Deep learning-based physics-informed autoencoder for model order reduction of hydraulically actuated multibody system

Abstract

Hydraulically actuated multibody systems are widely used in heavy machinery, robotics, and industrial automation, where achieving accurate yet computationally efficient simulations re-mains a major challenge. Conventional formulations based on generalized coordinates often lead to large-scale differential-algebraic systems that are computationally intensive and are susceptible to numerical drift. This thesis introduces a deep-learning-based physics-informed autoencoder framework for nonlinear model order reduction of such systems. In the proposed framework, the mechanical subsystem is modeled using generalized coordinate formulations with Baumgarte stabilization, whereas the hydraulic subsystem is represented using lumped fluid theory. These mechanical and hydraulic subsystems are coupled through a monolithic integration scheme. The autoencoder learns a mapping between generalized and minimal coordinates, enabling efficient simulation of the system dynamics. A hybrid loss function combining reconstruction and simulation losses is utilized to ensure both geometric accuracy and physical consistency. A hydraulically driven slider-crank mechanism is used as a case example to validate the methodology adopted. The proposed framework effectively reduces model complexity while maintaining dynamic fidelity, providing robust real-time simulation and control of a hydraulically coupled multibody system.