Abstract
Anomaly detection for predictive maintenance is a significant concern for industry. Unanticipated failures cause high costs for experts involved in maintenance policy. Traditional reconstruction-based anomaly detection methods perform well on multivariate time series but they do not consider the diversity of samples in the training dataset. An abrupt change of operating conditions, which is labeled as anomaly by experts, is often not detected due to the lack of sample diversity. Besides, obtaining large volumes of labeled training data is cumbersome and sometimes impossible in practice, whereas large amounts of unlabelled data are available and could be used by unsupervised learning techniques. In this paper, we apply the principles of contrastive learning and augmentation in a self supervised way to improve feature representation of multivariate time series. We model a large variety of operating conditions with an innovative distance based stochastic method to prepare an anomaly detection downstream task. Our approach is tested on NASA SMAP/MSL public dataset and shows good performance close to the state-of-the-art anomaly detection methods.