Neural Network-based Direct Speed Control Of Pmsm Drive Systems

Modern permanent magnet synchronous motor servo systems require highdynamics on speed control. However, the bandwidth of speed loop using conventionalcascade linear controller is limited due to the mutual influence between speed loop andcurrent loop, which results in poor dynamic performance in speed and position control.Model predictive direct speed control (MP-DSC) breaks the limitation of traditionalcascade control structure and improves the dynamics of the system response. However,MP-DSC uses the linear mathematical model of permanent magnet synchronous motorto predict the future states of system, which results in the low robustness of system. Onthe basis of MP-DSC, a neural network based direct speed control (NN-DSC) methodsis proposed to improve the accuracy and robustness of the control system. The mainresearch contents of this paper are showed as follow.First, the control structure, the control process and relative algorithms of MP-DSCis studied in-depth, and the weakness, such as its high and changeable switchingfrequency, its poor robustness and poor dynamic in speed control when disturbed byload torque, are analyzed. The state limitation measure is adopted to reduce theswitching frequency and a load torque estimator is designed to improve the poordynamic of speed control when disturbed by load torque. And the correctness ofMP-DSC control algorithm and the effectiveness of improvement measures are verifiedby simulation and experiment.Second, on the basis of MP-DSC, a new kind of control strategy NN-DSC isproposed, using online learning BP neural network to approach the local dynamic modelof the motor and predict the future states of system, which replace MP-DSC’smathematical model prediction method and improve the prediction accuracy androbustness. In this paper, the structure of the neural network parameters and parameteradjustment algorithm is designed, and the online sliding window learning method isapplied in BP neural network training, and then the conjugate gradient method isadopted to improve the convergence rate and the training time of neural networks withinone single learning cycle is limited to improve the dynamic performance of the system.The advancement of NN-DSC algorithms is verified by simulation and experiment.Finally, a system experiment platform using AD5435real-time emulator as thecontrol core and PS21867integrated power module as the driven core is designed andbuilt. And MP-DSC and NN-DSC was tested on the experiment platform and the resultsare analyzed contrastive. Experimental results show that, MP-DSC has a good dynamicperformance and the modified MP-DSC improves the response speed of the disturbance load torque significantly. NN-DSC not only can achieves good control accuracy anddynamic response, but also has excellent adaptability to parameter perturbation andexternal disturbance. The control accuracy and robustness has greatly improvedcompared to the MP-DSC.