Reinforcement Learning-Based Online Estimation of the Inductances of the Permanent Magnet Synchronous Machines

This paper proposes a novel method to estimate the d- and q-axis inductances of the permanent magnet synchronous machine based on reinforcement learning algorithms. The main challenges in online inductance estimation are the non-linearity of the inverter and the complex dependence of machine inductances on the operating d- and q-axis currents. A feedforward neural network (FNN) was developed and trained to compensate for the non-linearity of the inverter in the measured d- and q-axis voltages. The Twin-Delayed Deep Deterministic Policy Gradient (TD3) reinforcement learning-based estimator was developed and trained to identify the machine inductances in real-time using the compensated d- and q-axis voltages, the measured d- and q-axis currents, and the rotor speed, allowing high flexibility to adapt to variations in the system during operation by retraining with updated data. The proposed technique was experimentally validated and demonstrated high estimation accuracy under critical operating conditions, including full load variation, field weakening, and speed changes. Simulation and experimental results also showed that the proposed approach outperforms the conventional Recursive Least Squares (RLS) and Adaline Neural Network (ANN) methods, achieving over 10% improvement in estimation accuracy and enhanced robustness against the inverter’s non-linearity.