An Application of Feature Engineering and Machine Learning Algorithms on Condition Monitoring of SiC Converters

This paper proposes an approach for noninvasively fault detection in a wide range of power electronics applications. Simulations were used at this stage for gathering the large amount of data needed, but the conclusions can be applied as-is to experimental data sets. These data were fed to the machine-learning algorithm for training and validation. A buck converter based on silicon-carbide MOSFETs is selected as the case study. The beauty of the proposed approach is that it only uses the output terminal signals for diagnostics. Our approach starts with extracting features from the output voltage signal, then inferring criteria to evaluate and rank these features in terms of relevance (feature engineering phase). The final step in this phase is reducing the space dimension by selecting sets of meaningful features. Finally, several artificial neural network structures demonstrate the effectiveness of this approach.