Evaluation of Chinese freight train bearing condition based on spatiotemporal feature extraction

Abstract

Bearing failure in freight trains can directly affect operational safety. Therefore, bearing condition evaluation is of practical and critical significance. In this paper, a new method of evaluating the condition of the bearings is proposed based on spatiotemporal feature extraction with bearing temperature data. First, the temperature time series is divided into many sub-sequences with a sliding time window. Second, based on the forms of series anomaly, point anomaly, and pattern anomaly in time series, the spatiotemporal features extraction framework is presented, which combines manual feature extraction with machine learning methods. Specifically, series classification, distribution-based and increment-based outlier detection, model-based time series anomaly detection methods are utilized to extract abnormal features in temporal, spatial, and spatiotemporal dimensions. Third, a penalty vector is constructed by an ergodic accumulation of penalty values for the outliers. And the weights to each element in the penalty vector are allocated using correlation analysis. The penalty vector and the weights are then employed to calculate the bearing health indicator, which can quantify the health condition and determine the severity of the potential fault. Finally, the validity of the proposed evaluation method is verified with on-site historical temperature data of 15,120 bearings on 35 trains, which demonstrates an accuracy of more than 94%. The method can provide early warning for an average of 161 h before hotbox alarms. The results indicate the proposed method can effectively evaluate the bearing condition and provide supportive information for condition-based maintenance.