Online training of parallel neural network estimators for control of induction motors

This paper presents an adaptive parallel control architecture, using an artificial neural network (ANN) which is trained while the controller is operating online. The proposed control structure incorporates five-multilayer feedforward ANNs, which are online trained using the Marquardt-Levenberg least-squares learning algorithm. The five networks are used exclusively for system estimation. The estimation mechanism uses online training to learn the unknown model dynamics and estimate the rotor fluxes of an inverter-fed induction motor. Subsequently, the estimated stator currents are fed into an adaptive controller to track the desired stator current trajectories. The adaptive controller is constructed as a feedback signal (a nonlinear control law), depending on estimated stator currents supplied by the neural estimators and the desired reference trajectories to be tracked by the output. The control of the direct and quadrature components of the stator current successfully tracked a wide variety of reference trajectories after relatively short online training periods.