Browsing > By author > Mélot Adrien

Gears are an essential component of kinematic chains used in many mechanical systems and a wide range of technical applications. However, they may cause noise pollution. Accurately predicting vibroacoustic response is of paramount importance for the design, optimization and health monitoring of gear transmission systems. System identification is therefore needed to reach a sufficiently high level of accuracy. However, this usually comes at the cost of high computational burden.

It is widely accepted the dynamic and vibroacoustic response of geared systems is mainly induced by the gear static transmission error (STE) and time-varying mesh stiffness. These physical quantities are governed by the local contact conditions between the gear teeth. An accurate computation of these physical quantities is therefore crucial. However, this is a difficult problem as gear contact resolution is intrinsically nonlinear and multiscale.

Classical approaches rely on semi-analytical computation methods, which are computationally efficient due to their reliance on simplifying assumptions. However, this efficiency is achieved at the cost of reduced accuracy. To overcome the limitations of these assumptions, a direct finite element (FE) contact resolution can be implemented, capturing the full flexibility of the gearbox. While this method provides higher accuracy by minimizing assumptions, it results in significantly higher computational costs, often requiring several hours or even days to complete.

To address this drawback, this work introduces a novel framework based on a multifidelity deep learning approach for predicting static gear contact conditions. The proposed method involves training multiple neural networks using extensive numerical data generated by a semi-analytical approach. Additionally, a transfer learning (TL) method is employed to incorporate supplementary data from a limited number of full FE simulations into neural networks built from semi-analytical data. The developed multifidelity model offers a significant computational speedup, enabling near real-time predictions compared to FE analysis. It enables further dynamic optimization tasks that fully integrate the flexibility of the gearbox.