Enhancing Sensorless Speed Estimation Accuracy Through Global Parameter Identification and Neural Network-Based Residual Compensation

Sensorless speed estimation replaces fragile shaft encoders in cost-sensitive Permanent Magnet Direct Current (PMDC) motor drives, but classical model-based observers degrade under brush friction, commutation ripple, and thermal drift, while purely data-driven estimators sacrifice physical interpretability. This paper presents a Hybrid Physics-Data-Driven Observer (HPDDO) that couples an identified lumped-parameter electrical model with a compact multilayer-perceptron residual compensator, which is executed in real time on a low-cost ESP8266 microcontroller. Global parameters are identified from a short labeled recording, after which the network corrects only the nonlinear residual that the physics model cannot explain. Under a strictly time-series-aware evaluation (chronological 80/20 split), the proposed estimator achieves an average root mean square error (RMSE) of 4.11 RPM across dynamic PWM sweeps, abrupt load transitions, and a long-duration thermal-drift test, outperforming an extended Kalman filter (9.49 RPM), a sliding mode observer (10.89 RPM), and a pure neural-network estimator (7.14 RPM) implemented on the identical dataset. An ablation study shows that accuracy is insensitive to network size, with a 0.9 kB variant matching the deployed model, and a residual-clamping safeguard bounds the estimation error under unseen operating conditions. The framework provides an accurate, interpretable, and computationally lightweight solution for industrial PMDC drives without dedicated speed sensors.