Contributed Session: Process Monitoring & Control
Statistical Process Control of Multivariable Dynamical Processes Using Automated Time Series Analysis
Wallace E. Larimore
Nancy J. Kirkendall
This paper discusses automatic methods for fitting multivariate time series models for use in the routine on-line application of statistical process control (SPC) to multivariate dynamical process data. The method is based on canonical variate analysis (CVA) between the process past and future. The major computation is a singular value decomposition that is always stable and avoids nonlinear parameter optimization usually employed. The method has been shown to be near optimal in selecting model structure and order and in parameter estimation accuracy even for quite small sample sizes. The CVA directly determines the system state vector that describes the dynamical order and behavior of the system. The CVA procedure has been extensively applied to complex multivariable dynamical processes such as chemical processes, vibrating systems, and industrial and manufacturing systems that include feedback, unstable dynamics, nonstationary disturbances, and up to 100 states. The SPC monitoring and diagnosis of a process can be performed directly and efficiently using projections on the low dimensional canonical variables. This leads to more powerful tests about changes in process behavior.
[Wallace E. Larimore, Adaptics, Inc, 40 Fairchild Drive Reading, MA 01867 USA; firstname.lastname@example.org ]
Automated System Monitoring & Diagnosis via Singular Value Decomposition
John L. Maryak
We present a new methodology for use in automated monitoring of a system to detect abnormal or degraded behavior. The methodology uses the singular value decomposition technique of matrix algebra to discover relationships between the elements of data observed from a normally operating system. It then tests these relationships against newly acquired data to detect system malfunctions. Distribution theory developed specifically for this application allows the system monitor to properly set threshold levels so as to control the false alarm rate. The algorithm is numerically stable and relativly easy to apply, and was tested successfully on data from a motorized globe valve.
[John L. Maryak, Applied Physics Laboratory, Johns Hopkins Univ., Laurel, MD 20723-6099 USA; email@example.com ]
Model-Free Control of General Processes
James C. Spall
Consider the pervasive problem of developing a controller for a process or system where the governing equations are not well known. This paper presents an approach based on estimating a controller without building or assuming a mathematical model of the system to be controlled. The controller is constructed through use of a function approximator (FA) such as a neural network or polynomial (no FA is used for the unmodeled system equations). This involves the estimation of the unknown parameters within the FA. However, since no functional form is being assumed for the system equations, the gradient of the loss function for use in standard optimization algorithms is not available. Therefore, this paper considers the use of the highly efficient "simultaneous perturbation stochastic approximation" algorithm, which requires only system measurements (not a system model). It is shown that this approach is powerful and can lead to practical controllers for a broad range of complex processes. An example of wastewater treatment will be presented. Information on other applications areas such as process and quality control, scheduling, and traffic management will also be briefly provided.
[James C. Spall, Applied Physics Laboratory, Johns Hopkins Univ., Laurel, MD 20723-6099 USA; firstname.lastname@example.org ]
Efficiency & Robustness of Discrete PI Control Schemes
Proportional-integral (PI) control schemes are used extensively for process control and variation reduction. For example, the well-known exponentially weighted moving average (EWMA) control schemes are special cases of PI schemes. Despite their widespread use, the efficiency and robustness properties of PI schemes have not been systematically investigated. In this paper, we study the behavior of discrete PI control schemes for different stationary and nonstationary disturbance processes and compare them with minimum mean-squared error (MMSE) control schemes Both theoretical and numerical results are obtained to show that the PI schemes have reasonably high efficiency and are also quite robust, especially in the presence of non-stationarity. This is joint work with Huaiqing Wu and Vijayan N. Nair.
[Fugee Tsung, 2203-6 Cram Place, Ann Arbor, MI 48105 USA; email@example.com ]
Date created: 6/5/2001