A neural observer for dynamic systems

Abstract A new class of observers is presented to estimate the states of dynamic systems, based on the rapidly growing technology of neural networks. The observer, with its parallel processing capabilities, provides a viable means for constructing, in real time, the states of complex dynamic systems from a small number of sensors. It has also the ability to learn about the system dynamics, generalize from previous learning and abstract essential characteristics from contaminated inputs. Unlike conventional Luenberger observers, where the design calculations are rather intensive and the implementation is a challenge in real time even for low order systems, the design and implementation of the neural observer is very simple and straightforward. It can equally be designed for linear as well as non-linear systems, whereas the Luenberger observer is limited only to linear systems. Numerical examples are presented to illustrate the merits of the proposed neural observer in estimating the states of simple linear and non-linear systems.