Machine Learning-Based Methods for Automatic Compensation of Dead-time Distortion: A Comparative Study

Abstract

In many power converters, a compensated voltage reference for dead-time distortion is typically preset during the design stage. However, this reference may deviate from the optimal value when semiconductor devices are replaced or when device characteristics change due to operating conditions. As a result, there is a growing need for active adjustment of the compensated reference. This paper evaluates the effectiveness of machine learning techniques for this task, and compares three approaches: (i) an image classification model, (ii) a waveform feature-based linear regression model, and (iii) a one-dimensional CNN regression model using time-series data. Results show that method (iii) outperforms the others.