DC motor identification based on Recurrent Neural Networks

Abstract

The paper describes system identification by using Artificial Neural Networks that is applied to a permanent magnet DC motor. To identify its dynamic behavior an experimental setup has been developed that enables to measure data of the system input (armature voltage) and output (current and rotor speed). Generally, the identification methods can be classified as parametric and non-parametric. We use a non-parametric method (black box). A recurrent neural network was used and the Nonlinear AutoRegressive network with eXogenous inputs network (NARX) has been selected. Parallel architectures have been used in training the NARX network. The scaled conjugate gradient training algorithm, using the first and second derivatives of error to train the network to minimize the error function, has been selected. The network architecture which has been used to create the dynamic model of the motor consists of three hidden layers, a single input neuron, and two output neurons. The modeled and measured normalized data were compared with good conformity.