A Deep Neural Network-Based High Performance Robust Position Controller for Servomechanisms

Abstract

A new high-performance deep neural network-based robust quasi-time optimal servomechanism (DNN-RQTOS) method for position control is proposed in this study. The controller is trained by a dataset produced by using Robust Quasi-Time-Optimal Servomechanism (RQTOS) method. RQTOS approach is founded on discrete-time sliding mode position control (DSMC) method. This approach is implemented to a vector-controlled induction motor servo-system, utilizing quasi-optimal time-varying sliding surfaces to provide both robust and high performance control. The dataset used to train the new DNN-based controller is created in a wide range of uncertainty and position references. The DNN-based controller successfully learns the behavior of the fixed structure RQTOS controller and reveals superior performance. Specifically, the new DNN-based controller offers better transient response and load torque rejection under parameter uncertainty, compared to the sliding mode controller from which the training data were produced. Error and performance metrics of DNN-RQTOS are seen better than those of RQTOS. In addition, the chattering behaviors, which is the major drawback of sliding mode control, has been effectively eliminated.