Time series anomaly detection based on VAE-LSTM hybrid model

Abstract

This paper introduces VAE-LSTM model which is a hybrid model for detecting anomalies in time series based on unsupervised learning. Anomalies in time series often indicate potential failures or abnormal behaviors, which are of great importance in industrial monitoring, network security, financial forecasting and other fields. The VAE-LSTM model introduced in this paper improves and extends the traditional VAE model by introducing the LSTM module, thereby improving its capacity to capture long-range dependencies and helping to overcome the limitations of VAE in detecting long-term anomalies. This model learns time series' latent representations of local segments by VAE and model long-term temporal dependencies by LSTM. Therefore, with the VAE-LSTM architecture, anomalies spanning different temporal lengths—both short and long—can be accurately detected. Experiments are performed on publicly benchmark datasets including NAB and PSM. This thesis uses three sequences from the NAB dataset to make anomaly detection in univariate time series: cpu_utilization, nyc_taxi, and ec2_request, achieving F1 scores of 0.8958, 0.9789, and 0.9936, respectively. Multivariate time series anomaly detection experiments were conducted on the PSM dataset, and the F1 score was 0.8351. This confirms that the model introduced in this paper performs well in anomaly detection.