Abstract: With the continuous increase in the complexity of railway networks, the scheduling difficulty of railway operation departments is increasing. It is urgent to study methods for accurately predicting the duration of railway failures, in order to improve the ability of railway dispatch systems to cope with various risks and accidents. This paper was based on the text data of "Safety Supervision Report 1", combined with natural language processing techniques such as Jieba word segmentation and Word2vec word vector model, to construct a railway fault duration prediction model based on Convolutional Neural Network (CNN). The model was tested based on actual generated data from China Railway Shenyang Group Co. Ltd. The experimental results show that this prediction model can quickly and accurately obtain the duration and probability distribution of railway faults, and provide reference for train operation adjustment.