Research On SDN Fault Prediction Based On Deep Learning

With the increasingly close relationship between computer network and people,new needs are born constantly,and the performance requirements of network are gradually improved,but some traditional networks can not meet the requirements gradually.Compared with traditional Networks,Software Defined Networks(SDN)can separate the control plane from the data plane,reduce the complexity of the network,and facilitate development and management.However,like traditional networks,SDN also has faults such as switch equipment failure and link congestion.Therefore,the study of SDN network fault’s management is also an important research direction of SDN network.Compared with fault detection and diagnosis,fault prediction provides early warning before a fault occurs,allowing network maintenance personnel more time to prepare,and has high practical value in improving network reliability.Deep learning has shown good prediction ability in the field of fault prediction.Meanwhile,in view of the difficult problem of insufficient fault data encountered in the fault research of SDN network,the following work is carried out in this thesis:(1)Build a custom SDN network based on the Mininet simulation platform,design network services to generate different network faults,and collect and store current network fault data.Perform a state assessment of the current network and combine existing data to obtain the original dataset.(2)Due to the contingency and uncertainty of faults,the occurrence frequency is small,leading to the small number and types of fault samples obtained in the normal operation of SDN.The data set used for fault prediction is not perfect,which has a great impact on the results.Therefore,based on the data characteristics of Generative Adversarial Networks(GAN),which can generate approximately real data after learning the real data,a fault data generation algorithm for SDN networks is designed.Combined with the convolution idea and Wasserstein distance with gradient penalty as a loss function,a DCWGAN-GP model was designed to generate fault data and obtain more network fault data.After testing,compared with traditional GAN models,the generated data of DCWGAN-GP is closer to the real data.(3)According to the characteristics of time series presented by SDN fault data,CNNLSTM and TCN-Bi LSTM fault prediction algorithms were designed by relying on the prediction ability of cyclic neural network to time series and the extraction ability of convolutional network to data.Convolutional Neural Networks(CNN)and Temporal Convolutional Network(TCN)combining Attention mechanism realize feature extraction.Long Short-Term Memory(LSTM)and Bi-directional Long Short-Term Memory(Bi LSTM)learn from network fault data to find deep characteristics.To achieve future time failure prediction.After testing,compared with LSTM,CNN-LSTM and other models,TCN-Bi LSTM network combined with Attention mechanism has excellent predictive ability.