AI-Based Surrogate Models for Multibody Dynamics Systems

Abstract

Abstract Advancements in computer sciences and technology allow for implementation of detailed numerical models of a system such as the Finite Element (FE) or Multibody Dynamics (MBD) models. Complex mechanical systems can easily be modelled in detail, yielding accurate results. This opportunity provided by these high-fidelity numerical models has led to the broad application of such methods in development and prototyping of mechanical systems, their optimization and fault analysis and so on. The capability of detailed modelling however usually comes at a great computational cost, with the simulation time needed for a problem in many cases rising exponentially, rendering these models impractical. This problem becomes even more profound when one considers the recent integration of model-based data in data-driven methods where a large number of datasets is usually required, and multiple iterations of the same model must be simulated in order to produce the desired number of samples. To mitigate these short-comings, surrogate modelling has been extensively used in applications including large systems or repetitive runs in the form of Reduced Order Models (ROMs) to reduce the computations time and render these simulation-driven methods more viable. Use of these ROMs however is limited to cases where low loss of information is ensured, and the features lost due to the model simplification are insignificant. The developments in Artificial Intelligence (AI) and its applications have demonstrated its potential to accurately describe the relationships between a model’s inputs and outputs and as such using an AI algorithm as a surrogate model is a promising alternative. A properly trained AI algorithm can usually fit to FE and MBD models and yield accurate results at a fraction of the computational burden. To this end, an AI-based surrogate modelling framework is proposed in this work, with application on an experimental gear drivetrain system. A detailed MBD of the actual system is initially constructed and optimized via a black box optimization method in order to better simulate the physical system. A variety of supervised AI algorithms such as regression models and Convolutional Neural Networks (CNNs) is then examined as a surrogate to the various mechanisms of the system, aiming to replace them with the goal of reducing the simulation time while maintaining the high accuracy and fidelity of the original model. The various algorithms are then compared in terms of time reduction and accuracy both to each other and to the initial MBD model in order to conclude to the best suited for the application. The results are also compared to the measured response data of the physical system to ensure the validity of the models and prove the viability of the proposed method through its use on a relatively complex model. The proposed framework provides an alternative to the commonly used ROM methods and the presented application acts as a benchmark case for its implementation to more complex systems and different operating conditions.