LAUSR

As the preferred mode of travel, the high-speed train (HST) and its healthy operation have received extensive attention. In the long-term service of HST, the track irregularity and wheel-rail wear may cause all kinds of faults toward the components of the bogie, which is the only connection between the train body and the track. Taking into account the unknownness of bogie fault during actual operation, it is inappropriate to simply convert the issue of bogie fault diagnosis to the problem of group classification of known faults, as conducted in almost all reference works. In this article, aiming at resolving the defect that supervised learning cannot identify unknown categories, the fractional Brownian motion (FBM) is integrated into a 1-D convolutional neural network (1-D-CNN) to distinguish unknown bogie faults from known ones. The method takes the advantage of both 1-D-CNN and FBM. On the one hand, the limitation in extracting features artificially is broken by using a convolution algorithm, and the deep features of original signals are extracted through stacking convolution kernels. On the other hand, different from the one-class theory, e.g., one-class support vector machine (OCSVM), FBM brings randomness into the network to make the model sensitive to unknown faults. Finally, four diagnosis strategies, i.e., the neural stochastic differential equation model (SDE-Net), the particle swarm optimization (PSO)-support vector data description (SVDD), the CNN-long-short term memory (LSTM)-fuzzy C-means (FCM), and OCSVM-extreme learning machine (ELM), are introduced to compare with the method proposed to verify the effectiveness and superiority. Experimental results reveal that the fractional-Brownian-motion-based network (FBM-Net) can not only classify known faults efficiently but also distinguish unknown faults with an accuracy of more than 93%.