Machine-learning-based measurement, modeling, and control of spatial variability in advanced manufacturing

Spatial and spatiotemporal variabilities commonly exist in many advanced manufacturing processes and systems. High-performance characterization, modeling, and control of such variabilities are crucial for enabling next-generation smart decision-making in quality inspection, process control, machine health monitoring and prognostics, etc. However, fundamental challenges exist in these decision-making tasks. Although the recent development in measurement technologies has made it possible to acquire spatial and spatiotemporal data at different scales, it is generally expensive and time-consuming to use such technologies in real-world production settings. Additionally, there is a lack of effective and data-efficient methods for analyzing high-dimensional, heterogeneous spatial and spatiotemporal data in manufacturing. This dissertation presents novel machine-learning-based approaches for the measurement, modeling, and control of spatial and spatiotemporal variabilities. The effectiveness of these approaches is demonstrated using real-world data collected from different manufacturing processes and systems at different length scales including ultrasonic metal welding, additive manufacturing, and two-photon lithography. A hierarchical measurement strategy is developed to cost-effectively acquire spatiotemporal data by optimally allocating the measurement efforts in both spatial and temporal domains. Determining the observation times and measurement locations is formulated as a two-level decision-making problem. To expedite the solution search process, hierarchical genetic algorithm is adopted and implemented using high-performance computing. A hybrid multi-task-learning approach is created for accurate and cost-effective response surface modeling. This approach recognizes the differences and similarities between multiple manufacturing processes and accordingly constructs physics-informed self-learning and multi-task-learning models. Data efficiency and learning performance are thus improved through the transfer of information across processes. A hybrid hierarchical modeling method is developed to data-efficiently predict the feature-level geometric accuracy of parts produced by multiple identical additive manufacturing machines. The modeling method decomposes the geometric variability into a large-scale part-level trend and a small-scale feature-level term, which are characterized by a hierarchical Bayesian linear model and a Gaussian process model, respectively. A geometric compliance improvement framework for two-photon lithography is established to quantify the spatial variation in the geometric accuracy of 3D structures and generate compensation designs to improve geometric compliance. It is revealed for the first time that both systematic and random geometric errors exist in 3D structures fabricated by two-photon lithography.

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

2022

Author dc:creator

• Yang, Yuhang

Contributors dc:contributor

• Shao, Chenhui
• Ferreira, Placid M
• King, William P
• Wang, Pingfeng

• Machine learning
• Spatial process
• Spatiotemporal process
• Gaussian process
• Surface measurement
• Sampling design
• Measurement strategy
• Big data analytics
• Data fusion
• Smart manufacturing
• Advanced manufacturing
• Additive manufacturing
• Quality control
• Hierarchical modeling
• Data-efficient learning
• Intelligent metrology

Statement dc:rights

• Copyright 2021 Yuhang Yang

Language dc:language

en, eng

Handle dc:identifier

https://hdl.handle.net/2142/115503

OAI record oai:identifier

https://www.ideals.illinois.edu/oai-pmh?verb=GetRecord&metadataPrefix=etdms&identifier=oai%3Awww.ideals.illinois.edu%3A2142%2F115503