Adaptive Memory Networks with Self-supervised Learning for Unsupervised Anomaly Detection

Unsupervised anomaly detection aims to build models effectively detect unseen anomalies by only training on the normal data. Although previous reconstruction-based methods have made fruitful progress, their generalization ability is limited due two critical challenges. First, dataset contains patterns, which limits model ability. Second, feature representations learned existing often lack representativeness hampers preserve diversity of patterns. In this paper, we propose a novel approach called Adaptive Memory Network with Self-supervised Learning (AMSL) address these challenges and enhance in unsupervised detection. Based convolutional autoencoder structure, AMSL incorporates self-supervised learning module learn general patterns an adaptive memory fusion rich representations. Experiments four public multivariate time series datasets demonstrate that significantly improves performance compared other state-of-the-art methods. Specifically, largest CAP sleep stage 900 million samples, outperforms second-best baseline 4%+ both accuracy F1 score. Apart from enhanced ability, also more robust against input noise.