Motor Thermal Modeling with Deep Learning
Recurrent (LSTM, GRU) and feed-forward virtual sensors for online motor temperature estimation on the public Paderborn PMSM dataset. Methodology generalizes my industrial work on rare-earth-free EESM motors to an open benchmark.
Overview
A clean, reproducible Python implementation of three deep-learning model families for predicting hot-spot temperatures inside a running motor from its electrical and thermal sensor stream:
- GRU — single-layer, last-output regression head
- LSTM — single-layer, last-output regression head
- MLP — feed-forward baseline with flattened windowed input
All three are trained and evaluated with leave-one-profile-out cross-validation on the public Paderborn dataset, using train-only target standardization and early stopping on validation MAE.
Headline result
Cross-validation on the pm (permanent-magnet rotor temperature) target — 10 profiles × 5 random seeds = 50 runs per model.
| Model | RMSE [°C] | MAE [°C] | MaxAE [°C] | Latency [ms/batch] |
|---|---|---|---|---|
| MLP (baseline) | 10.14 ± 4.82 | 8.48 ± 4.47 | 24.98 ± 9.69 | ~0.1 |
| LSTM | 10.61 ± 5.55 | 9.02 ± 5.45 | 27.18 ± 10.10 | ~30.7 |
| GRU | 11.45 ± 6.13 | 9.97 ± 6.05 | 26.68 ± 10.28 | ~24.9 |
Related publications
- F. Tatari, M. M. Aligoudarzi. Deep Learning-Based Rotor Temperature Estimation for Rare-Earth-Free Motors. NDIA GVSETS 2026.
- F. Tatari et al. Data-driven Thermal Modeling for Electrically Excited Synchronous Motors — A Supervised Machine Learning Approach. IEEE ITEC 2024.