Deep learning of multibody minimal coordinates

Abstract

• Multibody models typically include redundant coordinates and constraints leading to differential algebraic equations (DAEs). • An autoencoder neural network is trained to find the multibody minimal coordinates. • The obtained neural network function is used for the non-linear model order reduction eliminating the constraints and leading to ordinary differential equations (ODEs). • In order to avoid overfitting, the multibody physics information is embedded in the training. • This permits to combine the obtained reduced order model with common integration, control or estimation algorithms such as the Kalman filter.