Direct neural network adaptive observer control for PMSM

Rotor position detection is necessary for phase commutation and current control in high performance PMSM. Traditional detection methods are based on resolver, absolute encoder etc. The paper presents a position and velocity sensorless control algorithm based on a direct neural model reference adaptive observer. The proposed observer comprises two neural networks which are trained to learn electrical and mechanical models respectively. By using current prediction error, adaptation is realized by new fast orthogonal triangular decomposition based RLS training method. Various advantages of this estimating scheme over other sensorless control scheme, such as robustness, nonlinear adaptation and learning ability is shown by extensive simulations