Bayesian Optimized Long-Short Term Memory Recurrent Neural Network for Prognostics of Thermally Aged Power Mosfets

Abstract

Power semiconductor switches such as MOSFET are widely used in electric vehicles, electric trains, and power conversion systems. The prognostics and health monitoring (PHM) and reliability of the MOSFETS are critical in power electronic systems. The ON-state resistance of the MOSFETS is the primary and significant precursor of failure. In this paper, we present a data-driven approach to predict the change in ON-state resistance using Long-Short Term Memory (LSTM) Recurrent Neural Network (RNN). The experimental data are obtained by power cycling the MOSFET in thermal stress. The LSTM network's hyperparameters are optimized using Bayesian Optimization, which improves results and reduces time to develop the model. Our proposed Bayesian Optimized LSTM RNN (BO-LSTM RNN) gives good prediction accuracy using less number of layers. The model's performance is also analyzed using a small percentage of data used to train the LSTM RNN. The proposed model is also tested on different MOSFETS degradation data to prove the universality of the model using the same hyperparameters.