DTC drive with ANN-based stator flux estimator

In this paper, we investigate the possibility to implement a novel ANN (artificial neural network)-based stator flux estimator in the DTC-SVM (direct torque controlled-space vector modulated) drive. Proposed estimation scheme does not exploit pure integration, therefore there is no problem with drift and initial conditions. Moreover, it does not require any stator resistance identification algorithm. A multilayer perceptron (MLP) is trained off-line to approximate stator flux space vector. The approximation space is spanned by stator voltages and currents preprocessed with different low-pass filters. The value of stator resistance does not occur openly in elaborated estimator, thus some level of robustness to its variations can be observed. The effectiveness of the novel approximation scheme has been confirmed in simulation as well as by experimental results. The results indicate its superiority to selected solutions based on back EMF integration (under assumption of no additional stator resistance identification algorithm). The performance of the DTC drive with ANN-based stator flux estimator has been compared to the analogous drives with PLPF (programmable low-pass filter)-based estimator and improved integration scheme with an amplitude limiter in polar coordinates. If the windings' temperature rises, there is observable speed dynamics deterioration during reversals of the latter ones, whereas there is no significant change in performance of the former one