Retrieving highly structured models starting from black-box nonlinear state-space models using polynomial decoupling

Research output: Contribution to journalArticle

Abstract

Nonlinear state-space modelling is a very powerful black-box modelling approach. However powerful, the resulting models tend to be complex, described by a large number of parameters. In many cases interpretability is preferred over complexity, making too complex models unfit or undesired. In this work, the complexity of such models is reduced by retrieving a more structured, parsimonious model from the data, without exploiting physical knowledge. Essential to the method is a translation of all multivariate nonlinear functions, typically found in nonlinear state-space models, into sets of univariate nonlinear functions. The latter is computed from a tensor decomposition. It is shown that typically an excess of degrees of freedom are used in the description of the nonlinear system whereas reduced representations can be found. The method yields highly structured state-space models where the nonlinearity is contained in as little as a single univariate function, with limited loss of performance. Results are illustrated on simulations and experiments for: the forced Duffing oscillator, the forced Van der Pol oscillator, a Bouc-Wen hysteretic system, and a Li-Ion battery model.

Original languageEnglish
Article number106966
Number of pages23
JournalMechanical Systems and Signal Processing
Volume146
DOIs
Publication statusPublished - 1 Jan 2021

Keywords

  • Black-box modelling
  • Model reduction
  • Multivariate polynomial decoupling
  • Nonlinear system identification

Fingerprint Dive into the research topics of 'Retrieving highly structured models starting from black-box nonlinear state-space models using polynomial decoupling'. Together they form a unique fingerprint.

Cite this