Power Device Degradation Estimation by Machine Learning of Gate Waveforms

Abstract

The emitter resistance (R <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">E</sub> ), the junction temperature (T <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">J</sub> ), the collector current (I <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">C</sub> ), and the threshold voltage (V <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">TH</sub> ) of power devices are key parameters that determine the reliability of power devices. Adding dedicated sensors to measure the key parameters, however, will increase the cost of the power converters. To solve the problem, power device degradation estimation methods by the machine learning of gate waveforms are proposed. Two methods are shown in this paper. First, in order to detect the bond wire lift-off of power devices, the estimation of the number of the connected bond wires using the linear regression of two feature points extracted from the gate waveforms of a SiC MOSFET is shown using SPICE simulations. Then, in order to detect the power device degradation, the estimation of R E, T <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">J</sub> , I <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">C</sub> , and V <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">TH</sub> using the convolutional neural network (CNN) with the gate waveforms of an IGBT for input is shown using both simulations and measurements.