Neural Network Control by Error-Feedback Learning for Hydrostatic Transmissions with Disturbances and Uncertainties

Abstract

This paper presents a decentralized control approach based on a neural network for a hydrostatic transmission. The bent-axis angle of the hydraulic motor is adjusted by a pure feedforward control law based on identified physical parameters, whereas the corresponding motor angular velocity is controlled using a combination of a generalized proportional-derivative (PD) controller and a multilayer perceptron with one hidden layer that is trained by an error-feedback learning approach and uses only measurable input variables. In this observer-free control structure, the neural network learns the inverse dynamics by minimizing the PD controller output and, as a consequence, an accurate tracking of the desired trajectory is achieved. As no physical modelling is required for the motor velocity control design, it can be considered as model-free. The tracking performance shows the robustness of the overall control structure for the hydrostatic transmission despite disturbances and uncertainties. The proposed control scheme is investigated by simulations first. Second, experimental results are presented taken from a dedicated test rig at the Chair of Mechatronics, University of Rostock. Finally, an experimental comparison with results from previous work is provided.