Study Of Learning Feed-Forward Control Method For Linear Servo System

Based on the Liaoning Province Natural Science Foundation project named "Research of Control Methods for High Speed and High Precision Linear Servo System (No. 20022037)" and considered the characters of the controlled object, the control strategies for reducing the tracking error of the high precision permanent magnet linear synchronous motor (PMLSM) servo system are represented . And a lot of works were studied in this control strategy.Learning Feed-Forward Control (LFFC) is a form of Feedback-Error-Learning control, i.e. the control system consists of a feedback controller, and a feed-forward part that is implemented as a function approximator. The learning feed forward component is a neural network based controller , comprised of a one hidden layer structure with second order B-Spline basis functions, In LFFC, a B-Spline Network (BSN) is chosen for the feed-forward part. During operation, the feed-forward controller is trained by the output, u_C, of the feedback controller, insuch way that the tracking error, e , decreases.When the plant has to perform repetitive motions, time-indexed LFFC should be considered, i.e. the case that the periodic motion time is the only input to the feed-forward part. In this case, LFFC is similar to Iterative Learning Control (ILC). Based on the convergence results of ILC, a stability analysis of time-indexed LFFC has been performed. We have derived conservative stability conditions for the B-spline distribution and for learning rate. The simulation results prove that the proposed system has strong robustness, and LFFC reduce effective the influence of external disturbances, end-effect, friction and parameters variation, and enhance the track performance of the PMLSM servo system.In case of random reference motions, path-indexed LFFC should be applied, i.e. the reference signal and derivatives thereof should be selected as inputs for the feed-forward part. For ensuring stability and convergence, we proposed to add a stabilizing measure to the path-indexed LFFC, known as regularization. Regularization is performed by filtering the learning signal with a time-indexed BSN. The simulation results prove that this form of regularizationused in path-indexed LFFC can not only reduce effective the influence of various disturbances, but also yielded accurate and stable learning.