PAI: Preserving Amplitude Information in Representation-Based Time-Series Anomaly Detection

Researchers propose PAI, a novel anomaly scoring scheme that addresses a critical limitation in representation-based time-series anomaly detection by explicitly preserving amplitude information in learned embeddings. The method achieves significant performance improvements, with average gains of 98.4% on TSB-AD-U-Eva and 36.8% on TAB UV datasets, suggesting that amplitude retention is crucial for robust anomaly detection.

Time-series anomaly detection has become increasingly important across industries, from financial monitoring to infrastructure management. Representation-based approaches have dominated recent research due to their superior performance on diverse tasks, yet they systematically discard a critical feature: amplitude information. This oversight means existing methods fail to detect anomalies where the key indicator is a change in magnitude rather than pattern—a significant blind spot for real-world applications.

The PAI scheme introduces an elegant two-stage solution. First, a diagnostic module tests whether amplitude data already exists in learned representations by comparing cosine similarity (which ignores magnitude) against Euclidean distance (which preserves it). If amplitude information is lost, the second module augments anomaly scores using point-wise median absolute deviation and local mean-shift detection. This approach respects the original method while retrofitting missing signal.

The performance gains are substantial and consistent. Achieving 98.4% improvement on one benchmark represents meaningful progress in a maturing field. The fact that gains hold across multiple underlying representation methods—not just one technique—suggests PAI addresses a fundamental architectural issue rather than exploiting a specific weakness.

For practitioners deploying anomaly detection systems, this work validates the importance of verifying what information their models actually retain. Organizations using representation-based approaches should evaluate whether amplitude-related anomalies represent meaningful false negatives in their use cases. The availability of open-source code enables rapid adoption, making this a practical contribution beyond academic interest.

• →Representation-based time-series anomaly detection methods systematically lose amplitude information, creating blind spots for magnitude-related anomalies
• →PAI's diagnostic module identifies whether amplitude is missing in learned embeddings by comparing cosine vs. Euclidean distance metrics
• →The proposed method achieves 98.4% average improvement on TSB-AD-U-Eva dataset and works consistently across all evaluated representation methods
• →Amplitude preservation is an underemphasized requirement in existing anomaly detection scoring schemes despite its practical importance
• →Open-source availability enables practical adoption across deployed anomaly detection systems in finance, infrastructure, and monitoring applications