Introduction
Ensuring the stability of control systems under uncertainty is a cornerstone of robust control engineering. Traditionally, this is achieved by calculating the structured singular value ($\mu$), a metric that quantifies the smallest uncertainty required to make a system unstable. However, exact calculation of $\mu$ is NP-hard, leading engineers to rely on upper and lower bounds derived from precise mathematical models. In many modern applications, such models are either unavailable or prohibitively complex. This paper presents a novel, fully data-driven approach to estimate the lower bound of $\mu$ directly from experimental input-output data, bypassing the need for an explicit nominal model.
Methodology: Data-Driven Power Iteration
The proposed method adapts the classic power iteration algorithm—typically used with frequency-domain models—to the time-domain experimental setting. The algorithm treats the physical system as a "black box," iteratively redesigning input signals based on previous outputs to focus energy on the system's most vulnerable frequencies.
A significant challenge in data-driven $\mu$ estimation is the requirement for the system's "adjoint" (the complex conjugate transpose). Since physical systems cannot be "transposed," the authors employ a mathematical technique to simulate the adjoint response by processing experimental data backwards in time. The integration of Discrete Fourier Transforms (DFT) allows the algorithm to transition between the time-domain execution and the frequency-domain definition of $\mu$.
Experimental Validation
The method was validated using randomly generated Multi-Input Multi-Output (MIMO) systems of varying dimensions (up to $n=8$) and eight distinct uncertainty structures, ranging from simple scalar gains to complex full-block interactions. Results were benchmarked against the mussv command in MATLAB’s Robust Control Toolbox, the industry standard for model-based $\mu$ analysis.
Key Insights and Results
The data-driven approach demonstrated high accuracy, with lower bounds matching model-based benchmarks almost perfectly in noise-free environments. Convergence was typically achieved within 15 to 30 iterations, making the method practical for real-world testing.
A key observation is the "frequency concentration" effect: as iterations progress, the test signal's energy naturally clusters around the critical frequency where the system is most susceptible to instability. This behavior is illustrated in the frequency response comparison below.
Furthermore, the algorithm's performance varies with the uncertainty structure. It excels with "full block" or complex uncertainties, which provide more directional information during experiments. Conversely, repeated scalar uncertainties can lead to less reliable convergence, a limitation shared with traditional model-based power iteration methods.
Limitations and Future Directions
While this research successfully establishes a data-driven lower bound for $\mu$, several challenges remain:
- Upper Bound Estimation: Robust stability certification requires an upper bound to guarantee safety. Developing a data-driven counterpart for the upper bound is a critical next step.
- Noise Sensitivity: The current iteration is sensitive to measurement noise. Future work may incorporate instrumental variables to mitigate noise-induced bias.
- Transient Dynamics: The method assumes steady-state conditions. For systems with slow dynamics, long experimental durations may be required to avoid transient errors.
In conclusion, this work demonstrates that experimental iterations can effectively reveal stability margins previously thought to require complex mathematical models, paving the way for more resilient, model-free control systems.
Summary updated