Abstract:

Safety domain can be used to describe bearing health condition from the perspective of the concept of domain. It is used to identify whether a bearing is normal in a feature space established by the operation condition. However, high-speed rail vehicles usually operate in normal condition, that is, the collected data are usually normal samples rather than faulty ones. Actually, it is costly to get faulty samples for high-speed rail vehicles. Therefore, we know very little or even nothing about their fault behaviors. It means that the conventional methods, based on both normal and fault samples, cannot be applied to fault diagnosis of high speed rail vehicles. Support vector data description (SVDD) is a tool to use only normal samples in the feature space to build the so-called safety domain; however, the SVDD based models are easily influenced by the penalty parameter. Particularly in the context of big data, it is crucial to select the penalty parameter in a reasonable and quick way so that it helps to improve accuracy of both the safety domain boundary and the anomaly detection. For this reason, a penalty parameter selection method based on the distance entropy in the kernel space of safety domain is proposed. Based on the position distribution of samples in the kernel space, the method calculates the distance between every sample and the kernel center, computes the distance entropy of the training data, finds the relationship between the optimal penalty parameter and the distance entropy on the basis of a large number of reference data points, and finally provides an empirical formula to determine the optimal penalty parameter. Experimental results show that, comparing with grid search methods and genetic algorithm, the proposed method can be easily interpreted and can also significantly improve efficiency of safety domain modeling without much accuracy loss. The proposed approach is applied to anomaly detection of a wheel set bearing for high-speed rail vehicles.

Key words: anomaly detection, penalty parameter, rolling bearing, support vector data description, safety domain