On the design of persistently exciting inputs for data-driven control of linear and nonlinear systems

Mohammad Alsalti, Victor G. Lopez, Matthias Albrecht Müller

Abstract

In the context of data-driven control, persistence of excitation (PE) of an input sequence is defined in terms of a rank condition on the Hankel matrix of the input data. For nonlinear systems, recent results employed rank conditions involving collected input and state/output data, for which no guidelines are available on how to satisfy them a priori. In this paper, we first show that a set of discrete impulses is guaranteed to be persistently exciting for any controllable LTI system. Based on this result, for certain classes of nonlinear systems, we guarantee persistence of excitation of sequences of basis functions a priori, by design of the physical input only.

Introduction

Data-driven control aims to design controllers directly from experimental data, bypassing the need for complex mathematical modeling. However, the success of this approach hinges on the quality of the data: it must be "rich" enough to capture the system's underlying dynamics. This property is known as Persistence of Excitation (PE).

While PE is well-understood for linear systems, nonlinear systems present a significant challenge. Currently, researchers often collect data first and check its quality a posteriori. If the data fails the necessary rank conditions, the experiment must be repeated. This paper addresses this gap by providing a priori guidelines to design input sequences that guaranteed PE for linear, Hammerstein, and SISO flat nonlinear systems.

Methodology: Structured Excitation

The authors move away from traditional random noise, proposing structured, sparse input sequences—often timed pulses—to guarantee PE. Their approach covers three main system categories:

  1. Linear Time-Invariant (LTI) Systems: The authors prove that a simple sequence of discrete impulses (a pulse every $L$ steps) is sufficient to satisfy PE conditions, offering a more sample-efficient alternative to continuous noise.
  2. Hammerstein Systems: By selecting specific pulse magnitudes that ensure a "basis function matrix" is invertible, the authors guarantee PE for systems where nonlinearities act on the input before entering a linear stage.
  3. SISO Flat Nonlinear Systems: For these complex systems, the authors utilize monomial basis functions and a "mosaic Hankel matrix" approach. By systematically probing the system's reachability space, they ensure the resulting data matrix maintains full rank.
The structured mosaic Hankel matrix W used for nonlinear system identification.

Key Insights

  • Structure over Randomness: Structured pulses are mathematically guaranteed to work where random noise might fail, which is critical for systems with bandwidth constraints or sensitive delivery requirements (e.g., medical applications).
  • Collective PE: To prevent unstable nonlinear systems from diverging during long experiments, the authors use "collective persistence of excitation," which aggregates data from multiple short experiments.
  • Inverting the Basis: The fundamental mechanism for nonlinear systems involves ensuring that input magnitudes, when processed through basis functions (like $u, u^2, u^3$), result in an invertible matrix.

Experimental Results

The proposed methods were evaluated using numerical simulations of a second-order SISO flat nonlinear system, comparing the designed input against a standard uniform random distribution.

Comparison of PE levels and stabilization errors across different controllers.

As shown in the results table, the designed input (Controller iii) significantly outperformed the baselines:

  • Higher Information Content: The minimum singular value ($\sigma_{min}$) of the data matrix—a proxy for data richness—was 0.0399 for the designed input, compared to just 0.0083 for the random input.
  • Superior Control: The richer data led to more accurate control, with an average stabilization error for the first state ($x_1$) of 0.0553, compared to 0.0620 for the random baseline.
  • Safety: The structured approach resulted in fewer unstable experimental runs, suggesting it is a safer choice for sensitive nonlinear dynamics.

Limitations and Future Work

While providing a robust framework for a priori PE design, the method currently requires knowledge of the system's basis functions and assumes local reachability at the origin. While the LTI and Hammerstein results are general, the specific design for flat systems is currently limited to SISO cases. Future research will focus on extending these guarantees to Multi-Input Multi-Output (MIMO) nonlinear systems and investigating performance under heavy measurement noise.

Summary updated

Abstract

In the context of data-driven control, persistence of excitation (PE) of an input sequence is defined in terms of a rank condition on the Hankel matrix of the input data. For nonlinear systems, recent results employed rank conditions involving collected input and state/output data, for which no guidelines are available on how to satisfy them a priori. In this paper, we first show that a set of discrete impulses is guaranteed to be persistently exciting for any controllable LTI system. Based on this result, for certain classes of nonlinear systems, we guarantee persistence of excitation of sequences of basis functions a priori, by design of the physical input only.
BibTeX
@inproceedings{alsalti2023design,
  author = {Mohammad Alsalti and Victor G. Lopez and Matthias Albrecht Müller},
  title = {{On the design of persistently exciting inputs for data-driven control of linear and nonlinear systems}},
  booktitle = {IEEE Control. Syst. Lett},
  year = {2023},
  url = {https://openreview.net/forum?id=RQDJkATUPW},
}
Figure