Application of NN-based and MRAS-based FFC to electromechanical motion systems

Neural Network (NN) based Learning Feed-Forward Control (LFFC) is an attractive control paradigm for motion systems. The use of LFFC can improve not only the disturbance rejection, but also the stability robustness of the controlled system. One of the main drawbacks of the NN-based LFFC is the requirement that the training motions have to be chosen carefully, such that all possibly relevant input combinations are covered. This requirement may be quite restrictive in practical applications. But Model Reference Adaptive Systems (MRAS)-based LFFC can be used to overcome such problem. We address the problem relating to the precision control of permanent magnet linear motors to track random motion trajectories. By implementing both controllers on a `Tripod' setup, the performances of both methods are compared. Also the combination of the two is considered. The simulation and experimental results show that both control algorithms reach almost the same tracking error after convergence and are superior to the classic PD controller. However, after convergence the MRAS-based LFFC is able to generate a much better feed-forward control and hence obtain about a 5 times smaller maximum tracking error than the NN-based with an untrained reference motion. Moreover, MRAS-based LFFC is simpler to implement. The resulting control laws are simple and thus interesting for practical use.