EAGER: Real-D: Integrating Data-Driven Methods and Engineering Models in Manufacturing Systems

Georgia Tech Research Corporation, Atlanta GA

Investigators

Abstract

This EArly-concept Grant for Exploratory Research (EAGER) will contribute to national prosperity and economic welfare by studying new methods for combining machine learning with engineering knowledge to improve performance of manufacturing systems. Machine learning has attracted much attention as it offers the potential for analysis of massive data in various application domains, including engineering and manufacturing. However, the reliance on data-driven methods alone, outside of the context of engineering knowledge and physical principles, can led to misspecified models with low accuracy and/or black-box models with poor interpretability. On the other hand, engineering models generally rely on assumptions that may not hold in practice, leading to bias and poor predictive capability. This award supports fundamental research that bridges the gap between pure data-driven methods and those based purely on engineering models by introducing a novel framework that integrates statistical machine learning methods with physics, engineering and first principles to create more accurate analytical models for manufacturing systems. This research creates an analytical framework enabling a better understanding of manufacturing system performance through the fusion of data with engineering principles. This approaches is expected to improve product quality, increase machine availability and reduce manufacturing costs by identifying and controlling critical process factors. New methodologies developed in this research will be incorporated into the STEM educations curriculum and teaching activities. Going beyond existing machine learning techniques, this research will integrate data analysis and engineering modeling to provide more accurate methods for data analysis and prediction. These new methods are expected to outperform both data-driven and first principles models when they are used separately. The project will create a new sampling strategy for conducting experiments and collecting data that, unlike current design of experiment and optimal design methods, uses engineering models to guide the sampling direction and select the sampling points that most improve accuracy. The project will also build a new real-time dimension reduction and feature extraction method from streaming data that can extract both low-dimensional spatial and temporal features embedded in data streams leading to effective dimension reduction. A set of computationally efficient estimation algorithms will be developed that enable the real-time feature learning and analysis for high-velocity streams. From a quality improvement viewpoint, the project will enable researchers in the quality engineering community to reexamine quality monitoring and improvement methods with a new perspective based on the fusion of data and engineering models. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.