A Contrastive-Representative-Learning-Based Method for Multivariate Time Series Anomaly Detection

Time series anomaly detection is extremely important in many fields. Using a series of observation data recorded at a certain time interval can detect deviations from normal behavior in a timely and effective manner. Transformer-based time series anomaly detection methods perform very well due to their self-attentive mechanism that can capture long-range dependencies in time series, but when there are complex patterns in the data, reconstruction alone may not capture enough information. In this paper, we propose a contrastive representation learning model that utilizes a novel complementary mask design to generate augmented views, constructs pairs of positive and negative samples, and learns the similarity between them, enabling the model to learn representations that distinguish between normal and abnormal data. We use a joint training optimization method of contrastive loss and reconstruction loss to enable the model to maintain a high detection accuracy even in the face of unseen abnormal patterns, thus greatly enhancing the performance and generalization ability of the model. Comprehensive experiments demonstrate that the model attains leading results across various benchmark datasets for time series anomaly detection.