Abstract
In nonlinear system identification, the quality of the learned model is heavily dependent on the coverage of the experimental data across the system's operating space. This paper introduces a novel input design strategy that leverages Gaussian Processes (GPs) to generate "space-filling" signals. By minimizing the average GP posterior variance across a predefined set of anchor points, the proposed method ensures that the resulting data thoroughly explores the region of interest. We demonstrate that this approach provides a smooth, differentiable cost function for optimization and guarantees specific space-filling properties. Evaluations on simulated linear and nonlinear systems show that the GP-based design significantly outperforms standard techniques like Schroeder phase multisines in terms of state-space coverage.
Introduction
Building accurate mathematical models of nonlinear systems from experimental data—a process known as system identification—remains a significant challenge. While input design for linear systems is well-understood through the maximization of the Fisher Information Matrix, nonlinear systems require a different approach. Because nonlinear behavior is state-dependent, a model's performance is often limited by generalization error: the failure to accurately predict behavior in regions of the state space where data is sparse.
To address this, researchers advocate for "space-filling" inputs that force the system to visit a wide variety of state and input combinations. However, existing methods often struggle with non-smooth cost functions or are limited to specific signal types. This paper proposes a flexible, GP-based framework to design smooth, space-filling trajectories for nonlinear systems.
Methodology
The core innovation of this work is the use of Gaussian Processes as a proxy for measuring space-filling quality. Rather than modeling the system itself, a GP prior is placed on a hypothetical system model.
Anchor Points and Variance Proxy
The strategy involves defining a grid of "anchor points" within the desired region of interest (ROI). In GP regression, the posterior variance at any given point serves as a measure of uncertainty; it is high in regions far from existing data and low near data points. The authors define a cost function as the average GP posterior variance across all anchor points.
Optimization and Space-Filling Guarantees
Minimizing this cost function effectively pushes the system's trajectory to pass near the anchor points. Because the GP variance is smooth and differentiable, standard gradient-based optimization can be used to tune input signal parameters, such as the amplitudes and phases of a multisine wave.
The paper provides a mathematical proof that if the cost function reaches zero, the dataset satisfies the $\epsilon$-space filling property, meaning every point in the ROI is within a distance $\epsilon$ of a collected data point.
Experimental Evaluation
The method was tested on two simulated systems:
- Linear Time-Invariant (LTI) System: A second-order system used to validate the theoretical relationship between anchor point density and coverage.
- Nonlinear Mass-Spring-Damper System: A complex 3D model representing a mass on a rail connected to a ceiling spring. This system served as a benchmark for handling nonlinearities and spectral constraints.
Results
The GP-optimized inputs were compared against unoptimized signals and the industry-standard Schroeder phase multisine.
In the nonlinear mass-spring-damper case, the "filling distance" ($\rho$)—a metric where lower values indicate better coverage—was reduced from an initial 1.2756 to 0.3029 after optimization. This significantly outperformed the Schroeder multisine, which achieved a filling distance of 1.1811.
Discussion and Conclusion
The GP-based approach offers several advantages:
- Smooth Energy Landscape: Unlike distance-based metrics, GP variance provides a smooth surface for optimization.
- Flexibility: The method can optimize various signal classes, including multisines and piecewise constant steps.
- Distribution Agnostic: It focuses on coverage rather than forcing data into a specific probability distribution.
However, limitations remain. The computational cost of GP variance scales cubically ($O(N^3)$) with the number of data points, which may limit its use for very long experiments. Additionally, the method requires a prior model of the system to simulate trajectories during the design phase. Future work will focus on scaling the approach for high-dimensional systems and evaluating its direct impact on the accuracy of various nonlinear identification algorithms.
Summary updated