Synthesis of Data-Driven Nonlinear State Observers using Lipschitz-Bounded Neural Networks

TL;DR

This paper introduces a data-driven method for synthesizing nonlinear state observers using Lipschitz-bounded neural networks within the Kazantzis-Kravaris/Luenberger framework, enabling robust, model-free state estimation for nonlinear systems.

Contribution

It proposes a novel neural network-based observer synthesis approach that bounds the Lipschitz constant for robustness, linking generalization loss to Lipschitz bounds and demonstrating on a chaotic system.

Findings

Effective state estimation on Lorenz system

Robustness achieved via Lipschitz constraint

Model-free neural observer synthesis demonstrated

Abstract

This paper focuses on the model-free synthesis of state observers for nonlinear autonomous systems without knowing the governing equations. Specifically, the Kazantzis-Kravaris/Luenberger (KKL) observer structure is leveraged, where the outputs are fed into a linear time-invariant (LTI) system to obtain the observer states, which can be viewed as the states nonlinearly transformed by an immersion mapping, and a neural network is used to approximate the inverse of the nonlinear immersion and estimate the states. In view of the possible existence of noises in output measurements, this work proposes to impose an upper bound on the Lipschitz constant of the neural network for robust and safe observation. A relation that bounds the generalization loss of state observation according to the Lipschitz constant, as well as the $H_2$-norm of the LTI part in the KKL observer, is established, thus reducing the model-free observer synthesis problem to that of Lipschitz-bounded neural network training, for which a direct parameterization technique is used. The proposed approach is demonstrated on a chaotic Lorenz system.