(42e) Forming Distributed Estimator Networks from Decentralized Estimators

Complex, tightly integrated chemical process networks become more and more common in the modern process industry due to their economic efficiency. Such a process network usually consists of several unit operations (subsystems), which are connected together through material, energy and information flows. Because of the significant interaction between different subsystems, it poses great challenges in the design of automatic control systems. Process networks are traditionally controlled in the decentralized framework. During the past decade, much attention has been given to development of distributed process control systems (in particular, distributed model predictive control). In a distributed control framework, local controllers communicate to exchange information to collaborate their actions. The distributed framework has been demonstrated to be a promising framework for handling the scale and the interaction of these integrated chemical processes. Compared to distributed process control, its dual problem â?? distributed state estimation â?? has received relatively less attention within process control. In this work, we focus on distributed state estimation.

In the literature, the existing algorithms on distributed state estimation are primarily developed in three frameworks: the deterministic observer framework, the Kalman filter (KF) framework and the moving horizon estimation (MHE) framework. Most of these algorithms are developed for linear systems and few for nonlinear systems. A prominent feature of all these algorithms is that they all require the local observers/estimators to be of the same type (e.g., all the local estimators should be MHEs) and each of the developed algorithm can be only applied to a specific type of local observers/estimators (e.g., distributed KF algorithms cannot be used to coordinate distributed MHEs). It is desirable to develop a more general method that is capable of connecting observers/estimators of different types together while ensures overall convergence of the estimates. The motivation of this kind of method can be seen when a process includes subsystems of different types (e.g., linear and nonlinear subsystems) or subsystems posing different process/numerical constraints (e.g., estimation constraints need to be considered in some of the subsystem).

Another feature of the existing distributed state estimation algorithms is that in the design of these algorithms, no consideration is given to the potentially existing (decentralized) implementation of control/estimation algorithms in a process. If a decentralized state estimation algorithm has already been implemented in a process, the above distributed algorithms essentially require a completely redesign of the existing implementation which means high capital investment. This also motivates us to develop a general method that is capable of connecting (decentralized) observers/estimators together to improve the overall estimation performance.

Motivated by above considerations, in this work, we propose a method to connect decentralized observers/estimators (which can be of different types) to form distributed state estimation schemes. One key idea of proposed method is to develop a local compensator for each subsystem, which is used to compensate for the interactions between a subsystem and its neighbor subsystems. Each compensator is connected to the corresponding subsystem observer/estimator designed without considering interaction to form a complete local estimator. Within the proposed framework, the local estimation algorithms for different subsystems can be independently selected and each local estimator can be more easily maintained without deteriorate the structure of the entire estimation system. The proposed method is demonstrated to be effective via the application to a simulated chemical process. It is also demonstrated that the proposed method has the potential to be used in distributed output feedback control which is an issue that has not been addressed in the research of distributed model predictive control.