Fault Analysis Of Train Real-time Ethernet Based On Data-driven

With the continuous improvement of train intelligence,train communication networks need to undertake more various data transmission tasks.Due to the advantages of high transmission rate,large amount of transmitted data and strong integration ability,real-time Ethernet has become the future solution of train communication networks.However,the failure of the train communication network has a great impact on the safe and stable operation of the train.Once the failure of the equipment in the train can not be solved in time,it will cause huge hidden dangers and economic losses.The fault of existing train communication network is often solved by the expert experience of the maintenance personnel.The fault diagnosis efficiency is low and the subjective factors to judge the fault are too many,so the network fault can not be diagnosed timely and accurately.With the support of the school-enterprise cooperation project,this dissertation collects the data of normal operation and fault injection of train real-time Ethernet,extracts network features,and carries out fault diagnosis of network by combining machine learning method.The main work of this dissertation is as follows:(1)The layered data acquisition method of train real-time Ethernet is studied and implemented,and features are extracted to represent different network states.It is divided into two parts to collect data.One part takes the train cable as the test object,the physical layer characteristics are collected,and the electrical characteristic parameters,which can reflect the cable state are extracted and analyzed.In the other part,TRDP is used as the test object to collect data above the data link layer.A scheme based on SNMP and Wireshark is designed to collect data together.Through the processing and analysis of the original data,the characteristic parameters,which can reflect the state of the network are selected to monitor the state of the network.(2)In order to improve the fault diagnosis data set,the injection of common faults in each layer of the network is realized.For the physical layer wiring system,five kinds of fault injection are realized: cable open circuit,short circuit,jumper,virtual connection and stress;on the link layer,six kinds of fault injection are realized: link blocking,device disconnection,address assignment fault,switch connection fault,port unsubscribed and application program fault.(3)In order to avoid the defect that the traditional fault diagnosis method relies too much on the expert experience of maintenance personnel,a random forest fault diagnosis model for train real-time Ethernet is established.In order to supplement network state data set and simplify model calculation,data preprocessing methods such as balanced sampling and normalization are used.Cross validation and grid search are used to optimize the parameters.Based on python,the interface of fault diagnosis software is designed.(4)Combined with the random forest fault diagnosis model,the real-time Ethernet of train is tested.For the physical layer cabling system,the data set is the actual cable test of a train.On the data link layer,the TRDP experiment platform is built to simulate the train communication network topology in the laboratory environment,and the network state feature data set is formed.Experimental results show that,compared with the two commonly used classification algorithms of support vector machine and decision tree,the random forest fault diagnosis model has higher classification accuracy.It proves that the diagnosis model has good effect on the real-time Ethernet data diagnosis of trains,and can provide relevant basis for train network fault maintenance.