage energy predictive modelling based on different design-relevant parameters of AM processes. Advanced data analytics and DL techniques are leveraged to develop more accurate and efficient models for predicting and optimising energy consumption in AM systems. However, traditional energy modelling in AM systems has significant limitations, particularly in handling large and complex datasets in these dynamic environments in AM systems and capturing valuable insights from non-linear relationships between energy consumption and various parameters. This has led to a worldwide recognition of the problems associated with establishing energy predictive models in AM systems. Design for Additive Manufacturing (DfAM) considers energy efficiency with its functionality and manufacturability, and integrating data-driven systems can optimise AM systems. The precise information required to optimise designs could be provided by energy consumption modelling. DfAM helps manufacturers consider energy efficiency at the design stage, leading to more economical and sustainable manufacturing by managing the energy consumption of various design options and optimisation support. DL provides an alternative to building the framework of energy management and optimisation support. Compared to conventional analytical approaches, Deep Neural Networks (DNNs) take significant advantages in handling different data, revealing and predicting complex patterns or insights in AM systems. Differently from other accelerating devices, Field-Programmable Gate Arrays (FPGAs) are often reprogrammable to perform new types of computing tasks. This is due to the iii computing capabilities and flexibility, which allow them to work collaboratively with Central Processing Units (CPUs) in terms of training and inference. However, the complexity and memory requirements of predictive models pose challenges, failing to perform edge computing on FPGA platforms directly. This research is established based on a comprehensive framework for managing and optimising energy consumption and design-relevant parameters, integrating design-relevant parameters with image data to optimise overall energy consumption. The framework is organised into three key topics, which consider a Selective Laser Sintering (SLS) system as the case study. Firstly, a data-driven approach using multi-scale feature fusion techniques is proposed to predict energy consumption from different layer-wise image data, providing insights into the valuable energy consumption patterns. Secondly, to address the challenges of complexities of the model, Knowledge Distillation (KD) is employed, compressing a cumbersome teacher model to a lightweight student model, thereby deploying on the edge device. Finally, Particle Swarm Optimisation (PSO) utilises insights from the lightweight model to optimise design-relevant parameters, providing optimisation support for the case study. The framework offers a potential method to achieve an efficient design with optimal parameter combinations with adjustment of the energy consumption of different prototypes. The framework improves the accuracy of energy consumption predictions, facilitating more energy-efficient AM processes and sustainable manufacturing practices.
Item Type: | Thesis (PhD) |
---|---|
Date Type: | Completion |
Status: | Unpublished |
Schools: | Schools > Engineering |
Uncontrolled Keywords: | 1). Additive Manufacturing 2). Deep Learning 3). Knowledge Distillation 4). Energy Consumption 5). Optimisation Algorithms 6). Interdisciplinary Studies |
Date of First Compliant Deposit: | 15 April 2025 |
Last Modified: | 15 Apr 2025 15:15 |
URI: | https://orca.cardiff.ac.uk/id/eprint/177700 |
Actions (repository staff only)
![]() |
Edit Item |