Abstract
An accurate lifetime prediction of power MOSFET devices is vital for critical applications such as hybrid electric vehicles, high-speed trains and aircrafts. These devices are subject to thermal, electrical and mechanical stresses on the field and hence the reliability study of these devices is of utmost concern. The performance of model-based methods depends on strong assumptions of the initial values for the parameters and also on the choice of the degradation model. In this work, we propose to use a data-driven method using the feedforward neural network for prognosis of power MOSFET devices with large noise. The experimental data consists of accelerated aging tests done on these devices, extracted from recently published work. The impact on modifying the complexity of the neural network framework on the prognostic metrics such as relative accuracy and computational time are analyzed and quantified. The results demonstrate that the neural network model yields good prediction results even for a highly noisy dataset and also for degradation trends that are strikingly different from the training dataset trend.
•A Feedforward Neural Network (FFNN) framework is presented for prognosis on Power MOSFET ON-state resistance degradation.•The FFNN learns degradation trend of one device and predicts RUL for remaining devices with similar degradation trends.•The effect of noise and number of hidden neurons on the prediction accuracy is analyzed.•The RUL prediction accuracy increases up to a given number of neurons and subsequently decreases again.•The robustness of the method is analyzed by varying the size of the prediction dataset.