Torque estimation in switched reluctance motor drive using artificial neural networks

Measurement of torque developed by an electric machine is an essential part in many industrial applications. However, commercially available torque transducers are expensive and not very accurate at high speeds. Elimination of the external torque sensor is an attractive issue in many applications. An electric machine processes its input current based on its physical characteristics and converts it to electromagnetic torque and hence can be considered as a current to torque transducer. However obtaining the mapping between electrical and mechanical terminals of an electric machine by analytical tools is almost impossible because of the nonlinearities involved in the process. In this study, artificial neural networks (ANN) have been used to find the above mapping for an 8/6 SRM drive system. In this paper, the torque production mechanism in SRM is first reviewed to get a closer look at current to torque transformation. Linear estimation techniques are then used for predicting the electromagnetic torque and their shortcomings due to the inherent nonlinearity of the SRM drive are shown. In the next step neural networks have been used to implement nonlinear dynamic predictors. Regressor structure, network topology and learning algorithms have been tuned to introduce a novel torque sensing scheme which can process the phase currents and predict the electromagnetic torque. Certain special considerations for various operations have been taken into account