Improving Monitoring and Diagnosis for Process Control Using Independent Component Analysis

Thaddeus T. Shannon, III

ABSTRACT
Statistical Process Control (SPC) is the general field concerned with monitoring the operation and performance of systems. SPC consists of a collection of techniques for characterizing the operation of a system using a probability distribution consistent with the system's inputs and outputs. Changes observed in the empirical distribution over time are attributed to changes in the system's operation. The proper design of an SPC scheme for a particular system requires that the disturbances external to the system be distinguished from internal system disturbances.

Classical SPC monitors a single variable to characterize the operation of a single machine tool or process step using tools such as Shewart charts or Cumulative Sum (CUSUM) charts. This traditional approach works well for simple small to medium size processes, but becomes inefficient when applied to large (many step) processes. For these cases a number of multivariate SPC techniques have been developed in recent decades. These advanced SPC methods work well in many circumstances but suffer from several disadvantages compared to univariate techniques: they tend to be statistically less powerful, and they tend to complicate process diagnosis when a disturbance is detected.

This research proposes a general method for simplifying multivariate process monitoring in such a manner as to allow the use of traditional SPC tools while facilitating process diagnosis. The objective is to develop latent variable representations of complex processes which are directly identified with process steps or segments. Our method models disturbances in the process rather than the process itself. The basic tool used is a relatively new method for data analysis known as Independent Component Analysis (ICA). The methodology is illustrated on the problem of monitoring Electrical Test (E-Test) data from a semiconductor manufacturing process. On the sample problem, the methodology successfully distinguishes between the four different types of disturbances it is designed for. Development and production data from a working semiconductor plant are used to estimate a factor model that is then used to develop univariate control charts for particular types of process disturbances. Detection and false alarm rates for data with known disturbances are given. The charts correctly detect and classify all the disturbance cases with a very low false alarm rate.

I also introduce a method for performing an ICA like analysis using possibilistic data instead of probabilistic data. This technique extends the general ICA framework to apply to a broader range of uncertainty types. Further development of this technique could lead to the capability to use extremely sparse data to estimate ICA process models. Such a capability would allow rare or costly disturbances to be modeled in our framework.


Friday, May 11, 2007
10:00 a.m.-12:00 Noon

Dissertation Committee
Andrew M. Fraser, Chair
George G. Lendaris
James McNames
Martin Zwick
Serge Preston, Graduate Studies Rep.