A Faster and Cheaper Method of Implementing States Observers using Artificial Neural Networks

Abstract

In modern control theory, a system is modeled as a functional block that receives an input and produces a corresponding output that is based on the internal states of the system. However, the internal states are generally not directly accessible to due to high costs and physical limitations; hence they are estimated from an observer. The states observer is a mathematical model that estimates the unavailable internal states from the system input and output data. The majority of the differential equation systems encountered in control applications are ordinary differential equations (ODE). There have been previous attempts to use artificial neural networks to solve the ODE, but they were never implemented as multi-dimensional state observers in particular. Moreover, these attempts have high memory requirements and also required powerful general purposed processors to perform the required computation which can be rather costly to implement. This paper proposes solve ODE for a states observer using an artificial neural network rather than general purpose processor. The use of artificial neural networks to solve ordinary differential equations in the observer takes advantage of a set of unique constraints and opportunities that earlier proposals overlooked. This implementation appears to be a faster and cheaper method and has the potential application of solving differential models on real-time embedded system such as smart sensors and self-diagnostic systems.