Speaker
Description
We present a framework for developing non-intrusive reduced-order models (ROMs) to predict nonlinear dynamical systems' frequency response functions (FRFs), focusing on gear transmission systems. This approach addresses the challenge of modeling complex, multi-valued, and non-monotonic FRFs across a wide frequency range of up to 10 kHz. The methodology combines the Craig-Bampton method with harmonic balance analysis to efficiently generate high-fidelity snapshots. Data preprocessing uses parametric spline interpolation for consistent representation of non-monotonic responses. Autoencoders are employed to reduce the dimensionality of response data, while deep feedforward neural networks map input parameters to the reduced latent space. The methodology is validated against high-fidelity simulations of a single-DOF gear system with non-monotonic solutions and a real multi-DOF gear transmission system. The results show that the framework accurately predicts FRFs for nonlinear systems while minimizing computational costs. The proposed method can efficiently accelerate the optimization of nonlinear dynamical systems.
