Source-linked AI summary

A trajectory-based framework for data-driven system analysis and control

Julian Berberich, Frank Allgöwer

arXiv:1903.10723v4eess.SY

TL;DR

The paper addresses how measured data can support rigorous system analysis and control without explicitly identifying a model. It translates single-trajectory representation into classical state-space control, extends it to certain nonlinear systems, and applies the extension to data-driven simulation with kernel methods. The resulting framework constructs unknown-system trajectories from measured data and reports good noisy-output estimation in the kernel-based simulation example.

  • Problem

    Data-driven control needs rigorous methods that avoid relying exclusively on identified models while addressing limitations of some learning-based approaches, including data demands, reproducibility, and guarantees.

  • Method

    The paper characterizes system trajectories from a single measured input-output trajectory, extends the construction to systems linear in suitable known nonlinear coordinates, and uses kernel methods for simulation.

  • Results

    All trajectories of an unknown LTI system can be constructed from one persistently exciting measured trajectory, and the nonlinear extension supports data-driven simulation with good output estimates under noise.

  • Takeaways & Limitations

    A measured trajectory can contain the information needed for system analysis, controller design, and certain data-driven simulation tasks without explicit model identification.

Abstract

from arXiv · show

The vector space of all input-output trajectories of a discrete-time linear time-invariant (LTI) system is spanned by time-shifts of a single measured trajectory, given that the respective input signal is persistently exciting. This fact, which was proven in the behavioral control framework, shows that a single measured trajectory can capture the full behavior of an LTI system and might therefore be used directly for system analysis and controller design, without explicitly identifying a model. In this paper, we translate the result from the behavioral context to the classical state-space control framework and we extend it to certain classes of nonlinear systems, which are linear in suitable input-output coordinates. Moreover, we show how this extension can be applied to the data-driven simulation problem, where we introduce kernel-methods to obtain a rich set of basis functions.

I. INTRODUCTION

The paper develops a data-driven control framework that uses measured trajectories directly rather than explicitly identified models. It translates behavioral-system results into classical control and extends them toward nonlinear systems and further applications.

  • Data-driven control seeks rigorous alternatives to model-based methods and to learning approaches that often require much data and lack reproducibility or closed-loop guarantees.
  • The framework characterizes all trajectories of an unknown system using a single measured input-output trajectory.
  • Behavioral systems theory is especially suited to purely data-driven analysis because it defines systems through their spaces of trajectories.
  • Prior work applies the trajectory-based result to data-driven MPC, stochastic analysis, power systems, stability, robustness, and closed-loop parametrization.
  • The paper translates the main result to classical state-space control, improves persistence-of-excitation requirements by weaving trajectories, extends it to nonlinear systems, and addresses data-driven simulation.

II. SETTING

The setting defines finite input-output trajectories of discrete-time LTI systems and introduces persistent excitation as the central data condition. The framework assumes measured data are available while the unknown system order is known only through a potentially rough upper bound.

  • Persistent excitation: Persistent excitation of order L requires the associated Hankel matrix to have full rank, rank(H_L(x)) = nL.
  • Persistent excitation: Persistent excitation requires sufficient signal length, not merely linearly independent time-shifts.
  • Data assumptions: The measured input-output trajectory may come from simulation or experiment, while the unknown system order is assumed known only through a potentially rough upper bound.
  • Trajectory and system setting: An input-output sequence is a trajectory when it is generated by an LTI realization from an initial condition and state sequence.
  • Trajectory and system setting: LTI trajectory sets form vector spaces, whose bases can be formed by time-shifts of one measured trajectory under persistent excitation.

III. TRAJECTORY-BASED REPRESENTATION OF LINEAR

The linear result represents every trajectory of an unknown LTI system using one persistently exciting measured trajectory, equivalently through a data-dependent Hankel matrix. Multiple trajectories can be woven together to extend the usable horizon, but persistence of excitation imposes a data-length limit.

  • Classical reformulation: The paper reformulates the behavioral trajectory-characterization theorem in the classical state-space control framework.
  • Single-trajectory representation: Every length-L trajectory of an unknown LTI system can be constructed from a single measured trajectory when the input is persistently exciting of order L+n.
  • Single-trajectory representation: The trajectory space equals the range of a data-dependent Hankel matrix, so the measured trajectory acts as a system representation without explicit model identification.
  • Persistence-of-excitation limitation: Theorem 3 requires N ≥ (m+1)(L+n)−1 to span trajectories of length L, limiting the horizon supported by a finite dataset.
  • Trajectory weaving: Multiple trajectories can be woven into an arbitrarily long trajectory when their internal states align over at least n steps at each intersection.

