Abstract
Accurate knowledge of temperatures in power semiconductor modules is crucial for proper thermal management. Precise prediction allows operating systems at physical limits while avoiding reliability issues. We propose a linear time-invariant state-space thermal model based on a mesh-based compartment representation. Identification is performed using an Expectation-Maximization (EM) algorithm from incomplete temperature data. By sharing parameters among compartments and constraining the process covariance matrix, we relax the need for measuring all active elements. We also suggest an algorithmic speed-up using steady-state covariances. Performance is demonstrated on simulated data.
What Problem Does This Paper Address?
In power electronics, heat is a primary failure mode. Operating a power module at its physical limit requires precise, real-time temperature knowledge. However, measuring temperature at every point inside a module is practically impossible due to protective packaging and cost constraints.
Engineers typically face a trade-off between:
- Detailed Numerical Models (FEM/FDM): Highly accurate but computationally intensive, making them unsuitable for real-time prediction.
- Lumped Parameter Thermal Networks (LPTNs): Fast "gray-box" models that often oversimplify physical layouts and require sensors on every active component.
This research bridges the gap by creating a model that respects physical geometry (a mesh) but remains efficient enough for identification and prediction using limited sensors.
Methods
Mesh-based Compartment Model
The module is represented as a Linear Time Invariant (LTI) State Space system. The physical structure is divided into a Cartesian grid of "compartments." Unlike standard high-resolution FEM, these compartments are relatively coarse, keeping the state space manageable (hundreds to thousands of states).
Identification via Expectation-Maximization (EM)
The EM algorithm serves as the core identification engine. It iteratively estimates hidden states (temperatures of unmeasured compartments) in the E-step and updates model parameters (thermal conductivities and power loss coefficients) in the M-step.
To improve convergence and prevent overfitting, the authors implement Parameter Sharing. Compartments with similar physical properties (e.g., those within the same copper layer) share the same thermal parameters. This physical intuition acts as a powerful regularizer.
Computational Speed-up
Standard EM identification utilizes the Rauch-Tung-Striebel Smoother (RTSS), which becomes memory-intensive as the state space grows. To address this, the authors introduce a version using steady-state covariances, solving Riccati and Lyapunov equations to drastically reduce computational overhead.
Datasets and Experimental Setup
The approach was tested using synthetic data generated from a model of the Semikron SK20 DGDL 065 ET power module.
- Model Structure: 817 compartments across layers including IGBTs, diodes, copper, ceramic, and heatsink.
- Data: 18,000 time steps.
- Sparsity: Only 40 out of 117 top-layer compartments were "measured" to evaluate the algorithm's ability to predict unobserved temperatures.
Key Results and Insights
- High Accuracy: Long-term predictions showed an error of less than 4% (maximum error of 0.3°C), even with significant temperature gradients.
- Parameter Convergence: The "strongly shared" parameter model (using 5 parameters) converged more reliably to ground-truth values than more complex models.
- Efficiency: The steady-state covariance optimization made it feasible to process models with hundreds of compartments that would otherwise be computationally prohibitive.
- White-Box Transition: By grounding the model in physical layout, the system can accurately predict temperatures in areas with zero sensor coverage.
Limitations and Future Work
Currently, the results are based on simulated data. While the simulation accurately reflects the geometry of real-world modules, real-world non-linearities and boundary conditions may introduce new challenges. Future work will focus on validation using real measured data and optimizing the discretization level to balance accuracy and computational cost.
Summary updated