SID-Net: Machine learning based system identificationframework for rigid and flexible multibody dynamics

Abstract

Modeling and simulation of dynamic systems is widely used in mechanical system design and control. System identification (SI), a process of correlation using experimental or target data, is essential for the reliability of implemented numerical models. To actualize the process, it is crucial to understand the relationship between numerous modeling parameters that affect the system responses. Modeling and simulating nonlinear systems such as multibody dynamics, is particularly difficult owing to their characteristic enormous assumptions and approximations; furthermore, the computational demands are also crucial during the iterative SI process. Therefore, we propose an effective SI framework for rigid and flexible multibody dynamics using artificial neural networks. The framework consists of two phases: a system metamodel and a SI network (SIN). Through supervised learning, a metamodel that can provide nonlinear time transient responses based on the input system design parameters in real time was generated. The loss minimization process for initializing a new neural network that, utilizes a SIN to find the input design parameters that generate the desired target output was also introduced. A new loss function was defined for minimizing the difference between the output of the trained system metamodel, following which the target output was defined. The performance of the proposed framework was evaluated in well-designed numerical and experimental examples.