An AI Architecture for Power MOS Rds-on-Driven Lifetime Estimation

The components lifetime estimation is one of the main challenge in all engineering process specially in power electronics. The problem involves several domains of interest and is particularly felt, because, the estimation of the residual life with an acceptable degree of confidence, allows to carry out preventive measures in order to avoid the consequences of the malfunction and failure of the device at the end of its life. Considering a Power MOS, the distinctive element for understanding the end of life of this component can be obtained by observing the
$Rds_{on}$
curve during its life. The
$Rds_{on}$
curve works in an operating range during the correct life cycle, this curve exceeds a certain threshold when the device begins to have the first signs of degradation. The proposed method allows to evaluate the variation of the
$Rds_{on}$
by making predictions for understand when the device may stop working well. The method is based on a deep learning model trained with a dataset containing the cycle of life of several Power MOS. The pipeline is based on Artificial Intelligence techniques with regularization of the Jacobian of the gradient, designed to compensate for the inevitable measurement and conversion noises present in the system based on ST MCU. The experiments concern predictions of the
$Rds_{on}$
at 10, 15 and 20 steps and allow to obtain a good degree of accuracy for predicting the
$\mathbf{RDS}_{-}\mathbf{on}$
at least 20 samples before. This allows to guarantee a preventive intervention when the prediction exceed the threshold considered for the correct operating range. The proposed pipeline allows to robustly characterize the threshold voltage of the device by means of a simple acquisition of voltages and currents at power-on, being able to monitor the evolution of this voltage over time and each time the device is used. The obtained results and the cross-validation performances are very promising and allow us to state that the proposed method is really robust.