Lightweight LSTM and GRU Design for Data-Driven Rotor Position Error Estimation in IPMSM Drives

High-performance motor drives rely on closed-loop controls that typically obtain rotor position from a physical encoder or a rotor position estimator. However, it is well established that there may be discrepancies between the measured/estimated position and the actual one. This may be due to the loosening of the encoder's mechanical fixing, initialization errors, sensorless estimation errors, etc. The rotor position error, if left uncompensated, may lead to torque fluctuation and reduced system efficiency. Different from the mainstream iterative or model-based methods introduced thus far, this paper focused on a data-driven solution that is based on the use of lightweight Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) neural networks, realized in conjunction with real-time embedded microcontrollers. Benchmarked against the familiar choice of Multilayer Perceptron (MLP), these emerging recurrent neural networks (RNNs), which have received tremendous attention in computer science subjects but much less in power electronic-based electric drives, are designed for estimating position errors with high accuracy. Upon careful consideration of the embedded data in the stationary and rotating reference frames, data down sampling, and real-time computing capability, this paper shows that these emerging RNNs are potentially more robust against measurement noises and harmonics inherently present in drive systems. They are proven to better generalize to non-training operating points or data, constituting an essential feature when dealing with closed-loop control's experimental data. The proposed lightweight LSTM- and GRU-based neural networks are extensively validated using a 2.2 kW interior permanent magnet synchronous motors (IPMSMs) through simulations and experiments for estimating the step- and ramp-type dynamic rotor position errors. The comparative evaluation against the classical iterative rotor position correction method confirms its superiority in terms of estimation speed and accuracy, suggesting a good potential of the data-driven concept in improving electric drives.