Estimation of remaining useful lifetime of power electronic components with machine learning based on mission profile data

Abstract

Reliability of power electronic components is essential to functionality and safety. In this paper, a data-driven method is presented to estimate the remaining useful lifetime of solder joints used in power modules of electric bikes. Temperature mission profile data is acquired from the electric bikes under different loading conditions and key temperature features are generated. Accumulated creep strains in solder joint of a chip resistor are evaluated by finite element analysis. A machine learning model, namely multilayer perceptron is first trained with the synthetically generated data from finite element analysis. The model is further introduced to creep strains generated under mission profile data by transfer learning methods. Results show that machine learning model trained with combination of mission profile and synthetic data has high accuracy with just 6.7% average error against unseen field data. Remaining useful lifetime is then evaluated based on predicted accumulated creep strains. This methodology provides a viable solution for real-time remaining useful lifetime estimation based on combination of synthetic and real-world data.