Model Reduction for Aperiodically Sampled Data Systems

Mert Baştuğ, Laurentiu Hetel, Mihály Petreczky

This paper, published in 2017, tackles the challenge of simplifying complex dynamical systems that are sampled at irregular intervals—a common scenario in modern networked and embedded control systems.

TL;DR

The authors propose two moment-matching methods to reduce the order of aperiodically sampled systems while guaranteeing "quadratic stability." Their key finding is that reducing the model after accounting for the sampling behavior (Approach 2) significantly outperforms reducing the continuous-time plant first (Approach 1).

What Problem Does This Paper Address?

In the world of control engineering, we often deal with high-dimensional state-space models. These models are accurate but computationally "heavy," making real-time simulation or control synthesis difficult. To solve this, we use Model Order Reduction (MOR) to create a simpler, lower-dimensional model that behaves like the original.

However, most MOR techniques assume the system is either purely continuous or sampled at a fixed, periodic rate. In reality, modern systems (like those communicating over a shared network) often experience aperiodic sampling, where the time between samples varies. This variation turns a simple Linear Time-Invariant (LTI) system into a more complex Linear Switched (LS) system. The challenge is to reduce these switched systems without losing accuracy or, more importantly, without making the simplified system unstable.

Methods

The authors investigate two distinct workflows for reduction, both based on moment matching (matching the coefficients of the system's impulse response, known as Markov parameters):

  1. Approach 1 (Reduce then Sample): You take the original continuous-time plant, apply standard LTI model reduction to get a smaller continuous plant, and then discretize it using the allowed sampling intervals.
  2. Approach 2 (Sample then Reduce): You first discretize the original high-dimensional plant into a Linear Switched (LS) system that accounts for all possible sampling intervals. Then, you apply a specialized moment-matching algorithm designed specifically for switched systems to shrink it.

A significant portion of the paper is dedicated to a mathematical proof showing that if the original continuous-time system is stable, both approaches will produce a "quadratically stable" reduced-order model.

Datasets

The paper uses two synthetic numerical benchmarks to evaluate the methods:

  • Example 1: A stable Single-Input Single-Output (SISO) system of order 50, reduced to order 18. It uses a sampling set of four intervals: {1, 1.5, 2, 3} seconds.
  • Example 2: An unstable SISO system of order 10, reduced to order 4. It uses a smaller sampling set: {0.1, 0.15, 0.2, 0.3} seconds.

Key Insights and Ideas

  • Information Awareness: The core "aha moment" is that Approach 2 is "aware" of the sampling intervals during the reduction process. Because the sampling interval acts almost like an additional control input, incorporating that information into the reduction step leads to a much better approximation of the system's true behavior.
  • Stability Preservation: Proving stability for reduced-order switched systems is non-trivial. The authors use Lyapunov theory to show that the projection-based reduction (using a matrix $V$ that spans the reachability space) naturally preserves the quadratic stability of the original plant.
  • The "Complexity" Trade-off: While Approach 2 is more accurate, it requires a more sophisticated reduction algorithm (handling Minkowski sums of vector spaces) compared to the "off-the-shelf" LTI reduction used in Approach 1.

Key Results

The authors compared the models using the Best Fit Rate (BFR), a percentage metric where 100% is a perfect match.

  • Superior Accuracy: In the 50th-order example, Approach 2 achieved a mean BFR of 98.4%, while Approach 1 only managed 53.8%.
  • Robustness: Even when the system was subjected to white Gaussian noise and random switching between sampling intervals, Approach 2 consistently tracked the original output with high precision, whereas Approach 1 showed significant deviation.
  • Stability Confirmation: As predicted by their proofs, the reduced models remained stable throughout the simulations, even when the sampling jitter was significant.

Limitations and Future Work

  • Lack of Formal Comparison: While numerical results strongly favor Approach 2, the authors note that they do not yet have a formal mathematical proof explaining exactly why or by how much Approach 2 is superior in all cases.
  • Finite Sampling Sets: The current method assumes the sampling interval comes from a pre-specified finite set. In many real-world applications, the sampling interval might vary continuously within a range (e.g., any value between 0.1 and 0.5 seconds). Extending this to continuous intervals is a logical next step.
  • Numerical Conditioning: Like many moment-matching methods, the algorithms can become numerically sensitive (ill-conditioned) as the number of matched moments increases, which was mentioned as a general implementation caution.

Summary updated

Abstract

Two approaches to moment matching based model reduction of aperiodically sampled data systems are given. In certain cases, such systems can be represented by discrete-time linear switched (LS) state space (SS) models. One of the approaches investigated in the paper is to apply model reduction by moment matching on the linear time-invariant (LTI) plant model, then compare the responses of the LS SS models acquired from the original and reduced order LTI plants. The second approach is to apply a moment matching based model reduction method on the LS SS model acquired from the original LTI plant; and then compare the responses of the original and reduced LS SS models. It is proven that for both methods, as long as the original LTI plant is stable, the resulting reduced order LS SS model of the sampled data system is quadratically stable. The results from two approaches are compared with numerical examples.
BibTeX
@inproceedings{bastug2017model,
  author = {Mert Baştuğ and Laurentiu Hetel and Mihály Petreczky},
  title = {{Model Reduction for Aperiodically Sampled Data Systems}},
  booktitle = {IFAC-PapersOnLine},
  year = {2017},
  doi = {10.1016/j.ifacol.2017.08.1134},
}

Loading related content...

Figure