摘要
基于铁路电务段轨道电路微机监测系统数据记录,提出一种基于注意力机制和双向长短期记忆网络的轨道电路故障预测方法。首先按照一定时间间隔记录并处理ZPW-2000A轨道电路微机监测数据,并使用合成少数类过采样算法进行数据扩充;然后基于注意力机制和双向LSTM网络构建轨道电路故障预测模型;考虑到神经网络超参数组合对模型性能有较大影响,使用基于贝叶斯采样的Hyperband算法优化轨道电路故障预测模型的超参数组合,再利用Adam算法优化模型参数,从而了解故障发生前轨道电路状态随时间的变化情况,实现对轨道电路故障状态的预测。实验表明:该模型能够准确预测轨道电路故障状态,从而指导电务段一线工作人员在故障发生前介入,提升轨道电路维护效率和运行效率。本研究为轨道电路故障预测相关研究提供一种新思路。
Based on the records of the track circuit microcomputer monitoring system in the communication and signal section,a method for predicting track circuit faults based on attention mechanism and bidirectional long short-term memory(Bidirectional LSTM)was proposed.First,a sequence of data over time was built based on the monitoring data of track circuit at a certain interval,followed by data expanding through synthetic minority oversampling technique.Then a fault prediction model was established based on attention mechanism and Bidirectional LSTM,to train the model parameters using Adam algorithm.Finally,considering the significant influences of the choice of the hyper parameters of neural network model on the performance of prediction model,the Bayesian optimization Hyperband algorithm was used to optimize the hyper parameters,to figure out the variation of track circuit status with time before the occurrence of faults,to realize the prediction of track circuit faults in the future.The experiments show that the model can accurately predict and discriminate the track circuit faults,therefore can help and direct staffs of communication and signal section to intervene before the occurrence of faults,to enhance the maintenance efficiency and operation efficiency of the track circuit.This study provides a new idea for the research of track circuit fault prediction.
作者
戴胜华
那岚
梁续继
DAI Shenghua;NA Lan;LIANG Xuji(School of Electronics and Information Engineering,Beijing Jiaotong University,Beijing 100044,China)
出处
《铁道学报》
EI
CAS
CSCD
北大核心
2023年第9期94-102,共9页
Journal of the China Railway Society
基金
国家自然科学基金(61833001)。
关键词
注意力机制
双向LSTM
故障诊断
故障预测
无绝缘轨道电路
attention mechanism
bidirectional long short-term memory
fault diagnosis
fault prediction
jointless track circuit
