Drivetrain Response Prediction Using AI-Based Surrogate and Multibody Dynamics Model

Abstract

Numerical models, such as multibody dynamics ones, are broadly used in various engineering applications, either as an integral part of the preliminary design of a product or simply to analyze its behavior. Aiming to increase the accuracy and potential of these models, complex mechanisms are constantly being added to existing methods of simulation, leading to powerful modelling frameworks that are able to simulate most mechanical systems. This increase in accuracy and flexibility, however, comes at a great computational cost. To mitigate the issue of high computation times, surrogates, such as reduced order models, have traditionally been used as cheaper alternatives, allowing for much faster simulations at the cost of introducing some error to the overall process. More recently, advancements in Artificial Intelligence have also allowed for the introduction of Artificial Intelligence-based models in the field of surrogates. While still undergoing development, these Artificial Intelligence based methodologies seem to be a potentially good alternative to the high-fidelity/burden models. To this end, an Artificial Intelligence-based surrogate comprised of Artificial Neural Networks as a means of predicting the response of dynamic mechanical systems is presented in this work, with application to a non-linear experimental gear drivetrain. The model utilizes Recurrent Neural Networks to accurately capture the system’s response and is shown to yield accurate results, especially in the feature space. This methodology can provide an alternative to the traditional model surrogates and find application in multiple fields such as system optimization or data mining.