Data-driven nonlinear modal identification of nonlinear dynamical systems with physics-constrained Normalizing Flows

TL;DR

This paper introduces a physics-constrained Normalizing Flows deep learning method for data-driven identification of nonlinear normal modes in complex dynamical systems, enabling modal analysis solely from measured response data.

Contribution

It develops a novel approach combining Normalizing Flows with physics constraints to identify nonlinear modes without requiring explicit system models or input-output tests.

Findings

Successfully identifies NNMs from response data.

Provides an invertible transformation between original and modal coordinates.

Captures complex nonlinear dynamics effectively.

Abstract

Identifying the intrinsic coordinates or modes of the dynamical systems is essential to understand, analyze, and characterize the underlying dynamical behaviors of complex systems. For nonlinear dynamical systems, this presents a critical challenge as the linear modal transformation, which is universal for linear systems, does not apply to nonlinear dynamical systems. As natural extensions to linear normal modes,the nonlinear normal modes (NNMs) framework provides a comprehensive representation of nonlinear dynamics. Theoretically, NNMs may either be computed numerically or analytically from the closed-form models or equations of dynamical systems, or experimentally identified from controllable input-output tests, both of which, however, are typically unknown or unavailable practically. In this study, we present a physics-integrated Normalizing Flows deep learning-based data-driven approach which identifies the NNMs and the nonlinear modal transformation function of NNMs using measured response data only. Specifically, we leverage the unique features of the Normalizing Flows model: 1) the independent latent spaces, naturally spanned by the Normalizing Flows, are exploited to facilitate nonlinear modal decomposition; 2) the invertible transformation through the Normalizing Flows, enabling efficient and accurate nonlinear transformation between original and modal coordinates transformation. Therefore, our framework leverages the independency feature and invertibility of Normalizing Flows to create a model that captures the dynamics of unknown nonlinear dynamical systems.