Abstract:

In order to improve data utilization rate of on-board equipment on high speed trains during operation and maintenance, with fault text data of the most representative CTCS3-300T equipment in CRH2 and CRH3 EMUs as an example, a fault diagnosis model based on BiLSTM and CBA was proposed Firstly, used Word2vec tool to train word vector for fault text. Secondly, for the problem of unbalanced distribution of fault data, small category text vector data were automatically generated by SMOTE algorithm. Then, BiLSTM was utilized to obtain fault text features. Finally, fault diagnosis was accomplished by CBA algorithm., and text data of on-board equipment of a railway bureau in the past 5 years were experimentally analyzed. The results show that the proposed model can make precision and recall rate of fault diagnosis reach 95.66% and 96.29% respectively. And compared with the model without SMOTE algorithm, its recall rate has increased by 11.77%, which not only guarantees accuracy of general classification, but also gets better classification performance of minority classes.

Key words: Bi-directional long short-term memory (BiLSTM), classification based on associations (CBA), on-board equipment, fault diagnosis, synthetic minority over-sampling technique (SMOTE)