Hybrid Learning Approach for Accurate Resolver Position Estimation Based on Error Metrics

Abstract

This study proposes a Reinforcement Learning (RL)-based hybrid machine learning model to enhance the accuracy of resolver position estimation and reduce error tolerance. Traditional methods struggle to maintain high accuracy under signal disturbances such as noise, DC offset, phase shift, amplitude imbalance, and speed variations. To achieve higher accuracy, the proposed hybrid model dynamically weights the Arctangent, Phase-Locked Loop (PLL), Angle Tracking Observer (ATO), and Artificial Neural Network (ANN) methods to ensure the lowest-error position estimation across different fault scenarios. The RL agent analyzes signal characteristics instantly, selecting the most suitable model and continuously updating the weights to minimize errors. MATLAB/Simulink-based simulations demonstrate that the proposed model achieves higher accuracy and stability compared to conventional methods. Additionally, the hybrid model dynamically adapts to various fault scenarios, effectively reducing error tolerance and offering a reliable alternative for real-time applications. This study highlights the benefits of hybrid machine learning approaches in improving accuracy for resolver-based control systems, providing a flexible and precise solution for automotive, robotics, and industrial applications. The hybrid model leverages strengths of conventional methods, dynamically adjusting to varying conditions with optimized performance.