Recovering nonlinear dynamics from non-uniform observations: A physics-based identification approach with practical case studies

Abstract

Uniform and smooth data collection is often infeasible in real-world scenarios. In this paper, we propose an identification framework to effectively handle the so-called non-uniform observations, i.e., data scenarios that include missing measurements, multiple runs, or aggregated observations. The goal is to provide a general approach for accurately recovering the overall dynamics of possibly nonlinear systems, allowing the capture of the system behavior over time from non-uniform observations. The proposed approach exploits prior knowledge by integrating domain-specific, interpretable, physical principles with black-box approximators, proving significant flexibility and adaptability in handling different types of non-uniform measurements, and addressing the limitations of traditional linear and black-box methods. The description of this novel framework is supported by a theoretical study on the effect of non-uniform observations on the accuracy of parameter estimation. Specifically, we demonstrate the existence of upper bounds on the parametric error resulting from missing measurements and aggregated observations. Then, the effectiveness of the approach is demonstrated through two case studies. These include a practical application with missing samples, i.e., the identification of a continuous stirred-tank reactor using real data, and a simulated Lotka-Volterra system under aggregated observations. The results highlight the ability of the framework to robustly estimate the system parameters and to accurately reconstruct the model dynamics despite the availability of non-uniform measurements.