Machine Learning Applied on Hydraulic Actuator Control

This paper presents a comparison of two different types of neural networks when used in the control of a hydraulic actuator. The advantages of using hydraulic actuators are pondered when facing the nonlinearities present in their model, which difficult their control difficult. The state of the art seeks several solutions, mostly in the use of neural networks. In this way, this paper addressed a study regarding the replacement of traditional sigmoidal networks by the use of wavelet networks in the representation of friction on the walls of hydraulic cylinders and reverse valve dynamics. Different architectures are tested and trained using the quickpropagation algorithm. Finally, the efficiency of the networks is compared regarding generalization for friction and reverse dynamics of the valve, as well as their use in a cascade neural control.