Research And Implementation Of Time Series Anomaly Detection Based On Self-learning Bidirectional GAN

As more and more network attack events and system failure events appear in various operation and maintenance systems based on time series databases,a small amount of abnormal data is mixed into the huge database,which has a certain impact on the stable operation of the system.Therefore,time series anomaly detection has become a hot issue in academia and industry.However,traditional threshold anomaly detection methods and general deep learning models are unable to make complete analysis and accurate anomaly detection for time series data.Therefore,this thesis takes time series data as the research object,proposes an improved anomaly detection algorithm,and builds an anomaly detection system based on this algorithm.The main contributions are as follows:(1)In view of the problem that the existing anomaly detection model has an imbalance of positive and negative class data during the training phase,the generative adversarial network is selected as the basic anomaly detection model.This thesis proposes a self-bidirectional learning training method based on GAN.The GAN model trained by this method can improve the efficiency of anomaly detection while optimizing the defect problems of the original GAN model.(2)Since the generative adversarial network itself does not have strong processing capability for time series data,a Self-GRU network architecture is proposed in this thesis as the basic model of each part of the generative adversarial network.Self-GRU integrates the self-attentive mechanism and the gated recurrent unit neural network to accurately capture the developing trends in time series data.(3)Based on the above research content,an anomaly detection system based on self-learning bidirectional GAN is designed and implemented by using front-end and back-end separation technology.The system can currently realize the functions of real-time monitoring of time series data,abnormality detection and alarm,etc.