Using Predictive Models in Engineering Design: Metamodeling, Uncertainty Quantification, and Model Validation
Author: Xiong, Ying
Supervisor: Chen, Wei
Institution: Northwestern University, Evanston, Illinois
This dissertation is motivated by the need to develop methods which connect the engineering and marketing domains to enable identification of the preferred engineering system configuration, considering the real complexities in engineering system design and the heterogeneity of consumer preferences for such systems. The research includes a design process tool, an experimental design approach for human appraisal experiments, a multivariate statistical analysis methodology for human appraisal data, and finally an integrated Bayesian hierarchical choice modeling method which rigorously considers consumer heterogeneity and the nature of complex system design. This research primarily uses an automotive vehicle occupant package design as a motivating example, to both illustrate the issues in system design and demonstrate the features of the proposed design approach. The research can be divided into four primary contributions. A new process tool called Product Attribute Function Deployment (PAFD) is introduced as a decision-theoretic, enterprise-level process tool to guide engineering design. The PAFD method is a model-based approach built upon established methods in engineering, marketing, and decision analysis to eliminate the need for user ratings and rankings of performance, priority, and attribute-coupling used in current process tools. To collect data necessary to support preference modeling, an algorithmic design of human appraisal experiments method is developed to identify the optimal human appraisal experiment for a given set of requirements. The advantages of this approach over competing approaches for minimizing the number of appraisal experiments and model-building efficiency are clearly demonstrated. An issue with human appraisal experiments is that the heterogeneity of the experimental respondents contributes to the response, and this heterogeneity must be understood to separate the influence of design factors from that of human factors. Multivariate statistical techniques are utilized to create a human appraisal analysis methodology to understand heterogeneity and preprocess the human appraisal data to enable preference modeling. The Integrated Bayesian Hierarchical Choice Model (IBHCM) framework provides a unified choice modeling approach for complex system design. It utilizes multiple model levels to create a link between qualitative attributes considered by consumers when selecting a product and quantitative attributes used for engineering design.