Anomaly Detection For Time Series Based On Generative Adversarial Networks

With the popularity and development of the Internet of Things(Io T),the stability of the Io T system has become more and more important.In order to ensure the stability of the Io T system,it is necessary to monitor the behavior of the Io T devices in real time,so that when an abnormality occurs,actions can be taken in time to solve the potential problems of the Io T system.In order to monitor the behavior of Io T devices,sensors on the devices collect a large amount of time series data,so it is very important to detect anomalies in the time series data.However,the anomaly detection of the time series data is difficult to meet expectations in some aspects due to its complex time dependence and dynamics,as well as the lack of tags in the data.In order to solve the above problems,in this thesis,a new unsupervised time series data anomaly detection method based on Generative Adversarial Networks(GANs)is proposed.This method can capture the distribution of normal time series data,and then use reconstruction Error to detect anomalies.The model captures the time dependence of time series data through Gated Recurrent Unit(GRU),and uses a generative confrontation network to extract the distribution of normal data.This method can not only fully consider the time dependence of time series data,but also detect abnormalities without data tags.Based on the existing basic model of generative confrontation network and combining it with time series data,this thesis proposes two new time series data anomaly detection algorithms,including:(1)Introducing a new encoder that can extract Input the hidden variables of the data,so that the generative confrontation network can reconstruct the data,and then detect anomalies.In addition,a cyclic consistency loss function is introduced to ensure the stability of the generated adversarial network,and this loss function can further improve the efficiency of anomaly detection.(2)On the basis of the algorithm proposed in the first part,an automatic threshold selection module is introduced,which enables the algorithm to update the threshold in real time,avoiding the error caused by manual selection of the threshold,and the algorithm uses the method of moments instead of maximum likelihood estimation.Calculating the parameters greatly improves the efficiency when updating the threshold.This thesis has done a lot of comparative experiments on open source real data sets.The empirical results show that the model proposed in this thesis can not only effectively detect anomalies in time series data,but also efficiently detect anomalies.In addition,the model can be used for different types of time series data and has strong versatility.