Super-Twisting Sliding Mode Observer With Physics Informed Neural Network Compensation for Ultra-Low-Speed Sensor Less PMSM Drives for Electric Vehicle Applications

Abstract

Ultra-low-speed sensor less control of Permanent Magnet Synchronous Motor (PMSM) drives remains a challenging problem because of weak back-EMF signals, parameter uncertainties, inverter nonlinearities, and external disturbances. Conventional Sliding Mode Observers (SMOs) provide strong robustness; however, they suffer from severe chattering and degraded estimation performance under low-speed operating conditions. To overcome these limitations, this paper proposes a Super-Twisting Sliding Mode Observer (ST-SMO) integrated with a Physics-Informed Neural Network (PINN) compensator for ultra-low-speed sensor less PMSM drives. The proposed ST-SMO employs a second-order sliding mode algorithm to reduce chattering while maintaining fast convergence and robustness against disturbances. Furthermore, a PINN-based compensator is incorporated into the observer framework to compensate for nonlinear estimation errors and parameter variations. Unlike conventional neural networks that rely solely on data-driven learning, the PINN incorporates the fundamental PMSM governing equations directly into the training process. By embedding physical knowledge within the network, the model produces estimations that remain consistent with the motor dynamics while exhibiting better generalization across varying operating conditions. The proposed hybrid observer combines the robustness of the Super-Twisting Sliding Mode Observer with the learning capability of the PINN to enhance rotor position and speed estimation, particularly in ultra-low-speed regions where conventional methods often face performance limitations. Simulation results indicate that the proposed approach achieves lower steady-state estimation errors, faster dynamic response, stronger disturbance rejection, and smoother estimation characteristics when compared with conventional SMO, ST-SMO and Extended Kalman Filter(EKF). Furthermore, the proposed ST-SMO–PINN framework is computationally feasible for real-time EV applications, as the PINN is trained offline and only lightweight online inference is required, resulting feasible overall observer execution time, making it suitable for embedded motor drive controllers.