TCAE: Temporal Convolutional Autoencoders for Time Series Anomaly Detection

Abstract

Prevalent recurrent autoencoders for time series anomaly detection often fail to model time series since they have information bottlenecks from the fixed-length latent vectors. In this paper, we propose a conceptually simple yet experimentally effective time series anomaly detection framework called temporal convolutional autoencoder (TCAE). Our model imposes dilated causal convolutional neural networks to capture temporal features while avoiding inefficient recurrent models. Also, we utilize bypassing residual connections in encoded vectors to enhance the temporal features and train the entire model efficiently. Extensive evaluation on several real-world datasets demonstrates that the proposed method outperforms strong anomaly detection baselines.