Graduation Year

2005

Document Type

Thesis

Degree

M.S.I.E.

Degree Granting Department

Industrial Engineering

Major Professor

Qiang Huang, Ph.D.

Committee Member

Jose L. Zayas-Castro, Ph.D.

Committee Member

Tapas Das, Ph.D.

Keywords

Variation decomposition, Change detection, Faulty condition, Discrimination, False alarm

Abstract

With current advances in sensors and information technology, online measurements of process variables become increasingly accessible for process control and monitoring. Such measurements may take the shape of curves rather than scalar values. The term Multichannel Functional Data (MFD) is used to represent the observations of multiple process variables in the shape of curves. Generally MFD contains rich information about processes. The challenge of process control in MFD is that Statistical Process Control (SPC) is not directly applicable. Furthermore, there is no systematic approach to interpret the complex variation in MFD. In this research, our objective is to develop an approach to systematically analyze the complex variation in MFD for process change detection and process faulty condition discrimination.

The main contributions of this thesis are: MFD decomposition, process change detection, and process faulty condition discrimination. We decomposed MFD into global and local components. The approach reveals global and local variations that are due to global signal shifts and local variations. Global variation was extracted using weighted spline smoothing technique, whereas, local variation was obtained by subtracting the global variation from original signals. Weights were obtained using the local moving average of the generalized residuals.

The proposed approach helps in process change detection and process faulty condition discrimination based on further MFD analysis using Principal Curve Regression (PCuR) Test. For process change detection, global variation component was used in the PCuR test. In-control global data sets were used as training data to detect process change that is due to global and local variation. On the other hand, for faulty condition discrimination purpose, local variation component was used in the PCuR test. In-control local variation data sets were used as training data in the PCuR test; therefore, process faulty condition that is due to local variations remains in control, whereas, process faulty condition that is due to global shifts appears as random out of control points in the PCuR test.

We applied our approach on real life forging data sets. A simulation study was also conducted to verify the approach and results are promising for wide applications.

Share

COinS