Optimisation of neural state variables estimators of two-mass drive system using the Bayesian regularization method

Abstract

The paper deals with the application of neural networks for state variables estimation of the electrical drive system with an elastic joint. The torsional vibration suppression of such drive system is achieved by the application of a special control structure with a state-space controller and additional feedbacks from mechanical state variables. Signals of the torsional torque and the load-machine speed, estimated by neural networks are used in the control structure. In the learning procedure of the neural networks a modified objective function with the regularization technique is introduced. For choosing the regularization parameters, the Bayesian interpretation of neural networks is used. It gives a possibility to calculate automatically these parameters in the learning process. In this work results obtained with the classical Levenberg-Marquardt algorithm and the expanded one by a regularization function are compared. High accuracy of the reconstructed signals is obtained without the necessity of the electrical drive system parameters identification. Simulation results show good precision of both presented neural estimators for a wide range of changes of the load speed and torque. Simulation results are verified by the laboratory experiments.