A New Adaptive Flux-Oriented Control Framework for Induction Motors with Online Neural Network Training

Abstract

Unlike conventional field oriented control methods, this paper presents a mathematically novel control strategy for induction motor drives, formulated using a two-loop nonlinear dynamic inversion (NDI) framework inspired by aeronautical control architectures. Sensorless operation is realized with a conventional rotor flux observer, while several additional enhancements are introduced to raise overall performance. In particular, a real time radial basis function (RBF) neural network is systematically embedded in a model reference adaptive system (MRAS), replacing the traditional PI adaptation loop with an online training mechanism that improves speed estimation accuracy under parameter variations and load disturbances. The single layer RBF network is trained by gradient descent and incorporated into the nonlinear observer without compromising closed loop stability. The complete controller was implemented on a 1.1 kW, 1430 rpm induction motor using a dSPACE DS1104 real time platform. Experimental results show clear superiority over classical FOC as well as DTSFC and DTRFC schemes, achieving the lowest measured flux ripple (0.002 Wb), minimal torque ripple (0.043 N·m), and the fastest torque response time (0.65 ms). The steady state speed error was reduced by 91 % (from 0.65 to 0.08 rad/s), settling times remained below 60 ms, and both RMSE and ISE metrics decreased appreciably across all tested conditions. Although the proposed design incurs moderate computational overhead, it is fully compatible with real time execution. Future work will examine scalability to high power drives, improved resilience to temperature induced parameter drift, and adaptation of the NDI based framework to permanent magnet machines.