Research On Data-driven Fault Diagnosis And Prediction For LTE-R Network

With the rapid development of China's heavy-duty freight railway and the continuous improvement of its safety performance requirements,the railway wireless network communication system,ensuring the safe operation of the railway,becomes to show its important value.In order to ensure the safe operation of heavy haul trains in the 20-thousand-ton class and provide support for railway business,the Shuohuang Railway Company has established the world's first LTE(Long Term Evolution)based communication network.At the same time,with the increasing complexity and scale of LTE-R network structure,the traditional network fault operation and maintenance method based on manual mode is no longer sui Table for the current production environment,while intelligent and automated network fault diagnosis and prediction has gradually developed into the objective needs of the current network business.Aiming at this problem and combining with the actual field operation and demand of the railway,this thesis follows the real business scenario of Shuohuang Railway and focuses on the research of network fault diagnosis and prediction during the operation and maintenance of the LTE-R network.Specifically,aiming at the current network fault problem,this thesis analyzes the network key parameters,and makes correlation analysis and mining of the dropping rate of network communication quality E-RAB(Evolved Radio Access Bearer)to realize the automation and intelligence of network operation and maintenance.Research products are as follows:(1)According to the characteristics of large amount of data and complex structure of LTE-R network,this paper analyzes the key parameters and faults of LTE-R network,and proposes a method of basic faults diagnosis of LTE-R network based on Spark.This method realizes the storage,reading and calculation of massive data by means of big data technology,and It can quickly and efficiently filtrate the network basic faults.(2)According to the characteristics that the key parameters of LTE-R network are time series data,a method of time series extraction shapelets based on genetic algorithm is proposed for feature extraction of LTE-R network parameters.This method can quickly calculate the most appropriate shapelets set in the time series,and can screen out the hidden shapelets time series features,which has better interpretability.(3)Aiming at the task of LTE-R network fault diagnosis,an attention-aware deep tree ensemble learning method is proposed for fault diagnosis.This method can effectively enhance the ability of representation learning for input data,and the introduced attention mechanism can significantly improve the performance of model in real-world LTE-R network fault diagnosis task.(4)Aiming at the task of LTE-R station network fault prediction,a transformer-based time series prediction method with local enhancement of causal convolution is proposed.This method can effectively mine the dependency among temporal features,and enhance the local attention-aware association in context.Furthermore,some reliable performance analysis is provided in E-RAB dropout rate fault prediction.