CATCH: Channel-Aware multivariate Time Series Anomaly Detection via Frequency Patching

Researchers introduce CATCH, a novel framework for detecting anomalies in multivariate time series data using frequency patching and channel-aware mechanisms. The method achieves state-of-the-art performance across 22 datasets by improving detection of fine-grained frequency patterns while identifying relevant channel correlations through a Channel Fusion Module.

CATCH addresses a fundamental challenge in anomaly detection for complex time series systems where multiple data streams exhibit correlated behaviors. Traditional reconstruction-based methods struggle to capture both detailed frequency characteristics and meaningful relationships between data channels simultaneously. The framework's innovation lies in its frequency patching approach, which discretizes frequency domains into interpretable bands rather than treating them as continuous spaces, enabling more precise detection of subtle abnormalities.

The Channel Fusion Module represents a significant technical advancement, employing patch-wise masking and attention mechanisms to dynamically determine which data channels are relevant for each frequency band. This selective focus prevents noise from irrelevant channels from degrading detection accuracy. The bi-level optimization algorithm drives iterative discovery of optimal channel correlations, effectively clustering similar channels while suppressing interference from unrelated ones.

The broader context reveals growing demand for sophisticated anomaly detection in industrial systems, financial markets, and network infrastructure where failures cascade through interconnected components. Current approaches either oversimplify channel relationships or fail to capture frequency-domain nuances, creating blind spots for complex, multi-modal anomalies.

For practitioners, CATCH's demonstrated superiority across real-world and synthetic datasets suggests immediate applicability to fault detection in manufacturing, fraud identification in financial systems, and security monitoring in telecommunications. The open-source release accelerates adoption and enables customization for domain-specific requirements. The framework's channel-aware design particularly benefits systems with heterogeneous sensors or data sources where selective fusion improves both accuracy and computational efficiency.

• →CATCH introduces frequency patching to capture fine-grained frequency characteristics in multivariate time series anomaly detection
• →Channel Fusion Module dynamically identifies relevant channel correlations while isolating adverse effects from irrelevant data streams
• →Framework achieves state-of-the-art performance across 22 real-world and synthetic datasets through bi-level optimization
• →Open-source code and datasets enable rapid adoption and customization for domain-specific anomaly detection applications
• →Channel-aware selective fusion improves both detection accuracy and computational efficiency in complex monitoring systems