Self-Supervised Anomaly Detection Using Outliers for Multivariate Time Series

Abstract

Due to the difficulty of having sufficient labeled data, self-supervised learning (SSL) has recently got much attention by many researchers in time series anomaly detection. The generative adversarial network (GAN) based autoencoder model, one of the SSL models, has good performance on anomaly detection but it tends to be too sensitive (i.e., predict normal data with a small anomalous value as abnormal). In this paper, we find that mispredicted normal data have values far from the average on some sensors. We call these data as outliers. Since these data are a few in the training set, the model struggles to reconstruct these data and incorrectly predicts them as abnormal. Based on these findings, we propose a robust self-supervised anomaly detection framework that finds outliers using a clustering based on correlation features and uses them for efficient training. To evaluate our method, we compare with various deep learning-based anomaly detection methods on the real-world pump dataset. The results demonstrate the superiority of our proposed method. Through our method, we maintain sensitivity to abnormal data while reducing sensitivity to normal data with a small anomalous value.