Short Course

Process Monitoring and Identification of

Dynamic Systems using Statistical Techniques

Instructors:
Wallace E. Larimore; Adaptics, Inc,
Dale E. Seborg, University of California, Santa Barbara


Friday, November 21, 1997, 8:30am - 5:30pm

during the American Institute of Chemical Engineers, Annual Meeting week

at the Sheraton Grand Hotel, Los Angeles, CA

SYNOPSIS

Traditional process monitoring in industrial plants is based on comparing measurements to specified limits, and on the experience of the plant personnel. More recently, statistical quality control techniques (SQC) methods have been widely used for product quality control. However, the standard SQC methods are based on the assumptions that the process dynamics are negligible and that the process disturbances are uncorrelated. But these restrictions are not valid for many real-time monitoring problems where the process dynamics are important and correlated disturbances are the rule, rather than the exception. Thus, more advanced monitoring methods are required for these applications. The potential payoff is the early detection of small changes in the process that can be corrected or compensated to reduce their ultimate impact on process performance.

In this short course, the traditional methods of process monitoring and statistical quality control will be reviewed, and several advanced methods for handling process dynamics and autocorrelated errors will be critically evaluated. These advanced methods include principal component analysis (PCA), projection to latent structures (PLS), and canonical variate analysis (CVA). They are based on multivariate statistics, time series analysis, and system identification. Numerous simulation and experimental examples are presented to illustrate key issues and to provide comparisons.

A new automated method is presented for identifying the system dynamics and autocorrelated error structure based on canonical variate analysis (CVA), a generalization of the PCA and PLS methods. The CVA approach allows for the completely automatic modeling of the process dynamics and error structure. From a software user's viewpoint, the modeling of the process dynamics and errors is no more difficult than fitting a static multivariate regression model. The CVA method is statistically optimal even in the presence of known or unknown feedback paths that are often present in process applications.

The CVA method is applied to a number of process monitoring problems using both simulated and industrial data, and is compared with alternative monitoring methods.

COURSE OUTLINE

  1. Overview of Process Monitoring Strategies
  2. Statistical Approaches to Process Monitoring
  3. Applications of Multivariate Statistical Monitoring Techniques
  4. Identification of Linear Dynamic Models from Input-Output Data
  5. Canonical Variate Analysis (CVA) Approach to Process Identification
  6. Process Monitoring Using Dynamic Models

Course Fee and Deadlines

The registration fee is $300 for reservations received before October 20, 1997 and $375 for reservations received after October 20. University students may register for $150. Full refunds will be made for cancellations received before October 20. A $100 cancellation penalty will be charged for cancellations received after October 20.

Registration Procedure

Please fill out the registration form and send it and a check to Dr. Larimore. (Credit cards cannot be used.) If you prefer to wire the registration fee, contact Dr. Larimore, preferably by e-mail.

Further Information: Contact either Dr. Larimore or Professor Seborg (seborg@engineering.ucsb.edu).
Dr. Wallace E. Larimore, President
Adaptics, Inc
1717 Briar Ridge Road
McLean, VA 22101
Phone: 703 532-0062; Fax: 703 536-3319
larimore@adaptics.com
Web Site: http://www.adaptics.com

Return to Adaptics Home Page