This code is the official PyTorch implementation of our ICLR'25 paper: CATCH: Channel-Aware Multivariate Time Series Anomaly Detection via Frequency Patching.
If you find this project helpful, please don't forget to give it a ⭐ Star to show your support. Thank you!
🚩 News (2025.1) CATCH has been accepted by ICLR 2025.
CATCH, a framework based on frequency patching, flexibly utilizing the channel correlations to reconstruct all the frequency spectrums in a fine-grained way to achieve remarkable detection accuracy. Technically, we propose a Channel Fusion Module (CFM), which features a patch-wise mask generator and the masked-attention mechanism. Driven by a bi-level multi-objective optimization algorithm, the CFM is encouraged to iteratively discover appropriate patch-wise channel correlations and cluster similar channels in the hidden spaces while isolate the adverse effects from irrelevant channels, which provides both the capacity and robustness of the attention mechanism.
Given a python environment (note: this project is fully tested under python 3.8), install the dependencies with the following command:
pip install -r requirements.txt
Prepare Data. You can obtained the well pre-processed datasets from OneDrive. (This may take some time, please wait patiently.) Then place the downloaded data under the folder ./dataset.
- To see the model structure of CATCH, click here.
- We provide the experiment scripts for CATCH and other baselines under the folder
./scripts/multivariate_detection. For example you can reproduce a experiment result as the following:
sh ./scripts/multivariate_detection/detect_label/MSL_script/CATCH.sh
sh ./scripts/multivariate_detection/detect_score/MSL_script/CATCH.shExtensive experiments on 10 real-world datasets and 12 synthetic datasets demonstrate that CATCH achieves state-of-the-art performance. We show the main results of all the 10 real-world datasets, and report the mean results of the 6 types of synthetic datasets:
If you want to test all baseline models, follow these steps:
Some models require Git to run. Please ensure that Git is installed on your system. If not, you can install it using the following command:
sudo apt-get install gitDownload the environment from OneDrive. (This may take some time, please wait patiently.) Then, create a directory my_env (we recommend creating it under conda/envs/) and extract the environment into this directory:
mkdir -p my_env
tar -xzf CATCH.tar.gz -C my_envsource my_env/bin/activateYou can clear the environment prefix using the following command. Now the environment is the same as one created directly with conda at this path:
(my_env) $ conda-unpackNow you can test all the baseline models. We provide experiment scripts for various baseline models under the folder ./scripts/multivariate_detection:
(my_env) $ sh ./scripts/multivariate_detection/detect_label/MSL_script/AnomalyTransformer.shIf you find this repo useful, please cite our paper.
@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}
}
If you have any questions or suggestions, feel free to contact:
Or describe it in Issues.