Network Fault Diagnosis Technology Based On Data-Driven

Nowadays,increasingly complex network environments and scenarios have put forward higher requirements for network fault diagnosis technology.Traditional manual fault detection methods cannot cope with minor faults that occur in large-scale network scenarios,and the existing network monitoring hardware or software will have an impact on the monitored network environment.In addition,these hardware and software have very high requirements on technical personnel regarding network protocols and network management techniques.After investigating and analyzing the related technologies of network fault diagnosis,this thesis focuses on the data-driven network fault diagnosis method.Firstly,this thesis analyzes the back propagation(BP)neural networks,radial basis function(RBF)neural networks and convolutional neural networks(CNN)in machine learning.The mini-batch gradient descent(MBGD)algorithm is introduced to optimize the parameters of the BP neural networks.The supervised learning method is used to modify and determine the center and expansion constant of the RBF neural networks.The MBGD algorithm,adaptive moment estimation algorithm and dropout algorithm are used to optimize CNN.Three network fault detection algorithms and models based on BP,BRF and CNN are established.The accuracy and feasibility of those algorithms are verified in the KDDCUP99 and NSL-KDD data sets.The network fault detection algorithm based on CNN has fast training speed,high accuracy and good applicability.Secondly,a passive network data collection method based on bypass mirroring is given,which collects network data without affecting the monitored network topology and without occupying the monitored network resources.This thesis builds a real network environment scenario,designs and produces physical disconnection,network congestion,and equipment downtime.After that,the WireShark software is used to collect network data packets,and the original network data packets of more than 200 hours are obtained.A data dictionary is constructed by analyzing the content of WireShark,and an initial data set containing 39 features is constructed by extracting the feature information of the transport layer and below in the original data packets.Finally,a scheme is proposed to extract the initial data set of network faults based on traffic characteristics.The statistical information in the initial data set is extracted through a one-second time window,and a network fault data set with 38 features is constructed.In this thesis,three algorithms based on BP,RBF,and CNN are used to test the data set.The algorithm based on BP can reach an accuracy rate of over 87.4%,but as the number of iterations increases,over-fitting will occur.The algorithm based on RBF can reach an accuracy rate of over 86.5%,while it needs a lot of time to train the model.The algorithm based on CNN can reach an accuracy rate of over 88.3%,which takes the least time to train the model.The experimental results verify that the data set constructed in this thesis has a certain application significance for network fault diagnosis.