The health management of railway vehicles is crucial to secure safety and efficiency in the long-term operation of high-speed trains. Meanwhile, complex components put forward a higher requirement for the robustness of condition monitoring systems, especially abilities to identify unexpected faults. The misidentification of infrequent faults could lead to unpredictable consequences for the vehicle's safety. This paper proposes a novel method for detecting unexpected faults of high-speed train bogie based on Bayesian deep learning. First, a Monte Carlo-Based perturbation is imposed on input samples, which can magnify the difference between unexpected faults and known ones. Then, through dropout-based Bayesian deep learning, the diagnosis result can be obtained as well as a Bayesian indicator of whether the anomalies belong to known classes. The proposed method can capture the uncertainty of model outputs and identify unexpected faults, requiring only a few samples of unexpected anomalies for calibration. Also, it is compatible with most existing neural network structures. The experiments compare the proposed method with existing methods on two real-world applications, which demonstrates the effectiveness and superiority of the proposed scheme.