Wheel condition assessment of high-speed trains under various operational conditions using semi-supervised adversarial domain adaptation

Abstract

Train wheels, among other components, are critical for the safety and ride comfort of high-speed rail systems. Various machine learning methods have been used together with onboard monitoring data to assess the wheel health conditions. However, only in some well-controlled experiments or authorized circumstances (source domain) can the well-labelled monitoring data for supervised learning be obtained. Even so, due to the difference in operational conditions, directly applying the model learned from this case to the case of interest (target domain) is not reliable. Facing this challenge, we propose an adversarial domain adaptation (DA) approach to transfer knowledge from a well-controlled monitoring test in one rail section to the rail section of interest. Since in the target domain, the data corresponding to components that are new or after reprofiling can be labelled as “intact”, the DA is modified to be semi-supervised rather than unsupervised. Two-level marginal and conditional DA is conducted in an adversarial manner, which can sufficiently eliminate the distribution discrepancy induced by the operational differences between two rail sections on which the train runs. Onboard monitoring data collected from the Lanxin high-speed rail section before and after wheel reprofiling is used as a case study. Results demonstrate the effectiveness of the approach as well as its superiority over three baseline models, and the underneath mechanisms are visualized. The study is expected to provide new thinking for the condition assessment for other key components when the train runs under various operational conditions.