Revisiting VAE for Unsupervised Time Series Anomaly Detection: A Frequency Perspective.

Time series Anomaly Detection (AD) plays a crucial role for web systems.Various web systems rely on time series data to monitor and identify anomaliesin real time, as well as to initiate diagnosis and remediation procedures.Variational Autoencoders (VAEs) have gained popularity in recent decades due totheir superior de-noising capabilities, which are useful for anomaly detection.However, our study reveals that VAE-based methods face challenges in capturinglong-periodic heterogeneous patterns and detailed short-periodic trendssimultaneously. To address these challenges, we propose Frequency-enhancedConditional Variational Autoencoder (FCVAE), a novel unsupervised AD method forunivariate time series. To ensure an accurate AD, FCVAE exploits an innovativeapproach to concurrently integrate both the global and local frequency featuresinto the condition of Conditional Variational Autoencoder (CVAE) tosignificantly increase the accuracy of reconstructing the normal data. Togetherwith a carefully designed "target attention" mechanism, our approach allows themodel to pick the most useful information from the frequency domain for bettershort-periodic trend construction. Our FCVAE has been evaluated on publicdatasets and a large-scale cloud system, and the results demonstrate that itoutperforms state-of-the-art methods. This confirms the practical applicabilityof our approach in addressing the limitations of current VAE-based anomalydetection models.