PMSM Torque Estimation Based on Machine Learning Techniques

Permanent Magnet Synchronous Motors (PMSM) are commonly used in robotic arms for light load applications such as industrial medical, and home service. These robotic applications are required an ability to aware their operating conditions due to safety reasons as they are likely to be operated close to humans. One of the major parameters to examine the situations is rotor torque which conventionally acquired by a torque transducer. However, using such device requires extra cost, bulky mechanical installation, and data acquisition electronics. Due to this trade-off, we propose a machine learning based method for rotor torque estimation. The information used in the estimation process are solely electrical signals generated by the motor itself without any actual sensor required. In this paper we present an investigation and feasibility study of the proposed sensorless torque estimation. A motor test bench has been developed for observing motor characteristics also for collecting information to create torque prediction models. Numerous statistical based machine learning methods have been applied in this work including Neural Networks regression, Linear regression, and Stepwise regression. The proposed system has been used to created prediction models according to the occupied regression methods. The estimation performance has been considered by comparing the estimated results with ground truths from an actual torque sensor. The estimation model based on Neural network regression has achieved highest accuracy at 0.11 of RMSE and 0.996 of R value. The results shown the potential of applying the proposed senseless torque estimation for robot application with acceptable performance.