Machine Learning Pipeline for Power Electronics State of Health Assessment and Remaining Useful Life Prediction

Abstract

Power electronic devices are integral to numerous applications in modern technology. The non-linear degradation and subsequent abrupt failure of these devices demand regular State of Health (SoH) checkups as well as accurate prediction of their Remaining Useful Life (RUL). SoH and RUL enable efficient maintenance planning, cost-effective resource allocation, and enhanced safety measures. Consequently, this paper proposes a computationally efficient machine-learning pipeline capable of accurately predicting the SoH and RUL of power electronic devices. To evaluate the effectiveness of the proposed approach, we conduct an exploratory analysis of MOSFET (Metal Oxide Field Effect Transistor) device degradation data. A total of 20 power MOSFETs are subjected to stress tests, mimicking real-world operating conditions. We then carry out pre-processing steps on collected data and train a two-stage machine learning pipeline. In the first stage, a non-parametric model classifies the state of the device as healthy or pre-failure. To save computational time, the second stage consists of a regression model that only triggers, when the device is classified as pre-failure, forecasting its resistance degradation trajectory. The combination of Random Forest classifier and Bayesian Ridge regressor effectively captures the non-linear degradation behaviour of the MOSFETs and accurately estimates SoH with an average accuracy of 80% for the classification step and an average RMSPE of 1.25% for the RUL regression step, respectively. Beyond power electronic devices the proposed pipeline can be adapted to determine the RUL of other electronic devices with non-linear degradation behaviour, thus becoming a methodology for future studies in this domain.

Keywords

• Accelerated life cycle test
• data-driven prognostics
• machine learning
• MOSFET
• MOSFET RUL
• predictive maintenance