Paper, 2026.03
Digital Twin-Based Motor Speed Ripple Control System with AI-Driven Virtual Sensing and Reinforcement Learning
Running an appliance compressor motor at higher frequencies improves energy efficiency, but load torque fluctuation then amplifies speed ripple and causes noise and vibration. The motor is sensorless, so torque cannot be measured directly, and training reinforcement learning on real hardware is risky and costly. This paper builds a digital twin from an Ansys TwinBuilder PMSM model, predicts load torque with a stacked Bi-LSTM virtual sensor, and uses an LSTM-TD3 reinforcement learning controller to choose the compensation current that reduces ripple. Joonho Seo is the second author and presented the paper at the conference. The virtual sensor reached an MSE of 0.2537 (a 69.46% improvement over prior work), and LSTM-TD3 reduced speed ripple by 27.5%. The paper was presented in the CPS session of ACM SAC 2026 in Thessaloniki, Greece, on March 27, 2026.
- Where
- ACM SAC 2026
- My role
- Second author and presenter
- Status
- published
Virtual sensor MSE 0.2537 (69.46% improvement); LSTM-TD3 reduced speed ripple by 27.5%