ML for Loop Gain Identification of DC/DC Converters

Abstract

Control loop identification is necessary for evaluating the stability of switched power supplies and is therefore an important step during design and verification. Analytical models of power supplies often yield inaccurate predictions of the loop gain; therefore, power engineers traditionally must conduct slow, invasive loop gain measurements on physical hardware. This thesis presents an alternate approach to loop gain identification in which a machine learning model infers the frequency-domain loop response from the quick and convenient time-domain measurement of a load step transient. We show that we can train a neural network to accurately infer the loop gain of a current-mode buck converter over a generalized set of configurations and illustrate the disruptive potential of such a model with example applications such as live Bode plot monitoring and automatic loop compensation.