Data-driven Neural Feedforward Controller Design for Industrial Linear Motors

In this paper we consider the problem of feedforward controller design for industrial linear motors. These motors are safety-critical high-precision mechatronics systems that pose stringent requirements on the feedforward design: safe and predictable behavior for the desired motion profiles, tracking performance within the 10 mu m range in the presence of nonlinear friction and real-time implementation within the 1ms range. We investigate and compare several possibilities to design data-driven feedforward controllers using neural networks (NN) and we show that a two-step inverse estimation method is the most suitable approach, due to robustness to noisy data. We also show that basic knowledge about the system dynamics and the friction behavior can be exploited to design neural feedforward controllers with a simple structure, suitable for real-time implementation in industrial linear motors. The developed data-driven neural feedforward controllers are tested and compared with standard mass-acceleration feedforward and iterative learning controllers in realistic simulations.