Anomaly Detection Framework With Contrastive Learning and Multiview Augmentation for Time-Series Domain Generalization

Time-series data are generated and collected in manufacturing processes. Time-series data have different distribution characteristics, with the underlying assumption being that all data originate from the same distribution. Time-series anomaly detection methods are trained on stationary data derived from the source domain. However, the performance of the trained model diminishes when confronted with a nonstationary distribution in the target domain. Consequently, there is a need for a model that can adapt to any distribution using only the source domain. In this study, we propose an anomaly detection framework with contrastive learning and multiview augmentation for time-series domain generalization. The proposed method uses multiview augmentation to learn three representations for time-series domain generalization. Moreover, the augmented samples use a temporal convolutional network (TCN) to extract these representations. Subsequently, the extracted representations are trained with a contrastive learning method inspired by simple triplet representation learning (SimTriplet). Therefore, the proposed method can attain the diversity and generalization performance of the source domain. The performance superiority of our proposed method is experimentally validated across various domain shift scenarios. Additionally, the effectiveness of the modules used in the proposed method is confirmed through qualitative analysis.