< Back to previous page

Project

A Physics-Informed Deep Learning Approach towards Multibody Digital Twins for State/Input/Parameter Estimation

We are entering in the digital twin era: physical components are accompanied by their digital doppelgänger with measured and numerical data flowing between the two. This brings clear benefits in engineering applications, with reduced times and higher accuracy in the design, prediction or maintenance phases.

In this case, measuring the relevant quantities is extremely important in order to have an up-to-date model which correctly represents the real system. However, whilst some quantities are relatively easy to measure, others may be difficult to obtain in an inexpensive, non-invasive manner. In mechanical systems, for instance, accelerometers are typically common and economical while for forces and torques sensors exist but they may be impractical to mount due to geometrical or economic reasons. In this case, the state-estimation discipline which combines the models with the available measurements may assist.

Nevertheless, not all models are equal: complex mechanisms are typically modelled as multibody systems, leading to the characteristic differential algebraic equations in redundant coordinates which cannot be straightforwardly employed in common estimation schemes.

In this work, a framework to reduce a generic, i.e. software and formulation agnostic, multibody model to independent coordinates and obtain ordinary differential equations is proposed.

Such methodology requires that the redundant--independent coordinate mappings are available: since this is seldom the case, it is then proposed to approximate them exploiting deep learning. A physics-informed neural network combines a reference simulation and the underlying multibody dynamics to perform the optimization avoiding over-fitting.

At this point, the attained ordinary differential equation model is embedded in a Kalman filter, precisely for the estimation of unknown input forces or torques; moreover the capacity of basing this prediction merely on acceleration information is demonstrated.

Finally, the framework is extended to build the multibody model purely from data: how to retrieve the mappings from tracked body points and approximate unknown inertia or force terms is shown.

For each of the aforementioned steps, application cases are presented as validation.

In summary, this work proposes to combine the reliability of the multibody physics information with the available measurement data by means of the approximating capabilities of deep learning, in order to obtain relevant (multibody) models and (state/input/parameter) estimations.

Date:27 Oct 2016 →  12 May 2021
Keywords:Multibody dynamics, Deep learning, Kalman filtering, Parameter estimation
Disciplines:Control systems, robotics and automation, Design theories and methods, Mechatronics and robotics, Computer theory
Project type:PhD project