Online assessment of degradation status for metro wheel with robust unsupervised tensor domain adaptation

Abstract

With on-board collected vibration signals, machine learning-assisted metro wheel’s degradation status assessment has received sustained attention in most recent years. Despite promising advances, the acquired vibration signals, heavily interfered by irregular noise due to various factors like road condition, vehicle load and uneven tread, etc., will block the deployment of current assessment methods in open environment. Moreover, existing methods are merely capable of realizing status assessment using offline data within a turning repair cycle, but fail to achieve online assessment that is crucial to actual maintenance. This paper incorporates the concept of transfer learning into metro wheel health management. We identify two major challenges in evaluating wheel degradation status: (1) imprecise representation of degradation characteristics with real-world signals, and (2) lack of common degradation trajectory on the same metro line. To address these challenges, this paper proposes an online assessment approach of wheel degradation status based on unsupervised deep transfer learning, including a robust unsupervised tensor domain adaptation network (RUTDAN) for cross-wheel degradation feature extraction and an online early warning mechanism based on the common health indicator of different wheels. Extensive experiments are conducted with the real-world monitoring data collected from Metro Line 6 of Beijing Subway. The degradation status of target wheel is evaluated with online sequential data, while the results precisely match the actual maintenance records in terms of wheel diameter value.