Measure Theory-based Approach for Remaining Useful Lifetime Prediction in Power Converters

It is well known that the reliability of each component in a power converter affects the reliability of the overall system. Due to the advancements in computing infrastructure and sensor technologies, data-driven approaches for the prediction of the health of power converters in real-time are slowly becoming popular. This paper proposes a new statistical approach using probability density functions (PDFs) and associated concepts in measure theory to predict the probability of system failure using individual components' degradation data. For this purpose, remaining-useful-life (RUL) is estimated for each power component (or sub-system) using qualification data, followed by an evaluation of a cumulative probability of survival for the converter. An artificial neural network (ANN) is then trained to quickly estimate in real-time, the probability of survival of the power converter in the future. While the algorithm involves multiple computation steps, the RUL prediction accuracy using the proposed method will be high due to the data-driven approach. Moreover, the machine learning-based model resulting from this approach to predict the probability of survival is light on memory utilization. It is envisioned that this approach can be used to create digital twins of power converters in practical circuits, optimize performance, and predict RUL. This paper explains the theory followed by an example analysis of an isolated DC-DC converter. An experimental qualification setup for device degradation test and system-level RUL measurement methods are provided.