Research On Anomaly Detection Method Of Time Series Data In Cloud Data Center Based On Deep Learning

The rapid development of cloud computing and big data has led to the trend of large-scale cloud data centers.In order to ensure the stable operation of the services required by users,anomaly detection of monitoring indicators is required in the cloud data center.However,due to the characteristics of many monitoring indicators and large data volume,data labeling,model selection,training,and parameter adjustment are required for anomaly detection on key performance indicators(KPI)of monitoring data,which consumes a lot of labor and economic costs.Traditional anomaly detection methods can no longer meet the needs of large-scale data.In order to solve the above problems,this thesis constructs deep learning anomaly detection algorithms for time series data based on feature fusion and data reconstruction.(1)At present,the characteristics of operational data are complex,and system abnormalities may be related to multi-dimensional internally associated monitoring data characteristics.The common anomaly detection algorithms mostly target single-dimensional data features and cannot fully capture the deep features of the data.If these characteristics are aggregated together for analysis and processing,anomaly detection can be improved the reliability of the algorithm.This thesis proposes an anomaly detection algorithm based on CNN-Bi LSTM-Attention feature fusion.CNN can effectively extract the spatial dimensional features of the data.LSTM obtains the temporal dimensional features of the time series data through the storage unit.The attention mechanism can effectively weight the spatial and temporal features,so as to effectively obtain the deep fusion features of the data.The algorithm has been used to detect anomalies on KPI monitoring data and achieved good results,with an accuracy rate of 97%,and F1-score of 0.94.The detection effect is significantly better than the result obtained by using a single model.(2)Based on the supervised anomaly detection algorithm,it takes a lot of manpower and economic costs to label the time series data of the cloud data center.At the same time,many algorithms cannot fully capture the related features in the time dimension.In this thesis,by introducing the VAE,an unsupervised time series anomaly detection algorithm based on the VAE-Bi LSTM hybrid model is proposed.The VAE module can extract stable local features in a short time window by reconstructing the features of the time series data,and the bidirectional LSTM can predict the feature window of the VAE from the context information,which improves the accuracy of the prediction.Therefore,the detection algorithm can identify a variety of time series anomalies.At the same time,the effectiveness of the algorithm is verified on the real data set,and its accuracy rate reaches 91%,and the comprehensive index F1-score can reach 0.94.At the same time,it is compared and verified with other unsupervised methods.The experimental results show that the proposed algorithm is better than others commonly used unsupervised anomaly detection methods.