PHYSICS-INFORMED MACHINE LEARNING FOR MODELING AND ANALYSIS OF MECHANICAL SYSTEMS

Abstract

Physics-informed neural networks (PINNs) provide an alternative to traditional solvers for differential equations, specifically in tasks such as system identification and inverse problems. This study applies the recently popularized paradigm of PINNs for system identification and surrogate modelling in Magnetorheological (MR) dampers. A task hindered by the nonlinear behaviour, hysteresis effects and the presence of latent variables in certain interpretations of the MR damper dynamic model. An experimental setup was developed to analyze empirical data collected from MR dampers, incorporating a voltage-controlled MR damper, with motions actuated through a linear actuator, and various sensors for capturing damping forces and motion profiles. A data collection pipeline was developed and allows for synchronous data collection from a multitude of devices, and database storage. Collected data is useful for validating developed models, as well as setting up a foundation for experimental validation of novel methods in future work. A literature review was performed, highlighting the limitations of existing models and the potential of PINNs, cases of deployments, and innovations by authors within the literature. The research additionally constructs and validates a discretized state space model from estimated parameters. Overall, this research demonstrates proof of concept for the application of PINNs in mechanical systems modelled by differential equations. Results demonstrated satisfactory accuracy in parameter identification, with implications for system behaviour prediction, demonstrating the potential and limitations of PINNs in this context.