- Conference Article
1
- 10.1109/edpe66853.2025.11224251
Embedded Benchmarking of Neural Network vs. Lookup Table-Based Current Reference Generation for PMSM Drives
- Sep 24, 2025
- Ebru Avci + 2 more +2
This paper presents an embedded benchmarking comparison between Neural Network (NN)- and Lookup Table (LUT)-based current reference generation for Permanent Magnet Synchronous Machine (PMSM) drives using Field-Oriented Control (FOC). Conventional control strategies such as Maximum Torque Per Ampere (MTPA) and Maximum Torque Per Voltage (MTPV) are often implemented using high-dimensional LUTs, which become inefficient as input dimensionality increases. As a scalable alternative, this work investigates an NN to replace these traditional methods. The NN was trained on loss-optimal reference data generated from constrained minimization. Both the NN and a corresponding 3D-LUT were implemented in C and deployed on an STM32H7B3I-DK microcontroller. Benchmarking shows that the NN offers a $97.2 \%$ reduction in memory usage while providing comparable accuracy to the LUT. The NN inference time of 0.5 ms comfortably meets the 2 ms realtime constraint of our application and falls well within the typical timing budget of a higher-level control loop. These results highlight the potential of NN-based control as an efficient and lightweight alternative to traditional methods for electric traction drives.
Read more