Prognostic Analysis of Switching Devices in DC-DC Converters

In this work, a monitoring technique for switching devices used in DC-DC converters is presented. The prognostic approach requires the processing of time domain measurements extracted from specific test points of the converter under test, with the purpose of assessing the drain-to-source resistance variations due to overheating before they lead to a total loss of functionality. The proposed monitoring method can be easily applied on different topologies while maintaining the same theoretical concept. In this study, a Zeta converter is considered to demonstrate the effectiveness of the analysis. The prognostic estimation proposed here is divided into two steps: the first phase involves a classification of the converter state to distinguish malfunctions of the active devices from faults of other passive components (capacitors and inductors); subsequently, a regression of the drain-source resistances is carried out to detect the extent of the problem. The two phases are performed using different Machine Learning methods and their performance is compared in both the classification and regression tasks. The results show a good performance both in classifying the faulty conditions, with an over 90% classification accuracy, and in predicting the drain-to-source resistance value, achieving an absolute RMSE of about 8 mΩ for the best performing algorithms, even in presence of highly variable circuit components, simulated through a Monte Carlo approach.