Multivariate statistical monitoring of batch processes

Extensive overload of data obtained from batch processes see the need for reduced dimensional analysis to gain better ground for fault detection and diagnosis to be carried out. Principal component analysis (PCA) is used to extract information from the unfolded matrix derived from a three-dimensiona...

Full description

Saved in:
Bibliographic Details
Main Author: Tong, Rhoda Min Ting.
Other Authors: School of Chemical and Biomedical Engineering
Format: Final Year Project
Language:English
Published: 2009
Subjects:
Online Access:http://hdl.handle.net/10356/16386
Tags: Add Tag
No Tags, Be the first to tag this record!
Institution: Nanyang Technological University
Language: English
Description
Summary:Extensive overload of data obtained from batch processes see the need for reduced dimensional analysis to gain better ground for fault detection and diagnosis to be carried out. Principal component analysis (PCA) is used to extract information from the unfolded matrix derived from a three-dimensional array of data obtained from the process measurements of a semiconductor etching process. The Multi-way PCA way of unfolding is applied. A reference model is first generated using batches representative of a successful operation and then used as the statistical reference framework to classify a new batch run as normal or abnormal. Multivariate Statistical Process Control charts are subsequently used to track the progress of new batch runs and detecting abnormalities. Global and local PCA models for post-analysis on historical batch data are developed and their robustness in fault detection abilities compared. Results from the scores plots show that samples with abnormal behavior are reflected more apparently by local models, with majority of the fault-induced samples projected beyond the 95% confidence ellipse in the local models while the same samples are positioned within the confidence ellipse in the global model with only a slight variation from its main cluster. Hotelling’s T2 value and Q-statistic charts present further insights into faulty batches inadequately captured or missed by the scores plots. Contribution plots provide a method of fault diagnosis for the faulty batches identified. Enhanced and timelier fault detection and diagnosis can occur with the implementation of online-process monitoring.