Permanent Magnet Synchronous Machines (PMSMs) are critical components in modern electric vehicles (EVs). However, they face significant thermal challenges; excessive heat in windings or rotors can lead to permanent magnet demagnetization and insulation failure. To mitigate this, engineers employ "thermal torque derating," which throttles motor power based on temperature. Traditional methods rely on physical sensors—which are costly and prone to failure—or complex physics-based Lumped Parameter Thermal Networks (LPTNs) that are difficult to calibrate and computationally intensive for automotive Electronic Control Units (ECUs).
This research, published in Energies (2025), proposes a hybrid solution: embedding a Deep Neural Network (DNN) within a Moving Horizon Estimation (MHE) framework. This approach combines the flexibility of data-driven modeling with the robustness of classical state estimation.
Methodology and System Architecture
The core of the proposed system is the integration of a Long Short-Term Memory (LSTM) network directly into the MHE optimization loop. Unlike standard models that predict absolute temperature, this LSTM is trained to predict the temperature gradient (rate of change), which is then integrated to find the state.
The architecture utilizes the acados framework to solve the non-linear optimal control problem required for MHE. By augmenting the MHE state vector with the LSTM's internal hidden and cell states, the estimator maintains a "memory" of past thermal behavior, effectively capturing the motor's thermal inertia.
Moving Horizon Estimation treats state estimation as an optimization problem over a sliding window of time, as illustrated below:
To ensure the model handles safety-critical edge cases, the researchers used high-fidelity synthetic data from a 70-node thermal model validated against a real-world DENSO PMSM. Training and validation datasets included customized WLTP Class 3 cycles and a simulated high-performance lap of the Nürburgring Nordschleife.
Key Results and Performance
The hybrid MHE-DNN system demonstrated exceptional accuracy and efficiency:
- High Accuracy: The system achieved a Root Mean Square Error (RMSE) of just 0.0373 °C/s for winding temperature gradients.
- Embedded Performance: On a 3.8 GHz dSPACE SCALEXIO system, the average computation time was 5.7 ms per 100 ms timestep. Even the peak computation time (28 ms) was well within the 100 ms limit.
- Real-Time Margin: The system demonstrated threefold real-time capability, meaning it uses only about 30% of the available processing window, leaving plenty of room for other vehicle control tasks.
- Robustness to Failure: A key finding was the system's resilience to sensor failure. When a massive negative offset was injected into the physical sensor data, the MHE/DNN hybrid successfully identified the anomaly and maintained accurate temperature estimates.
Discussion and Future Outlook
By placing a DNN inside an MHE framework, the authors bridge the gap between "black-box" AI and physics-based control. This hybrid intelligence allows for noise filtering and constraint handling while benefiting from the predictive power of neural networks. The use of temperature gradients rather than absolute values makes the model more stable and physically intuitive.
However, challenges remain. The "black-box" nature of DNNs complicates certification under automotive safety standards like ISO 26262. Furthermore, while the dSPACE system is an embedded platform, it is more powerful than standard production ECUs, necessitating further optimization for lower-cost hardware. Future work will explore transfer learning to adapt models to different motor configurations and the use of residuals for advanced sensor diagnostics.
Summary updated