Learning of feedforward torque for a servomechanism: empirical and model-based approaches

Hybrid feedforward-feedback controllers are a valid alternative to the standard feedback PID controllers which are usually adopted in Industrial Automation. A correct feedforward action can increase the controller performances and it can also function as a real time fault detector in order to avoid dangerous crashes, while feedback is necessary to guarantee controller stability. This thesis focuses on the problem of determining the correct feedforward torque to control a servomechanism by providing two different procedures. First, a model-based approach is discussed with the idea to evaluate, through a Gaussian process based machine learning algorithm, the proper feedforward action to steer the mechanism along a given trajectory, starting from a known training set. The realization of the ideal model through the Euler-Lagrange procedure is reported, as well as the the Gaussian process training and test procedures. Then, an empirical approach is presented, which is able to determine the mechanical parameters of the system in order to evaluate the necessary feedforward torque through the Euler dynamical equation. Finally, these procedures are tested on a physical test bench and on an specific axis motion in the Wrap 250 packaging machine from TMC. The results from these tests display an average reduction of the feedback current for both the model-based and the empirical approaches, proof that the feedforward branch is working in cooperation with the feedback control loop.