Physics-guided neural networks at DSPE Conference 2023

In high-precision measurement systems, the feedforward controller is designed to compensate for all known disturbances to a system, e.g., a desired position to be tracked. To compensate for all known disturbances to a system the feedforward controller is designed as an inverse model of the system. The performance that can be achieved by exploiting a feedforward controller hinges upon the quality of this model. Hard-to-model dynamics, such as nonlinear friction effects, are often neglected and therefore not compensated for.
 
The aim is to compensate for these hard-to-model dynamic effects using a novel feedforward controller that exploits a neural network in addition to a physics-based model, also known as physics-guided neural networks (PGNN). The addition of the neural network enables unprecedented performance levels in positioning accuracies as is demonstrated on an industrial linear stage in a comparison between a conventional feedforward strategy and the PGNN-based feedforward controller. The findings, along with the improvement factor achieved in this comparison, were presented by Nard.