Adapting CMAC neural networks with constrained LMS algorithm for efficient torque ripple reduction in switched reluctance motors

This paper presents a novel approach to learning control in switched reluctance motors (SRM's) for torque ripple reduction using a cerebellar model articulation controller (CMAC) neural network. The approach modifies the conventional LMS adaptive algorithm using a variable learning rate function over the rotor angle of the motor under control. The criteria and method for the development of current profiles suitable for use over a wide range of motor speeds are described. In particular, current profiles can be designed to possess desirable characteristics by selection of learning rate function with appropriate switching angles during the training of the network. The approach allows the generation of optimal current profiles in terms of minimizing torque ripple and copper loss as the motor operates at low speeds, and of minimizing torque ripple, copper loss and rate of change of current as the motor runs at high speeds. Experimental measurement of the torque production characteristics of a 4 kW, four-phase switched reluctance motor forms the basis of simulation studies of this approach. Substantial simulation results are reported and the performance of learned current profiles analyzed. These demonstrate that developing CMAC-based adaptive controllers following this approach affords lower torque ripple with high power efficiency, while offering rapid learning convergence in system adaptation.