Robust Control of Unknown Switched Linear Systems from Noisy Data

Wenjie Liu, Yifei Li, Jian Sun, Gang Wang, Jie ChenarXiv 2023 · arXiv · OpenReview

TL;DR

This paper tackles the challenge of controlling complex systems that switch between different behaviors (like aircraft or engines) using only data, even when that data is noisy. They propose a novel data-driven control strategy that intelligently decides when to update its control parameters by solving optimization problems, ensuring stability despite noise and unknown system dynamics.

What Problem Does This Paper Address?

Controlling complex systems, especially those with multiple operating modes that switch between each other (known as "switched systems"), is a significant challenge. Traditionally, this requires a precise mathematical model of the system. However, for many real-world applications like aircraft or industrial processes, obtaining such models is difficult or impossible. Data-driven control offers a way to design controllers using only observed input-output data, bypassing the need for explicit system modeling.

While data-driven control has shown promise for stable systems, it faces a major hurdle when dealing with noisy data and switched dynamics. Noisy data can corrupt the information extracted from system behavior, making it hard to reliably learn the system's dynamics. Switched dynamics add another layer of complexity because the system's behavior can change abruptly, and the sequence of these changes might not be fully represented in the offline data used for initial controller design. Existing data-driven methods often struggle to maintain stability and robustness in the presence of both noise and unknown switching. This paper aims to bridge this gap by developing a robust data-driven controller for unknown switched linear systems that can handle noisy data and adapt to changing dynamics.

Methods

The core of the proposed solution is an auxiliary function-based switching control law. This strategy intelligently manages when to solve a computationally intensive semi-definite program (SDP), which is used to derive the control parameters.

Here's a breakdown of the key methodological components:

  • Data-Driven SDP: The controller relies on solving an SDP that uses collected input-state data to find a stabilizing feedback gain. This is a standard approach in data-driven control, but it's sensitive to noise.
  • Auxiliary Function for Decision Making: A novel aspect is the introduction of an auxiliary function, $V(x(k)) = x(k)^T P(k) x(k)$, which acts as a "confidence meter" for the controller. This function helps decide when the system's state is "well-behaved" enough to warrant solving the SDP.
  • Intermittent SDP Updates: Instead of solving the SDP at every single time step (which can be computationally expensive and prone to errors with noisy data), the controller only updates its parameters by solving the SDP when the auxiliary function indicates sufficient confidence. This is crucial for robustness.
  • Exciting Signal for Persistency of Excitation: To ensure that the collected data is informative enough for the SDP to be feasible, an "exciting signal" $e(k)$ is strategically injected into the control input. This signal is designed to guarantee "persistency of excitation," a condition that ensures the data adequately covers the system's dynamics.
  • Input-to-State Practical Stability (ISpS): The ultimate goal is to guarantee ISpS, a robust stability notion that accounts for external disturbances (noise) and ensures that the system state remains bounded.

Datasets

The paper does not introduce new datasets. Instead, it relies on simulated data generated from the system models described in the numerical examples.

Key Insights and Ideas

  • Smart SDP Updates are Key: The most significant insight is that not all data points are equally valuable for learning a robust controller. By using an auxiliary function to intelligently decide when to solve the computationally intensive SDP, the controller can adapt to changing dynamics and noise without sacrificing stability. This is a departure from methods that attempt to solve the SDP continuously.
  • Noise and Switching are Intertwined Challenges: The paper highlights how noise can undermine the "persistency of excitation" condition, which is essential for data-driven control. This is exacerbated in switched systems where data from different modes can further complicate learning.
  • The Power of Auxiliary Functions: The auxiliary function $V(x(k))$ acts as a crucial decision-maker, guiding the controller's behavior. Its evolution provides a measure of confidence in the learned model and control parameters.
  • Balancing Open-Loop and Closed-Loop Control: The strategy cleverly uses an open-loop exciting signal $e(k)$ only when necessary to ensure data quality, while primarily relying on the data-driven state feedback $K(k)x(k)$ for stabilization. This balance is critical for robustness.

Key Results

  • Guaranteed ISpS: The proposed controller is theoretically proven to guarantee Input-to-State Practical Stability (ISpS) for the closed-loop switched linear system. This means the system remains bounded even in the presence of bounded noise and unknown switching dynamics.
  • Feasibility Conditions for SDP: The paper provides specific conditions on the noise bound and the switching dynamics (slow switching) that ensure the feasibility of the robust data-based SDP.
  • Numerical Validation: Two numerical examples (a flight control system and an aircraft engine system) demonstrate the effectiveness of the proposed controller. These examples show that the controller can successfully stabilize the system and handle noisy data, outperforming a disturbance-free controller in the presence of noise. The figures illustrate the convergence of the state and the auxiliary function, showcasing the controller's adaptive behavior.

Limitations and Future Work

  • Slow Switching Assumption: The theoretical guarantees of ISpS are derived under the assumption that the system dynamics switch "slowly enough." This means the time between switches is sufficiently long to collect enough data from a single mode. The paper acknowledges that this assumption might limit applicability in systems with very rapid switching.
  • Bounded Noise: The analysis assumes that the noise is bounded. While this is a common assumption in robust control, it might not hold for all real-world scenarios.
  • Data Length Requirement: The controller relies on a sufficient data buffer length ($T > 2N-1$). While this is necessary for theoretical guarantees, practical implementations might benefit from exploring methods that require shorter data buffers.
  • Future Work: The authors suggest exploring extensions to handle more complex scenarios, such as systems with faster switching dynamics or different types of noise. Further investigation into reducing the data buffer length requirement would also be valuable.

Summary updated

Abstract

This paper investigates the problem of data-driven stabilization for linear discrete-time switched systems with unknown switching dynamics. In the absence of noise, a data-based state feedback stabilizing controller can be obtained by solving a semi-definite program (SDP) on-the-fly, which automatically adapts to the changes of switching dynamics. However, when noise is present, the persistency of excitation condition based on the closed-loop data may be undermined, rendering the SDP infeasible. To address this issue, an auxiliary function-based switching control law is proposed, which only requires intermittent SDP solutions when its feasibility is guaranteed. By analyzing the relationship between the controller and the system switching times, it is shown that the proposed controller guarantees input-to-state practical stability (ISpS) of the closed-loop switched linear system, provided that the noise is bounded and the dynamics switches slowly enough. Two numerical examples are presented to verify the effectiveness of the proposed controller.
BibTeX
@misc{liu2023robust,
  author = {Wenjie Liu and Yifei Li and Jian Sun and Gang Wang and Jie Chen},
  title = {{Robust Control of Unknown Switched Linear Systems from Noisy Data}},
  year = {2023},
  eprint = {2311.11300},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2311.11300},
}
Figure