Gaussian Mixture Variational Autoencoder with Whitening Score for Multimodal Time Series Anomaly Detection

Abstract

Time series anomaly detection has attracted great attention due to its widespread existence in real life. With the increasing development and advancement of deep learning, many unsupervised deep learning methods have been proposed for time series anomaly detection since labeling time series is prohibitively expensive. In this paper, we propose an unsupervised anomaly detection method: Gaussian Mixture Variational Autoencoder with Whitening Distance Anomaly Score (WGVAE). Concretely, we employ an LSTM-based variational autoencoder to capture the long-term dependence of time series and learn the low-dimensional feature representation and distribution, in which the Gaussian mixture prior are used to characterize multimodal time series. Further, whitening distance anomaly scores are used to make the multidimensional time series independently and identically distributed among each dimension, which combines the distribution characteristics of the samples to measure the degree of outliers. When the anomaly score is below the threshold, the sample is detected as anomalous. Finally, comprehensive experiments are given to verify the effectiveness of our method.