Research on anomaly detection of multivariate time series data based on VAE-LSTM
With the popularity of the Internet of Things,more and more application areas need to process multivariate time se-ries data.This type of data is characterized by multiple simultaneous dimensions and complex temporal dependence.Water quality monitoring is one of the typical application scenarios,and a variety of water quality parameters(such as temperature,pH,dissolved oxygen,conductivity,etc.)need to be monitored and recorded in real time.These data are not only massive,but also have high di-mensions and strong time series correlations,which are difficult for traditional detection methods to handle.In order to solve this challenge,this paper proposes a VAE-LSTM multivariate time series anomaly detection method,which captures the latent distribu-tion characteristics and time-dependent characteristics of data by training a variational autoencoder and a long short-term memory model.By labeling data with large reconstruction errors,the model can effectively detect abnormal data points.Experimental re-sults show that the F1 score of the proposed method on the water quality monitoring dataset of the Internet of Things is 88.5%,which shows better anomaly detection performance than the KNN,VAE and LSTM models.
multivariate time seriesanomaly detectionvariational autoencoderlong short-term memory networksunsuper-vised learning
万方数据
