Time Series Intelligent Anomaly Detection In SDN

With the development of artificial intelligence technologies such as big data and machine learning,the deployment of B5G(Beyond Fifth Generation)and 6G(Sixth Generation)intelligent networks has promoted the development of intelligent operation in the network field.SDN(Software Defined Networking)provides functions such as automation management and network programming.However,with the expansion of network scale and the gradual complexity of information transmission,maintaining such dynamic characteristics of the network requires intelligent operation and maintenance mechanisms.This paper proposes to implement the closed-loop control of SDN based on AIOps(Artificial Intelligence for IT Operations)technology.Modeling through time series indicators such as link delay,network throughput,and device CPU utilization in the network can detect anomalies in the network.The current supervised detection algorithm for time series data regards the anomaly detection task as a binary classification problem,and usually extracts the features in the time series window as the basis for distinguishing anomalies.However,the existing algorithms mainly focus on the time series data itself,which is not universal for time series data with various forms.This article applies the saliency detection in the computer vision field to time series data.Transformation and coordination of time series data based on spectrum information can weaken background data and enhance the saliency of abnormal data in time series,which can improve the effect of abnormal detection.Regarding the feature processing of time series modeling,this paper adopts wavelet decomposition and threshold coordination method to reduce the influence of noise on time series data modeling.The PTA-DNN algorithm proposed in this paper uses saliency analysis features,statistical features,time series modeling features,and wavelet features as feature sets to perform anomaly detection on time series data.The F-score on 25 different types of time series data sets is 0.948,which is better than the current supervised anomaly detection algorithm.Supervised algorithms rely on labeling time series data,which consumes a lot of human resources.For the time series data associated with the topology map in the network,it is difficult to obtain the comprehensive characteristics of the time series based on the time dimension characteristics alone.This paper proposes a graph-based gated convolutional codec unsupervised anomaly detection model GAD.Taking into account the real-time nature of the network topology detection scene and the spatial topology connection relationship,this paper uses gated convolution to extract features in parallel for time series data and relies on graph convolutional networks to mine spatial dependencies.After the encoder composed of the spatiotemporal feature extraction module encodes the time series,the decoder composed of the convolutional network is used to reconstruct the time series.Finally,the residuals of the input time series data and the reconstructed time series data are used to detect anomalies.Experiments on public data sets and simulation platforms show that the GAD model has a higher recognition accuracy than the unsupervised anomaly detection baseline algorithm.