On dual control and adaptive Kalman filtering with applications in the pulp and paper industry.

摘要

This thesis is about dual control of time-varying stochastic processes in the pulp and paper industry with parameter estimation carried out via adaptive Kalman filtering. It can be divided into three parts.; The first part deals with the control of a chip refiner to produce wood pulp, where the process gain between the refiner motor load and the plate gap is both nonlinear and time-varying, with reversal in the sign of the gain indicating the onset of pulp pad collapse. The control objective is to regulate the motor load while avoiding pad collapse. The problem is principally stochastic in nature, since the gap at which gain reversal occurs can wander unpredictably. The proposed control strategy consists of an active suboptimal dual controller coupled with an adaptive Kalman filter for parameter estimation. The controller minimizes a nonlinear performance index designed especially to reflect the peculiarities of the process. Thus, no heuristic logic is needed. Simulations show the superior performance offered by this strategy.; The second part is concerned with CD coat weight control on bent blade coaters. The coater is a coupled multivariable process whose gain drifts over time and often switches sign. Current industrial practice is to switch off automatic control when the loop becomes unstable due to gain sign reversal. Because of this, the standard industrial controller is rarely on for more than half of the blade life. The proposed control strategy relies on the general framework introduced for the chip refiner problem. The controller takes into consideration the loading and bending limitations imposed by the actuators and the blade. An approximate analytical control law was derived to allow for easy implementation. The proposed strategy was successfully applied to an off-machine industrial coater. The results of a series of trials show the advantage of including probing in the control signal, and that the adaptive Kalman filter was capable of tracking gain variations. The dual controller yielded substantial quality improvement and was able to control the process throughout the entire blade life. The developed strategy was well accepted by the company.; The adaptive Kalman filter used for both applications had the disadvantage that one has to wait for some time before starting to use the noise variance estimates for parameter estimation. In the third part, a novel adaptive Kalman filtering scheme is derived to resolve this problem. This scheme is based on the Expectation-Maximization (EM) algorithm. It uses a number of fixed-point smoothers running in parallel to on-line estimate the variances of the process and measurement noise for a general linear, discrete, time-varying stochastic system. This approach implies that the estimates are used in the Kalman filter. The approach is evaluated through a number of simulation experiments.