Short Text Mining for Fault Diagnosis of Railway System Based on Multi-Granularity Topic Model

摘要

Vehicle equipment is one of the core equipments of train control system in the high-speed railway, and it mainly depends on the experience of maintenance personnel to perform fault diagnosis when it fails, which is ineffective. The maintenance text data of vehicle equipment, which contains fault category information, is not fully utilized. The classification of maintenance text can well assist fault diagnosis of vehicle equipment. However, shortness and imbalanced class distribution of maintenance texts, which hinder the application of conventional text representation models and classification algorithms, pose challenges for classification task. In this paper, we propose a novel text feature selection algorithm based on multi-granularity latent Dirichlet allocation (LDA) to overcome shortness characteristic of maintenance text. To solve the problem of class imbalance, a cost-sensitive Support Vector Machine (SVM) is utilized to construct fault diagnosis model. Finally, we compare our proposed method with the state-of-the-art baseline over a vehicle maintenance text data set collected by Guangzhou railway corporation, and it outperforms traditional methods.