Nonlinear Profile Data Analysis for System Performance Improvement.

Abstract

The rapid development of distributed sensing and computer technologies has facilitated a wide collection of various nonlinear profiles during system operations, thus resulting in a data-rich environment that provides unprecedented opportunities for improving complex system operations. At the same time, however, it raises new research challenges on data analysis and decision making due to the complex data structures of nonlinear profiles, such as high-dimensional and non-stationary characteristics. In this dissertation, for the purpose of system performance improvement, new methodologies are proposed to effectively model and analyze nonlinear profile data. Specifically, three major research tasks are accomplished. First, the problem of informative sensor and feature selection among massive multi-stream sensing signals is discussed. In this research, a new hierarchical regularization approach called hierarchical non-negative garrote (HNNG) is proposed. At the first level, a group non-negative garrote is developed to select important signals, and at the second level, the individual features within each signal are selected using a modified version of non-negative garrote that can guarantee nice properties for the estimated coefficients. Second, a new methodology has been developed to analyze cyclic nonlinear profile signals for fully characterizing process variations and enhancing the fault diagnosis capability. In the proposed method, both within- and between-profile variations are taken into consideration. In order to accomplish this, a new mixed-effect model integrated with multiscale wavelets analysis has been developed. Third, the problem of modeling and monitoring of binary survival profiles has been studied. A general Phase I risk-adjusted control chart is being proposed based on a likelihood ratio test derived from a change-point model. Furthermore it is shown that binary survival outcomes depend on not only patients’ health conditions prior to surgery, but also other categorical operational covariates, such as different surgeons. The efficacy of the proposed methods in each chapter is validated and demonstrated by Monte-Carlo simulations and real-world case studies.