Artificial Intelligence Reduced Order Observer Design for Output Feedback Control

Abstract

In modern control engineering, the rising demand for streamlined state estimation in dynamic systems has made the development of reduced-order observers a critical area of study. This paper explores the integration of neural networks into observer design, utilizing their specialized ability to learn unknown dynamics and approximate complex nonlinear functions. In this paper, we propose the use of neural-network-based reduced-order observer designed whish maps the relationship between system inputs, measurable outputs, and hidden state variables. The method uses the physics-in-formed neural networks formulation, which provides a powerful paradigm for solving partial differential equations for large dimensions of dynamical systems. The method facilitates precise estimation even when a comprehensive analytical plant model is unavailable. By utilizing a reduced-order architecture, the framework minimizes computational overhead, making it ideal for real-time practical applications. Furthermore, the approach exhibits high resilience against parameter uncertainties and system nonlinearities. Simulation data confirms that this intelligent observer configuration delivers significant accuracy and strong convergence performance.