Reliability Monitoring and Predictive Maintenance of Power Electronics with Physics and Data Driven Approach Based on Machine Learning

Abstract

This paper proposes a new prognostics analysis approach for power electronics by combining physics-based and data-driven techniques. Starting with Weibull degradation model, machine learning (ML) techniques are applied to degradation data progressively for continuous reliability monitoring and predictive maintenance decision-making. No prior knowledge of components or mission profiles is required for model training and prediction. Extracted features from analysis can be used to cluster the iteration-based predictions effectively. Another advantage is abrupt change of operation condition can be captured through machine learning for potential lifetime improvement through predictive maintenance. Proposed method can be generalized to other hardware components beyond power electronics.