Phase correction for Learning Feedforward Control

Intelligent mechatronics makes it possible to compensate for effects that are difficult to compensate for by construction or by linear control, by including some intelligence into the system. The compensation of state dependent effects, e.g. friction, cogging and mass deviation, can be realised by learning feedforward control. This method identifies these disturbing effects as function of their states and compensates for these, before they introduce an error. Because the effects are learnt as function of their states, this method can be used for non-repetitive motions. The learning of state dependent effects relies on the update signal that is used. In previous work, the feedback control signal was used as an error measure between the approximation and the true state dependent effect. If the effects introduce a signal that contains frequencies near the bandwidth, the phase shift between this signal and the feedback signal might seriously degenerate the performance of the approximation. The use of phase correction overcomes this problem. This is validated by a set of simulations and experiments that show the necessity of the phase corrected scheme.