By having access to data for several previous occurrences, it is more likely that a person familiar with the process can discern important patterns and identify the underlying cause(s) for the abnormal condition. Suppose that it is desired to analyze an abnormal condition, which is represented by multivariate time-series data for key process variables (e.g., measurements of controlled and manipulated variables for several interacting control loops). The objective is to locate similar, previous episodes (if they exist) in a large historical database, using an unsupervised learning technique. The proposed method does not require a process model, training data, or planned experiments. Instead, the analysis is based on historical operating data, which may be compressed
Chemical manufacturing processes present many challenging control problems, including: nonlinear dynamic behaviour; multivariable interactions between manipulated and controlled variables; unmeasured state variables; unmeasured and frequent disturbances; high-order and distributed processes; uncertain and time-varying parameters; unmodelled dynamics; constraints on manipulated and state variables; and (variable) dead time on inputs and measurements. Further, reliable measurements of important variables to be controlled, such as quality related variables, are often difficult to obtain on-line. A number of control approaches and algorithms that are able to handle some of the above process characteristics have been presented in the academic literature in resent years. Bequette (1991) gives a review of various approaches, such as: internal model approaches; differential geometric approaches; reference system synthesis techniques, including internal decoupling and generic model control; model predictive control approaches; and also various special and ad hoc approaches. Many of these Automatic Control approaches are not able to handle the various process characteristics and requirements met in industrial applications, and some of the approaches can only be applied for special classes of models.
Nonlinear model predictive control appears to be the only general approaches which can handle most of the common process characteristics and industrial requirements in a satisfactory way. It also seems to be the approaches, which are most suitable for the development of general and application independent software, which is essential for the development of cost-effective applications. For the above reasons this survey will focus on nonlinear model predictive control approaches presented in the open literature. Algorithms for nonlinear model predictive control are often extensions of linear model predictive control algorithms. For continuity the main characteristic features of linear model predictive control are briefly discussed.
A nonlinear model is used for predicting the effects of past inputs. Future input moves, however, are calculated from a linear model, by solving av. quadratic program at each sampling time. The computational burden is then comparable to the ordinary QDMC algorithm. In the case of a nonlinear state-space model, the linear model is obtained by linearizing the nonlinear model around the current state estimate. In the case of an input-output model the nonlinear model is also linearized, and a minimal state-space realization of the linear