Expectation-Maximization Algorithm for Identification of Mesh-based Compartment Thermal Model of Power Modules

TL;DR

This paper introduces a structured mesh-based compartment thermal model for power modules, utilizing an Expectation-Maximization algorithm to identify shared parameters from limited temperature data, reducing measurement requirements.

Contribution

It presents a novel structured compartment model with shared parameters and regularization, enabling temperature prediction with fewer measurements compared to previous methods.

Findings

Model accurately predicts unmeasured temperatures.

Parameter sharing improves identification robustness.

Simulation results demonstrate effectiveness on synthetic data.

Abstract

Accurate knowledge of temperatures in power semiconductor modules is crucial for proper thermal management of such devices. Precise prediction of temperatures allows to operate the system at the physical limit of the device avoiding undesirable over-temperatures and thus improve reliability of the module. Commonly used thermal models can be based on detailed expert knowledge of the device's physical structure or on precise and complete temperature distribution measurements. The latter approach is more often used in the industry. Recently, we have proposed a linear time invariant state-space thermal model based on a compartment representation and its identification procedure that is based on the Expectation-Maximization algorithm from incomplete temperature data. However, the model still requires to measure temperatures of all active elements. In this contribution, we aim to relax the need for all measurements. Therefore, we replace the previous dark gray-box approach with a structured compartment model. The structure of the model is designed by a mesh-based discretization of the physical layout of the module. All compartments are assumed to share parameters that are identified from the data of the measured elements. Temperatures of the unmeasured elements are predicted using the shared parameters. Identification of the parameters is possible only with suitable regularization due to limited amount of the data. In particular, the model tightening is accomplished by sharing parameters among compartments and by constraining the process covariance matrix of the model in this contribution. Applicability of the proposed identification procedure is discussed in terms of growing state-space and therefore speeding up of the identification algorithm is suggested. Performance of the proposed approach is tested and demonstrated on simulated data.