CATCH-RG: Channel-Aware Multivariate Time-Series Anomaly Detection via Patching with Reliability-Gated Channel Mask Generator
This work received an Excellence Award at KAMP 2025 (The 5th K-Artificial Intelligence Manufacturing Data Analysis Competition),
recognizing its robustness-oriented extension of CATCH : Channel-Aware multivariate Time Series Anomaly Detection via Frequency Patching (ICLR'25) for real-world anomaly detection scenarios involving imperfect and incomplete multivariate time-series data.
CATCH is a channel-aware multivariate time-series anomaly detection method that models cross-channel interactions through frequency-domain patching and channel-aware masking. It is evaluated under the TAB benchmarking protocol, which provides unified and rigorous evaluation for time-series anomaly detection. This repository provides a robust masking extension for CATCH, and proposed reliability-gated channel masking to better handle incomplete and noisy multivariate time-series data, while preserving the original CATCH architecture and interface.
In practical settings, multivariate time-series data often contain:
- missing or partially observed channels,
- noisy or corrupted sensor measurements,
- globally inactive segments due to sensor or communication failures.
The original CATCH framework does not explicitly model channel reliability, which may lead to unstable cross-channel interactions under such conditions.
We extend the channel masking mechanism of CATCH with a Reliability-Gated Mask, which dynamically controls cross-channel information flow based on data-driven channel reliability estimates.
The proposed mask:
- estimates channel reliability using simple, robust statistics,
- softly suppresses interactions involving unreliable channels,
- preserves self-channel connections,
- and collapses to an identity mask when all channels are globally inactive.
If you use this work or build upon it, please cite the original CATCH and TAB papers:
@inproceedings{wu2024catch,
title = {{CATCH}: Channel-Aware multivariate Time Series Anomaly Detection via Frequency Patching},
author = {Wu, Xingjian and Qiu, Xiangfei and Li, Zhengyu and Wang, Yihang and Hu, Jilin and Guo, Chenjuan and Xiong, Hui and Yang, Bin},
booktitle = {ICLR},
year = {2025}
}