Control of multi-mass system by on-line trained neural network based on Kalman filter

The purpose of this paper is to obtain on-line trained Artificial Neural Network Controller for PMSM multi-mass high dynamic drive. Structure of the controller with training algorithm and idea of Kalman Filter as observer are shortly described. The Resilient Back Propagation algorithm (RPROP) was chosen for ANN training process. There is assumed rotor position can be sufficient to the possibility of torsional vibration damping. The Neural Network Controller has been proposed instead classical form of control loop with speed sensor. The problem of controller synthesis is discussed and solved. The main advantage of proposed system is Kalman Filter algorithm using to obtain all necessary signals for ANN controller. The measurement of motor position is enough for good control strategy. The proposed control scheme guarantee good properties in scope of mechanical parameter changing. The speed response is parameters nearly independent. The estimation scheme and controller was tested on a single drive setup under its mechanical elements changing. There are presented an original combination of Artificial Intelligence method with classical form of mathematical filters. We proved that newest control structures work best with known behaviour of classical observers theory.