Adaptive Sensorless Vector Control of Permanent Magnet Synchronous Motors Using Deep Reinforcement Learning

Abstract

ABSTRACT Permanent Magnet Synchronous Motors (PMSMs) are widely integrated into industrial drives, electric vehicles, and precision motion systems due to their high power density and efficiency. High-performance field-oriented control (FOC) traditionally relies on mechanical speed and position sensors, which increase hardware costs, physical footprint, and susceptibility to environmental disturbances. Conventional sensorless techniques, such as Sliding Mode Observers (SMO) and Extended Kalman Filters (EKF), suffer from high-frequency chattering, estimation lag, and performance degradation under parameter variations and low-speed operation. To address these challenges, this paper proposes an adaptive sensorless vector control framework driven by Deep Reinforcement Learning (DRL). Utilizing a Twin Delayed Deep Deterministic Policy Gradient (TD3) continuous actor-critic architecture, the proposed DRL agent dynamically tunes the switching gains of an adaptive sliding mode observer while optimizing quadrature reference currents in real time. The proposed control structure eliminates chattering through dynamic boundary layer modulation and actively compensates for thermal stator resistance drift. High-fidelity Matlab/Simulink co-simulations confirm that the DRL-based adaptive scheme achieves zero-speed position estimation error within ±0.4 electrical degrees, reduces speed tracking response time by 42% relative to standard PI-SMO drives, and suppresses torque ripple to under 2.3%.

KEYWORDS: Permanent Magnet Synchronous Motor (PMSM), Sensorless Vector Control, Deep Reinforcement Learning (DRL), Twin Delayed Deep Deterministic Policy Gradient (TD3), Sliding Mode Observer, Field-Oriented Control (FOC), Thermal Drift Adaptation.