Control × AI
AI-Based Sensorless Rotor Position Estimation for PMSM
A CNN rotor-angle estimator benchmarked against an Extended Kalman Filter for inverter-fed PMSM operation, with emphasis on the low-speed observability problem.
7.12M
Simulation samples
89K
Windows
0.2316 rad
CNN low-speed RMSE
2.5316 rad
EKF low-speed RMSE
Problem
Conventional model-based observers become difficult to use at very low speed because back-EMF is weak and electrical observability degrades. The project asks whether a learned estimator can complement a classical EKF in that operating region.
System
The simulated drive combines an inverter-fed PMSM, field-oriented control, α–β current and voltage measurements, a CNN estimator and an EKF benchmark. Both estimators output electrical rotor angle.
CNN estimator
A 1D CNN uses windowed iα, iβ, vα and vβ signals. Angle is represented using sine/cosine targets, with offline training and online Simulink deployment. The training set was generated from time-domain simulation.
Result
The CNN was especially strong at low speed, where online RMSE was 0.2316 rad versus 2.5316 rad for the EKF. The EKF remained stronger at high speed, showing complementary strengths rather than a single universal winner.
Evidence