Scalable Reliability Monitoring of GaN Power Converter Through Recurrent Neural Networks

Abstract

Reliability and operation of high-frequency Gallium Nitride (GaN) power converters are yet to be discovered. Coming with the reliability assessment and improving the life extension of power converters, the approach is to monitor semiconductor on-resistor changes as a precursor signature for diagnostic/prognostic. This paper presents a novel approach for hybrid condition-based prognostic and reliability monitoring of GaN devices. The proposed approach offers a multi-physics co-simulations solution for degradation fatigue modeling of the GaN power devices. With the availability of the most granular information deduced from the advanced devices, the paper develops deep learning based algorithms for online reliability in power electronics. The proposed algorithm is based on the prominent version of Recurrent Neural Network (RNN) named Long Short-Term Memory (LSTM). LSTM models are utilized for system training and simulation model calibrations, and eventually predicting the next states within the next time horizon.