Machine Learning Based Model Re-Identification of Complex Dynamical Systems on Autonomous Marine Surface Vehicle

In a plethora of research, data-driven methods are lauded for their aptitude in identifying nonlinear models that can adeptly capture complex process dynamics. The Sparse Identification of Nonlinear Dynamics (SINDy) technique, introduced by Brunton et al. in 2016, has particularly received significant recognition. However, its computational demands often inhibit its applicability for rapid system re-identification of multifaceted dynamical systems. To circumvent this drawback, we present a novel bifurcated strategy that enhances SINDy's efficacy for swift online model adaptation. Initially, during the offline stage, SINDy establishes a foundational dynamical blueprint based on historical empirical or simulated data. Thereafter, we introduce the groundbreaking SINDy-D framework, specifically designed to rapidly detect model inconsistencies. To facilitate SINDy-D's deployment, we rely exclusively on discrepancies between model predictions and actual system behaviors to recognize model divergences or process anomalies. This pioneering SINDy-D mechanism proves invaluable for understanding the nuanced dynamics of autonomous surface vehicles. Empirical evaluations underscore a remarkable decrease in computational duration using our recommended approach.