Self-Calibrating Loss Models for Real-Time Monitoring of Power Modules Based on Artificial Neural Networks

Abstract

This paper proposes a self-calibrating device loss model for real-time monitoring of power modules that applies artificial neural networks (ANNs). Combining loss estimating ANNs with junction temperature data in thermal observers allows continuous learning and self-calibration of the ANN on the basis of the loss estimation error of the observer. With the proposed ANNs, device loss characteristics can be modeled with superior accuracy over state-of-the-art parametric loss models. The learning-based self-calibration process introduces unique features: The device loss model can be applied to any power device technology without offline calibration and prior knowledge. Aging dependent changes of the losses are accurately captured and could even be used for device degradation diagnosis. The real-time generation of loss data allows flexible and effective converter efficiency measurements during operation. Overall, the proposed technology strongly reduces or even omits offline loss calibration of power device prior to operation. Furthermore, the application of the highly accurate device loss models in thermal monitoring systems allows monitoring thermal cycles with high precision. This is crucial for safe converter operation at thermal limits as well as accurate prediction of the remaining useful lifetime of power modules. Within this paper the proposed self-calibrating device loss model is described in detail and its static and dynamic performance is evaluated using measurements of an Infineon Hybridpack 2 IGBT module as an example.