Adaptive neural network based controller for direct torque control of PMSM with minimum torque ripples and EMI noise reduction

Purpose - The paper aims to provide an adaptive neural network controller for permanent magnet synchronous motor (PMSM) under direct torque control (DTC) algorithm to minimize the torque ripple and EMI noise. Design/methodology/approach - The design methodology is based on vector control used for electrical machines. MATLAB simulations supported with experimental study under C++ are used. Findings - The simulated and experimental results show that considerable torque ripple as well as current ripple and EMI noise reduction can be achieved by utilizing adaptive neural switching algorithm to fire the inverter supplying the PMSM. Research limitations/implications - This research is limited to PMSM, however the research can be extended to include other AC motors as well. In addition, the following points can be studied: the effects of harmonics in control signals on the torque ripple can be analyzed; the actual mathematical relation between the torque and flux ripple can be studied to set the flux and torque bands width in reasonable value; different neural network algorithms can be applied to the system to solve the similar problems. Practical implications - Based on existing DTC control system, it is only required to change the software switching algorithm, to provide smooth torque, given that the switching frequency of the inverter module is more than or equal to 15 MHz and the system is supplied with timers. In addition a relatively higher DC voltage may be required to achieve higher speed compared with the traditional DTC. Originality/value - In this paper, the stator flux position, and errors due to deviations from reference values of the torque and stator flux are used to select two active vectors while at the same time the absolute value of the torque error and the stator flux position are used neural network structure to adapt the switching of the inverter in order to control the applied average voltage level in such a way as to minimize the torque ripple, so instead of fixed time table structure, a neural network controller is used to calculate the switching time for the selected vectors and no PI controller is used as the case in the traditional space vector modulation. This work is directed to motor drive system designers who seek highly smooth torque performance with EMI noise reduction.