Enabling Online GaN Power Device Self-Health Monitoring With Analog SGD Supervised Learning and TJ-Independent Precursor Measurement

In order to mitigate reliability challenges on emerging wide-bandgap (WBG) gallium nitride (GaN) power devices, this article investigates a software–hardware codesign solution utilizing online device condition monitoring and machine-learning technologies, with an ultimate goal of achieving intelligent device self-health learning efficiently and effectively. Specifically, an on-die logarithm-based analog stochastic gradient descent (SGD) supervised learning engine is developed to train a GaN power circuit to establish and update its own device thermal model seamlessly through autonomously self-tested condition precursor
$r_{\text {DS}\_{}\text {ON}}$
. To remove the adverse temperature (
$T_{J}$
) impact on condition monitoring efficacy, a nonintrusive online
$T_{J}$
sensing scheme is introduced by evaluating the ambient temperature (
$T_{A}$
) and the instant power loss (
$P_{\text {LOSS}}$
) to accomplish a truly
$T_{J}$
-independent precursor measurement. To validate this research effort, a GaN-based switching power converter prototype was built. The controller and gate drivers are fabricated using a 180 nm high-voltage (HV) bipolar-CMOS-DMOS (BCD) process, with an active die area of 2.9 mm2. Operating at 3.3 MHz, the converter can support a wide range of
$V_{\text {IN}}$
from 5 to 40 V and deliver a maximum power of 9 W at 5-V
$V_{O}$
. The proposed on-die logarithm-based supervised learning engine and
$T_{J}$
-independent
$r_{\text {DS}\_{}\text {ON}}$
condition monitoring module occupy 0.89 mm2 silicon die area and consume up to 3.5 mW power. The design demonstrates a maximum prediction error of 2.77% over the temperature range from 0 °C to 120 °C.