One-class predictive autoencoder towards unsupervised anomaly detection on industrial time series

Anomaly detection is a fundamental problem of data science that aims at finding instances of unusual data. In recent years, due to the rapid expansion of the Industrial Internet of Things (IIoT), substantial amounts of high-dimensional industrial time series data have been generated. Detecting potential anomalies from such data is challenging and an important research topic. In this paper, we propose One-Class Predictive Autoencoder (OCPAE), a novel encoder-decoder approach with additional prediction and one-class branches to enhance the performance on detection of time series anomalies from different perspectives. The prediction branch can detect anomalies by learning the local temporal dependency while the one-class branch is suitable to learn the normal patterns from a global perspective. We evaluate our proposed approach on five public datasets and demonstrate the superiority of our approach over other state-of-the-arts methods. Lastly, we conduct ablation studies and in-depth analysis to show the effectiveness, efficiency, and robustness of our proposed method.