Research On SDN Fault Diagnosis Based On Machine Learning

With the continuous expansion of computer network scale and the emergence of various new applications,the traditional network infrastructure has become increasingly unable to meet the new needs of users.Therefore,software defined network(SDN)came into being.Different from the traditional network structure,SDN realizes the separation of control surface and data surface,reduces the complexity of network management,and provides programmability for the network,which has great application prospects.However,while bringing many advantages,SDN still faces the reliability problem that has always existed in traditional networks.Therefore,fault diagnosis has always been a research focus in SDN research.This research mainly focuses on the congestion fault of SDN network.At present,the biggest obstacle in the research of SDN network fault diagnosis is that there is not enough fault sample data for fault diagnosis based on machine learning algorithm,which leads to low accuracy of fault diagnosis and is only suitable for small-scale networks.Therefore,this thesis proposes a simulation algorithm of SDN link performance parameters based on Generative Adversarial Networks(GAN)and a SDN fault location algorithm based on Convolutional Neural Networks(CNN),which uses the characteristics of GAN model to generate more simulated link performance parameters close to the real link performance parameters for the training of CNN fault location model;The CNN model can compress the data while extracting the data features,find the characteristics of the corresponding relationship between the input data and the output data through a large amount of data training,find the mapping relationship between the link performance parameters before congestion and the link congestion type,and realize the congestion prediction of SDN network.Based on the above ideas,this research designs and builds a SDN fault diagnosis system based on machine learning,tests the fault diagnosis performance of the whole system through experiments,and compares the influence of the generated data of different GAN models on the accuracy of CNN fault diagnosis algorithm.The experimental results show that the GAN model can generate more simulation samples similar to the real samples through a small number of real samples,so as to ensure the training convergence speed and fault diagnosis accuracy of CNN fault location model.Among them,the simulation effect of Deep Convolutional Generative Wasserstein Adversarial Networks(DCWGAN)is the best.