A novel dynamic state observation method based on modified adaptive BP neural network for PMSM drive

The accuracy of angle feedback is crucial for ensuring the reliability and stability of motor control systems. When the angle values directly obtained from the magnetic encoder are used as feedback angle values, jump points typically occur at the boundaries of magnetic poles, and the high-frequency noise is introduced, which can adversely affect the reliability of angle feedback. The angle values predictive algorithm based on low-pass filtering (LPF) for magnetic encoder has been employed to block the occurrence of jump points at zero-crossing. However, the accuracy of angle feedback needs to be enhanced. Therefore, a modified second-order state observer based on modified adaptive backpropagation neural network (BPNN) is proposed, in which the adaptive weight coefficient of observed angle value of the LPF and the kinematic equation is established. Then, the appropriate weight coefficient is searched based on the convergence rate of the observed angle error, thereby decreasing the observation error is reduced. Experimental results show that the noise of the feedback angle values for the magnetic encoder can be effectively suppressed by the proposed algorithm. Meanwhile, tracking performance of the observer is enhanced by employing weight coefficients with superior tracking performance. At last, the accuracy of the observed angle achieves a precision of up to 0.54° at locations where the angle value of the control system changes smoothly.