RBF Neural Networks Model-Free Feedforward Control of Piezoelectric Actuator

Piezoelectric actuator is widely utilized in ultra-precision position fields. However, it is challenged by difficulty modeling dynamic properties, such as hysteresis, creep, vibration, etc. Therefore, this paper presents a model-free feedforward control strategy of Radial Basis Functions Neural Network (RBFNN) for piezoelectric actuator. RBFNN is used to estimate the inverse dynamic model of the piezoelectric actuator, and avoiding the difficult modeling and impossible to invert problems. This feedforward controller has model-free advantage that it does not depend on the specific physics-based model and is only data-driven by references and control signals. The parameters of RBFNN are training by offline methods and online learning strategy especially. The offline learning method can compensate the trained trajectory with high precision, but for the untrained trajectory, the compensation accuracy is insufficient. For unknown target trajectories, the online learning strategy is proposed, which utilizes current data for parameter iteration and updating to improve control accuracy. Finally, simulation verify that the proposed model-free feedforward control strategy is effective. Especially the online learning strategy has the adaptive ability and generalization ability for unknown trajectory targets.