Abstract
This paper presents two approaches to moment-matching-based model reduction for aperiodically sampled data systems. Aperiodic sampling, where the interval between samples varies within a pre-specified finite set, allows these systems to be represented as discrete-time linear switched (LS) state-space models. The first approach reduces the continuous-time linear time-invariant (LTI) plant before sampling, while the second reduces the LS model derived from the original plant. Both methods guarantee quadratic stability for the reduced models if the original plant is stable. Numerical results demonstrate that reducing the model after accounting for sampling behavior (Approach 2) yields significantly higher accuracy.
Introduction
Modern networked and embedded control systems often operate under aperiodic sampling, where communication delays or shared network resources cause irregular intervals between data samples. While high-dimensional state-space models provide accuracy, their computational cost hinders real-time applications. Model Order Reduction (MOR) aims to simplify these models. However, standard MOR techniques typically assume fixed sampling rates. This paper addresses the challenge of reducing systems with varying sampling intervals without compromising stability or accuracy.
Methodology: Two Approaches to Reduction
The authors explore two distinct workflows for reducing the order of aperiodically sampled systems using moment matching—matching the coefficients of the system's impulse response.
- Approach 1 (Reduce then Sample): The original continuous-time LTI plant is first reduced to a lower-order LTI model. This reduced model is then discretized based on the allowed sampling intervals to form a linear switched system.
- Approach 2 (Sample then Reduce): The original high-dimensional plant is first discretized into a complex LS system that captures all possible sampling behaviors. A specialized moment-matching algorithm for switched systems is then applied to produce the final reduced model.
A key theoretical contribution is the proof that both methods preserve "quadratic stability." Using Lyapunov theory, the authors show that if the original continuous-time system is stable, the resulting reduced-order switched models will also remain stable under any switching sequence within the defined sampling set.
Comparative Performance
The methods were evaluated using two numerical benchmarks: a 50th-order stable system reduced to 18th-order (Example 1) and a 10th-order system reduced to 4th-order (Example 2). Performance was measured using the Best Fit Rate (BFR), where 100% represents a perfect match to the original system's output.
As shown in the results, Approach 2 consistently outperforms Approach 1. In Example 1, Approach 2 achieved a BFR of 98.42%, compared to only 53.83% for Approach 1. This disparity arises because Approach 2 is "aware" of the sampling intervals during the reduction process. By treating the sampling interval as an inherent part of the system dynamics during reduction, it achieves a much more faithful approximation of the true behavior.
Conclusion
The study concludes that while reducing the continuous plant first (Approach 1) is simpler to implement using standard tools, reducing the switched system (Approach 2) is far superior for maintaining accuracy in aperiodically sampled environments. Future work may extend these methods to systems where sampling intervals vary continuously within a range, rather than being restricted to a finite set.
Summary updated