EURO 2024 Copenhagen
Abstract Submission

EURO-Online login

3758. State-aware online self-supervised domain adaptation for bearing remaining useful life prediction

Invited abstract in session WD-40: Experimental economics and game theory 3, stream Experimental economics and game theory.

Wednesday, 14:30-16:00
Room: 96 (building: 306)

Authors (first author is the speaker)

1. Kyeonggeun Park
Department of Battery-Smart Factory, KOREA UNIVERSITY
2. Jun-Geol Baek
School of Industrial Management Engineering, Korea University

Abstract

Transfer learning methods for remaining useful life (RUL) prediction of rolling bearings have rarely focused on online feature drift scenario. Bearing degradation trends are categorized into several states, as normal, slight degradation and severe degradation states, each with distinct feature distributions. During the degradation process, changes in state lead to online feature drift, which reduces the accuracy of RUL prediction. To overcome this challenge, this study suggests aligning feature distributions for each state using state-awareness to adapt to online domain shifts. State awareness is achieved by deriving the Health Index (HI) based on the Wasserstein distance, applying smoothing techniques, and statistically detecting the First Prediction Time (FPT) and acceleration point. This provides state label during online degradation process. The online self-supervised domain adaptation model leverages the current degradation data as target, with the source domain comprising previous run-to-fail data. Aligning feature distributions between these domains and states enhances RUL prediction accuracy by extracting domain-invariant and domain-specific features, as well as state-invariant and state-specific features. Experimental validation using IEEE PHM Challenge 2012 bearing data demonstrates the effectiveness of the proposed methodology in achieving state awareness and accurate RUL prediction through online self-supervised domain adaptation in dynamic operating environments.

Keywords

Status: accepted


Back to the list of papers