Discrete Time Neural Observers PDF Download

"An observer is a dynamic system that estimates the state variables of another system using noisy measurements, either to estimate unmeasurable states, or to improve the accuracy of the state measurements. The Modified State Observer (MSO) is a technique that uses a standard observer structure modified to include a neural network to estimate system states as well as system uncertainty. It has been used in orbit uncertainty estimation and atmospheric reentry uncertainty estimation problems to correctly estimate unmodeled system dynamics. A form of the MSO has been used to control a nonlinear electrohydraulic system with parameter uncertainty using a simplified linear model. In this paper an extension of the MSO into discrete-time is developed using Lyapunov stability theory. Discrete-time systems are found in all digital hardware implementations, such as that found in a Martian rover, a quadcopter UAV, or digital flight control systems, and have the added benefit of reduced computation time compared to continuous systems. The derived adaptive update law guarantees stability of the error dynamics and boundedness of the neural network weights. To prove the validity of the discrete-time MSO (DMSO) simulation studies are performed using a two wheeled inverted pendulum (TWIP) robot, an unstable nonlinear system with unmatched uncertainties. Using a linear model with parameter uncertainties, the DMSO is shown to correctly estimate the state of the system as well as the system uncertainty, providing state estimates orders of magnitude more accurate, and in periods of time up to 10 times faster than the Discrete Kalman Filter. The DMSO is implemented on an actual TWIP robot to further validate the performance and demonstrate the applicability to discrete-time systems found in many aerospace applications. Additionally, a new form of neural network control is developed to compensate for the unmatched uncertainties that exist in the TWIP system using a state variable as a virtual control input. It is shown that in all cases the neural network based control assists with the controller effectiveness, resulting in the most effective controller, performing on average 53.1% better than LQR control alone"--Abstract, page iii.