云边协同联邦计算方法在铁路信号系统故障检测中的应用

A Cloud Edge Coordinated Federated Computing Method for Fault Detection in the Railway Signal System

铁路信号系统是当下社会交通运力的主要承载系统,其对安全性有极高的要求.而由于铁路信号系统容易受到外界多种因素影响,易出现故障,需要设计一种针对铁路信号系统的实时故障检测方案,进而才能采取有效的维护措施.不同于传统的机器学习(ML)故障检测方法,采用双向编码器表示转换器(BERT)深度学习(DL)模型进行实时的智能故障检测.该模型能够在处理故障检测任务时获取双向上下文的理解,从而更准确地捕捉句子中的语义关系,使得其对故障描述的理解更为精准.采用了云边协同的联邦计算方法,使得各铁路运营单位的数据可以在本地进行初步处理,然后将汇总后的梯度上传至云端进行模型训练,最终将训练得到的模型参数发送回各边缘设备,实现模型的更新,突破了模型的训练数据分散的限制,同时允许多个铁路运营单位在保持数据隐私的前提下共同训练BERT模型.研究结果表明,采用联邦边云计算方法进行BERT模型训练,在解决数据保密性问题的同时,有效提升了轨道交通故障检测的准确性与可靠性,优于目前在铁路信号系统领域已有的故障检测方案.

The railway signaling system is the main carrying system of transportation capacity in the current society,and it has extremely high requirements for safety.However,due to the susceptibility of the railway signal system to various external factors and the possibility of malfunctions,it is necessary to design a real-time fault detection scheme for the railway signaling system,so as to take effective maintenance measures.Different from the traditional machine learning fault detection method,the bidirectional encoder representation from transformer(BERT)deep learning model is used for real-time intelligent fault detection.The model can obtain the understanding of the two-way context when dealing with the fault detection task,so as to capture the semantic relationship in the sentence more accurately,and make the understanding of the fault description more accurate.The federated computing method of cloud-edge collaboration is adopted,so that the data of each railway operator can be preliminarily processed locally,and then the summarized gradient is uploaded to the cloud for model training.Finally,the trained model parameters are sent back to each edge device to realize the update of the model,breaking through the limitation of scattered training data of the model,and allowing multiple operators to jointly train the BERT model under the premise of maintaining data privacy.The results show that the federated edge cloud computing method for BERT model training can effectively improve the accuracy and reliability of rail transit fault detection while solving the problem of data confidentiality,which is better than the existing fault detection schemes in the field of railway signaling system.