TDSRL : Time Series Dual Self-Supervised Representation Learning for anomaly detection from different perspectives

Time series anomaly detection plays a critical role in various applications, from finance to industrial monitoring. Effective models need to capture both the inherent characteristics of time series data and the unique patterns associated with anomalies. While traditional forecasting-based and reconstruction-based approaches have been successful, they tend to struggle with complex and evolving anomalies. For instance, stock market data exhibits complex and ever-changing fluctuation patterns that defy straightforward modelling. In this paper, we propose a novel approach called TDSRL (Time Series Dual Self-Supervised Representation Learning) for robust anomaly detection. TDSRL leverages synthetic anomaly segments which are artificially generated to simulate real-world anomalies. The key innovation lies in dual self-supervised pretext tasks: one task characterises anomalies in relation to the entire time series, while the other focuses on local anomaly boundaries. Additionally, we introduce a data degradation method that operates in both the time and frequency domains, creating a more natural simulation of real-world anomalies compared to purely synthetic data. Consequently, TDSRL is expected to achieve more accurate predictions of the location and extent of anomalous segments. Our experiments demonstrate that TDSRL outperforms state-of-the-art methods, making it a promising avenue for time series anomaly detection.