Fault classification method for on-board equipment of metro train control system based on BERT-CNN

Abstract

The on-board equipment of metro communication based train control (CBTC) is facing laborious maintenance problems, and its textual maintenance logs are criticized for having excessively fragmented information, ambiguous semantics and confused categorization, resulting in low classification metrics by traditional textual distributed representation with basic machine learning algorithms. A fault classification method based on bidirectional encoder representations from transformers - convolutional neural network (BERT-CNN) with the focal loss function is proposed to establish the relationship model between the 'fault processing and conclusion' and the 'fault phenomena'. The pre-trained bidirectional encoder representations from transformers (BERT) model is fine-tuned to fully capture the bidirectional semantics and focus on the keywords to produce better word vectors of the 'fault phenomena'. In order to counteract the classification performance degradation brought by data category imbalance, word vectors are trained using a convolutional neural network (CNN) model with the focal loss function. According to the experimental results conducted by the dataset from an on-board signaling department, the proposed method has the best classification performance among models of BERT-CNN, single BERT and word to vector - CNN (word2vec-CNN) using cross-entropy loss function, and it is also better to correctly classify categories with few samples and contributes to the development of a more comprehensive library of fault cases for intelligent operation and maintenance.

Keywords

• transportation engineering
• on-board equipment of metro train control system
• bert language model
• convolutional neural network (cnn)
• fault classification
• category imbalance