Research On Unsupervised Anomaly Detection Of Time Series Data Based On Variational Auto Encoder

In recent years,with the rapid development and large-scale application of computer technology,there are more and more Internet-based services.However,there are more and more abnormal events caused by machine failure,network interruption and user misconduct.In addition to bringing bad experience to users,these abnormal events may even cause business failure,equipment downtime and other problems.In order to ensure the stability and reliability of business,Internet companies usually need a large number of operation and maintenance personnel to monitor business indicators and machine indicators for a long time,and repair faults as soon as possible when abnormalities occur,which brings great challenges to the operation and maintenance scenario.Therefore,in order to share the pressure of operation and maintenance personnel,the research on reasonable and efficient anomaly detection strategy has attracted more and more attention.To meet the above challenges,this paper studies the unsupervised algorithm of time series anomaly detection based on Variational Auto Encoders(VAEs).The Variational Auto Encoders model is used to detect the time series data,determine whether there is an abnormal event according to the reconstruction probability of the data to be tested,and detect the abnormality as soon as possible to reduce the loss caused by the abnormal time.The main work of this paper is as follows:(1)An unsupervised anomaly detection algorithm based on Variational Auto Encoders and local outlier factor(LOF)is proposed.The algorithm mainly combines the Local Outlier Factor algorithm and Variational Auto Encoders algorithm to train without using any manual labels,and applies the trained model to the anomaly detection of stream data.In the preprocessing stage,the algorithm processes a small number of parts that are difficult to fit by Variational Auto Encoders through Local Outlier Factors,and optimizes the loss function accordingly to improve the ability of the algorithm to identify anomalies.(2)An unsupervised anomaly detection algorithm with high sensitivity based on conditional Variational Auto Encoders(CVAEs)is proposed.The algorithm preprocesses the data through time series information extraction and comparison,combined with Local Outlier Factor algorithm,and takes the extracted time information as an additional condition for training Variational Auto Encoders,so as to improve the sensitivity of the model.It can identify anomalies quickly and well without using any manual markers,that is,it can still show good performance in the case of low delay.