Phase I Analysis for Monitoring Nonlinear Profiles in Manufacturing Processes

Summary: [This abstract is based on the authors' abstract.]Unlike linear profiles that can be represented by a linear regression model, nonlinear profiles are frequently sampled into high-dimensional data vectors and analyzed by nonparametric methods, meanwhile generating huge historical data sets that must be analyzed. High dimensionality and data contamination create problems for the Phase I analysis of nonlinear profiles. A two-component strategy is proposed: a data-reduction component that projects original data into a lower dimension subspace, and a data-separation technique that can detect single and multiple shifts as well as outliers in the data. Simulated data and nonlinear profile signals from a forging process illustrate the proposed strategy.