Unsupervised Anomaly Detection in Multivariate Time Series through Transformer-based Variational Autoencoder

Modern industrial devices often use multiple sensors to detect the status of system, which produce a large amount of multivariate time series. Due to the complex temporal dependency of intra-channel and inter-correlations among different channels, few of proposed algorithms have addressed these challenges for anomaly detection in multivariate time series. Besides, previous work does not consider future dependency, which has been shown to be critical for sequential data modeling. In this paper, we develop an unsupervised anomaly detection algorithm TransAnomaly, which integrates Transformer, variational autoencoder (VAE) and nonlinear state space model. TransAnomaly not only reduces the computational complexity and allows for more parallelization but also provides explainable insights. To the best of our knowledge, it is the first model that combines VAE and Transformer for multivariate time series anomaly detection. Extensive experiments on several public real-world datasets show that TransAnomaly outperforms state-of-the-art baseline methods while training cost is reduced by nearly 80%.