Robust controlled artificial neural network flux estimation for induction motors.

摘要

Presented in this dissertation is a neural network flux estimator employing robust controllers for a field oriented controlled induction motor drive. Field oriented control (FOC), also referred to as vector control, is used to achieve high dynamic performance in inverter-fed induction motor drives, whereby it is necessary to know the instantaneous magnitude and position of the rotor flux. Rotor flux, as well as shaft speed, robust controllers are designed in terms of stability and performance criteria through a well defined, straight forward graphical technique, called loopshaping, taking into account the effects of external disturbances and motor parameter deviations from the nominal model. The simulated FOC drive system generates the open-loop mode training and test data, and becomes a means of performance verification in the closed loop mode, for the proposed neural network flux estimator.; The neural network architecture selected for this application is a three layer feed-forward network employing the backpropagation training algorithm. The proposed neural network flux estimator has 12 neurons in the input layer, 12 neurons in the hidden layer and 3 output neurons. The inputs to the flux estimator are the direct and quadrature stator currents, ids and iqs respectively, and their delayed values, as well as the delayed values of the flux magnitude Psi, and sine and cosine of the field angle &phis;. The outputs are the flux magnitude, sin&phis; and cos&phis;.; The neural network's ability to accurately estimate the flux magnitude and field angle under various load conditions and parameter variations is verified by a parameter sensitivity study and comparisons with alternative control schemes. The advantages presented by the neural network over conventional flux estimating methods include its adaptability (i.e. effective extrapolation), generalization (i.e. ability to estimate flux response lying outside the training data set), and the capability of handling time varying non-linearities. The primary disadvantage is the potentially long training process involving a large degree of trial and error investigation. Overall, the neural network responds reasonably well to both the training data and the test data, suggesting that the proposed robust controlled neural network FOC flux estimator may be practically realizable.