Scale-independent shrinkage broad learning system for wheelset bearing anomaly detection under variable conditions

Abstract

Wheelset bearing fault diagnosis is essential to guarantee safe rail system operation. However, real-time monitoring signals of high-speed trains are nonstationary under variable conditions. Additionally, wheelset bearings operate in a normal state most of the time, and therefore, only normal data are available for training in practice. Under such a scenario of variable operating conditions and lacking labelled data, it is still necessary to timely detect the possible abnormal states of wheelset bearings. Currently, one-class classification is a suitable approach for real-time anomaly detection, but its accuracy is affected by the linear relationship of network nodes. Concerning this issue, this paper presents a novel method named scale-independent shrinkage broad learning to detect the abnormal states of wheelset bearings in real time. First, scaled rotation regularization is presented to convert the output constraint into the node distance constraint. Then, the variance-guided mapping mechanism is constructed to solve the linear relationship between nodes. Meanwhile, the optimal network mapping parameters are generated iteratively according to a so-called high-polymerization strategy, with which new output weights are formed after the feedback calculation. This method eliminates the linear correlation of node features during the iteration and realizes one-class classification detection of abnormal states. Finally, wheelset bearing data collected from a test rig of a Chinese rolling stock company are utilized to demonstrate the feasibility and effectiveness of the proposed approach in variable conditions.