Physics-informed machine learning for smart decision-making in ultrasonic metal welding

Ultrasonic metal welding (UMW) is a versatile solid-state joining technique with various important industrial applications, including lithium-ion battery assembly, automotive body construction, and electronic packaging. Among the advantages of UMW over conventional fusion welding techniques are the ability to join dissimilar metals, short welding cycles, energy efficiency, and environmental friendliness. Despite of its numerous advantages, UMW is sensitive to variations in process conditions and has a narrow operating window. Moreover, process disturbances, including tool degradation and workpiece surface contamination, negatively impact the UMW joint quality and robustness. As such, industrial-scale UMW production calls for smart decision-making, e.g., process optimization, joint quality assessment, maintenance, real-time control. To this end, this dissertation develops a suite of physics-informed machine learning methods for intelligent decision-making in UMW. A machine learning-based response surface method is developed for multi-objective optimization of peel and shear joint strengths of UMW. Machine learning models are employed to characterize the response surfaces of peel and shear joint strengths, which are shown to have different patterns. Using the established response surface models, an optimal combination of process parameters is obtained to co-optimize peel and shear joint strengths. A hierarchical physics-informed ensemble learning (PIEL) framework is developed to incorporate both physical knowledge and in-situ sensing data for accurate online prediction of UMW joint strength. This framework decomposes the joint strength variability into a physics-informed global trend and a data-driven residual, which are modeled hierarchically. It is shown that the PIEL framework improves physical interpretability and prediction accuracy. A multi-functional few-shot learning (MF-FSL) approach is created to enable fast and cost-effective adaptation of online monitoring algorithms to new production scenarios with very limited data availability. MF-FSL utilizes model-agnostic meta-learning to learn and transfer the meta-knowledge between source and target domains. It is demonstrated that MF-FSL is effective in a variety of decision-making problems. To deal with extremely data-scarce cases, where no data is available in the new production scenario, a Similarity-based Meta-Representation Learning (SMRL) method is created for domain generalization. Compared with state-of-the-art methods, SMRL achieves significantly better generalizability and prediction performance. It is expected that SMRL will advance the generalizability, adaptability, and agility of decision-making algorithms, which are critically needed in modern and future manufacturing.

Name thesis:degree_name

Ph.D.

Level thesis:degree_level

Dissertation

Discipline thesis:degree_discipline

Mechanical Engineering

Grantor

University of Illinois at Urbana-Champaign

Year dc:date

2023

Author dc:creator

• Meng, Yuquan

Contributors dc:contributor

• Shao, Chenhui
• Ferreira, Placid M
• Salapaka, Srinivasa M
• Wang, Pingfeng

• Ultrasonic Welding
• Data-efficient Learning
• Smart Manufacturing
• Few-shot Learning
• Domain Generalization

Statement dc:rights

• Copyright 2023 Yuquan Meng

Language dc:language

en, eng

Handle dc:identifier

https://hdl.handle.net/2142/120351

OAI record oai:identifier

https://www.ideals.illinois.edu/oai-pmh?verb=GetRecord&metadataPrefix=etdms&identifier=oai%3Awww.ideals.illinois.edu%3A2142%2F120351