Abstract
This paper establishes a formal, input-centric definition of "persistence of excitation" (PE) for discrete-time linear switched systems. By linking PE to realization theory and Markov parameters, the authors provide a framework for designing switching signals and inputs that guarantee unique system identification from output data. This approach moves beyond previous definitions that relied on unobservable internal states or post-hoc data analysis, offering a proactive method for experiment design in control theory.
1. Introduction
In the field of system identification, the quality of a mathematical model depends heavily on the "richness" of the experimental data. This richness, known as Persistence of Excitation (PE), ensures that the input signal is sufficiently complex to reveal the full dynamics of the system. While PE is well-understood for standard linear systems, it presents significant challenges for Linear Switched Systems (LSS)—systems that transition between different linear behaviors, such as a vehicle shifting gears.
Existing research often defines PE based on measured input-output data or internal states, which necessitates a "post-hoc" validation of experiments. This paper addresses the fundamental gap of defining and designing input signals (both switching sequences and continuous inputs) that are guaranteed to be exciting enough to identify an LSS before the experiment is conducted.
2. Theoretical Framework and Realization Theory
The authors bridge the gap between PE and realization theory by focusing on the system's Markov parameters, which uniquely capture its input-output behavior. The core logic follows three steps:
- The "DNA" of a system's response is contained within its Markov parameters.
- An input is defined as "persistently exciting" if the resulting system response allows for the calculation of these parameters.
- Once Markov parameters are obtained, standard algorithms, such as the Ho-Kalman algorithm, can be employed to reconstruct the state-space model.
This input-only focus transforms PE from a data validation metric into a powerful tool for experiment design, independent of the internal state measurements.
3. Algorithmic Contributions
The paper introduces two primary algorithmic strategies for achieving PE in different system classes:
3.1 Finite PE for Reversible Systems (Algorithm 3)
For systems where dynamics can be inverted (reversible systems), the authors propose a "reset" strategy. Because the state of an LSS carries over between modes, the response of a current mode is often "polluted" by the history of previous modes. Algorithm 3 uses calculated inputs to drive the system state back to zero after testing a specific sequence, effectively resetting the system without a physical reset button.
3.2 Infinite PE for Stable Systems (Algorithm 6)
For stable systems, the authors demonstrate that if a switching signal is "rich" enough to visit every possible mode sequence infinitely often and the continuous input is "colored noise," the system can be identified with arbitrary precision. In these cases, the influence of the initial state naturally fades over time, allowing for asymptotic identification.
4. Numerical Evaluation
The theoretical framework was validated using simulated 3rd-order switched systems featuring two or four discrete modes. These simulations served as a proof-of-concept for the proposed algorithms. In one specific instance (Example 3), the authors utilized a binary white noise switching signal and a uniformly distributed input. After processing $2 \cdot 10^7$ data points, the identification algorithm achieved a Best Fit Rate (BFR) of 99.5%, successfully recovering the system matrices even when one subsystem was non-minimal.
5. Key Insights
- Combinatorial Complexity: Identifying an LSS requires more than just visiting every mode; it requires testing specific combinations of modes. The paper formalizes the necessary combinations based on the system's dimension.
- Necessity vs. Sufficiency: The authors argue that their definition is a necessary condition for any identification algorithm to function, whereas previous definitions were often merely sufficient for specific algorithms.
- The Reset Trick: The ability to "neutralize" system history through specific input design is a critical advancement for identifying switched dynamics in finite time.
6. Limitations and Future Work
Despite the theoretical success, several challenges remain. The length of a guaranteed PE input sequence grows exponentially with the system dimension ($n$), making it impractical for very large systems. Additionally, the current proofs are constrained to reversible or stable systems. Extending these results to unstable, non-reversible systems and improving robustness against noise in finite-sequence constructions are primary areas for future research.
Summary updated