MODELING OF QUALITY PROFILE DATA WITH APPLICATION IN MANUFACTURING AND BIOMEDICAL ENGINEERING

Abstract

The quality of the output of a complex system is often recorded as multidimensional profile data with panel structure. In such structure, the quality of each individual in the output is measured repeatedly based on time or other variables. In this dissertation, the quality profile data are modeled to address two types of problems: (a) to explore the underlying relationship between the parameter of interest in the complex system and the resulting quality under the condition that the principal mechanism is not fully known and (b) to quantify the uncertainties among the output. For the first type of problem, we consider a constrained semiparametric varying coefficient model. The system parameter of interest is treated as a covariate whose effect upon the resulting quality is modeled nonparametrically as a function of time. Any existing physicochemical knowledge related to other factors in the system that affect the resulting output quality is modeled parametrically as an additive term in the model. In the situation that expert knowledge about the effect of the parameter is available, some constraints can be incorporated in the model such that the estimated effect aligns with the given knowledge. For the second type of problem, mixed-effect model is developed to quantify the uncertainties among output using random effects. These random effects can be utilized for anomaly detection or for variation quantification where deviation among individuals is of interest depending on the context of the data. Three case studies from manufacturing and biomedical engineering domains are presented in the dissertation where the above two types of problems are discussed.