ExTAD: Embedding Exchange Inspired Time Series Anomaly Detection With Modal Consistency

Comprehensive learning of temporal representations is crucial for time series anomaly detection (TSAD). Frequency domain analysis has been proven to be an effective strategy for detecting diverse patterns of anomalies. However, existing methods ignore the exploration of complementary information of modalities and the anomaly misjudgments caused by time-frequency distribution shifts. In this article, we propose Embedding exchange-inspired unsupervised Time-series Anomaly Detection (ExTAD) with modal consistency to learn multimodal representations that improve detection performance. Specifically, we employ a weight-sharing dual-autoencoder (dual-AE) to extract the time-frequency embedding. Then, embedding exchange is used to learn robust perturbation-invariant representations for mining multimodal complementary information. Furthermore, we design a modal consistency strategy to align time-frequency representations to a unified scale, alleviating anomaly misjudgments caused by distribution shifts. Finally, we combine modal consistency loss and reconstruction errors to perform time series anomaly detection tasks. Experimental results on six benchmark datasets show that ExTAD achieves state-of-the-art performance compared to existing methods.