Physics-guided neural networks for high-precision mechatronics
The industry of high-precision mechatronics requires continuous innovation from a motion control perspective to facilitate the ever-increasing specifications on precision and throughput.
This motion control aims to design a system input (forces) to steer the output (position) to a desired reference. To do so, motion control relies on mathematical models that describe the dynamical behavior of the mechatronic system.
Conventionally, such a model is derived from physical knowledge of the system. However, the accuracy of such a physics-based model is limited: real-life systems exhibit complex dynamics that are difficult, if not impossible, to model using foundational knowledge. The limited accuracy of these physics-based models becomes more problematic when further increasing the specifications for precision and throughput.
Neural networks
With the aim of obtaining more accurate models, it is possible to employ neural networks. These neural networks are flexible models as they can learn much of the system dynamics from data. However, neural networks are sensitive to the training data, and often do not respect the underlying physics. This complicates the use of neural networks for safety critical applications, such as motion control of mechatronic systems.
As part of his PhD research, solved this problem by effectively merging neural networks with the available physics-based models. This yielded models with the same robustness as the physics-based model, as well as the high accuracy of neural networks. The resulting model has been denoted as the physics-guided neural network.
Application
One straightforward application of physics-guided neural networks in motion control is feedforward control.
Feedforward control reconstructs an input to the system by setting the output of the model equal the reference. Consequently, when the model describes the system more accurately, the resulting error between the system output and reference becomes smaller. It is important to define model accuracy in the context of feedforward control, and how we can use this to train the physics-guided neural networks from data.
The developed methodology was validated by Bolderman and his collaborators on multiple high-precision mechatronic systems that have applications in high-tech industries. The physics-guided neural networks demonstrated significant improvements with respect to the classical, physics-based approaches.
Title of PhD thesis: . Supervisors: Mircea Lazar and Hans Butler.