Comparative performance evaluation of neural network learning algorithms for magnetic characterisation of switched reluctance drives

This paper compares the performance of four different neural network learning algorithms for the purpose of characterising the magnetic behaviour of Switched Reluctance Motor (SRM) drives. The neural nets are trained off-line to predict the static dependence of flux linkage on both phase current and rotor position. This functional representation can be manipulated to produce a stand-alone analytical function representing the non-linear per-phase torque behaviour. The performance of the learning algorithms during training and the performance of the trained neural nets, illustrates advantages of each training algorithm. Performance is measured on rate of convergence and minimum RMS error.