Generalized feedforward control using physics-informed neural networks

The improvements in tracking performance resulting from inversion-based feedforward controllers are limited by the accuracy of the available model describing the inverse system dynamics. For this reason, the use of neural networks (NNs) as a model parameterization is growing in popularity. However, training black-box NNs to represent a general description of the inverse dynamics while respecting physical laws turns out to be troublesome, especially in situations where the training data does not cover the full domain of interest. In order to solve this, this paper adopts physics-informed neural networks (PINNs) for identification of the inverse system dynamics. Additionally, a method is proposed that enables a form of graceful degradation by having the PINN feedforward controller obey an a priori known physical model when it is operated on conditions that were not present in the training data.