Multi-Objective Optimized Computational Neural Network for Performance Enhancement in Non-Sinusoidal PMSM Drives

This work proposes a control scheme for tuning a neural network (NN) with a multi-objective optimizer in order to reduce the torque ripple and losses of a non-sinusoidal PMSM driver over a wide speed operation range. To achieve these goals, the neural network structure and the cost functions are defined, then a virtual machine running in cloud service massively executed the optimization algorithm to obtain several sets of dominant gains and the most appropriate one has been picked. Using these gains and feeding the inputs, the NN is able to add disturbances in both currents and voltages in dq current control loops, allowing torque ripple mitigation. Simulation results are presented to demonstrate the reduction in torque ripple and the losses when compared with approaches using only SPWM or DPWM.