Based on Dual Anomaly-aware for Time Series Anomaly Detection

Anomaly detection for time series has significant practical value, aiming to precisely identify anomaly from the normal distribution within temporal data. However, the inherent characteristics of time series data and the scarcity of anomaly bring great challenges to acquire a representative mapping function with strong discriminative power, leading to limited generalization capabilities. Meanwhile, contrastive representation learning constructs a discriminative feature space that clearly differentiates between instances, offering a more natural and promising solution for temporal anomaly detection. To address these challenges, this paper proposes a Dual Anomaly-aware Transformer model (DAT). DAT employs anomaly injection and anomaly awareness techniques to capture the features of anomaly during training. Additionally, a contextual contrastive loss function is adopted to guide dual-branch for extracting more intrinsic and discriminative semantic features between normal and anomalous samples. Comprehensive experiments conducted on five public datasets demonstrate the superiority of DAT over 18 baseline models. The results show that our model not only achieves average performance across all benchmark datasets but also performs SOTA on the MSL and PSM datasets.