IV. TRAJECTORY-BASED REPRESENTATION OF NONLINEAR

The paper extends trajectory-based representation beyond LTI systems to Hammerstein, Wiener, and more general systems that become linear in known nonlinear input-output coordinates. These extensions provide an alternative to identifying such nonlinear systems from data.

  • The nonlinear extension covers Hammerstein, Wiener, and other systems that are linear in suitably chosen and known input-output coordinates.
  • For these systems, a single measured trajectory can represent the system as an alternative to identification.

A. Hammerstein systems

The paper extends trajectory-based characterization from LTI systems to Hammerstein systems by transforming the static nonlinearity into known basis-function coordinates. Under persistence of excitation, measured data can represent system trajectories without identifying the nonlinear system.

  • A. Hammerstein systems: A Hammerstein system combines a static nonlinearity with a subsequent LTI system.
  • A. Hammerstein systems: The nonlinear input transformation ψ is represented using r known basis functions, producing an auxiliary input trajectory v.
  • A. Hammerstein systems: The transformed system can be treated as an LTI map from the auxiliary input v to the output y.
  • A. Hammerstein systems: If v is persistently exciting of order L + n, a length-L trajectory is characterized by a linear combination of time-shifts of the measured trajectory.
  • A. Hammerstein systems: The known basis-function representation requires N ≥ (r + 1)(L + n) − 1, limiting how many basis functions can be used explicitly.

B. Wiener systems

The paper applies the trajectory-based approach to Wiener systems by transforming outputs through an inverse static nonlinearity. This yields a corresponding trajectory representation, but with a weaker converse and more difficult basis-function selection than in the Hammerstein case.

  • B. Wiener systems: A Wiener system consists of an LTI system followed by a static nonlinearity.
  • B. Wiener systems: The inverse output nonlinearity φ^-1 is assumed to admit a decomposition using q known basis functions, which define an auxiliary output trajectory z.
  • B. Wiener systems: The transformed output z serves as the output of an equivalent LTI system for applying the trajectory-based argument.
  • B. Wiener systems: Under persistence of excitation of u of order L + n, the result gives a trajectory condition expressed through a coefficient vector α.
  • B. Wiener systems: Unlike the Hammerstein case, the Wiener result has no limit on the number of basis functions, but selecting basis functions for φ^-1 is more difficult.
  • B. Wiener systems: The “only if” direction generally fails because the map from u to z is not necessarily linear.

V. DATA-DRIVEN SIMULATION

The data-driven simulation problem computes an unknown system’s output from a prescribed input using only a measured input-output trajectory. The paper extends this trajectory-based approach to nonlinear systems through lifted coordinates and kernel methods, with regularization improving noisy simulation accuracy.

  • Trajectory-based simulation: Data-driven simulation predicts an unknown system’s output for a given input using only a previously measured input-output trajectory.The approach avoids requiring an explicit model.
  • Trajectory-based simulation: A persistently exciting input trajectory spans arbitrary LTI trajectories, allowing coefficients solved from the prescribed input to generate the predicted output.The remaining output is computed as ¯y = HL(y)α.
  • Kernel reformulation: Kernel methods implicitly provide rich basis functions through inner products, avoiding explicit construction of the nonlinear coordinates.A selected kernel can imply the basis functions used for the nonlinearities.
  • Nonlinear extension: The method extends simulation to Hammerstein-Wiener systems by representing trajectories in lifted input-output coordinates.For these systems, trajectory spaces are spanned by Hankel matrices of transformed inputs and outputs.
  • Numerical example: With 5% multiplicative output noise, a squared exponential kernel and λ = 10 produced a good estimate for a length-50 random input, whereas omitting regularization or using fixed polynomial bases significantly worsened accuracy.The experiment used N = 1000 measured samples and zero initial conditions.

VI. CONCLUSION

The paper presents a purely data-driven framework in which a single measured trajectory supports system analysis and controller design without explicit model identification. It extends the framework to certain nonlinear systems and applies the extension to kernel-based data-driven simulation.

  • Conclusion: All trajectories of an unknown system can be constructed from a single measured trajectory, which supplies the information needed for analysis and controller design without explicit model identification.The framework is first described in the classical control setting.
  • Conclusion: The nonlinear extension applies to certain systems that are linear in suitable input-output coordinates and supports kernel-based data-driven simulation.The paper identifies further applications and connections to kernel methods as future research directions.
Loading 1903.10723v4…