zqhang / MTGFLOW

Official implmentation of AAAI'23 paper 'Detecting Multivariate Time Series Anomalies with Zero Known Label'.

Detecting-Multivariate-Time-Series-Anomalies-with-Zero-Known-Label

This repository provides a PyTorch implementation of MTGFlow (Paper), which is the unsupervised anomaly dection and localization method. This repository is based on GANF.

Framework

Main results

Requirements

• python==3.8.5
• pytorch==1.7.1
• numpy==1.19.2
• torchvision==1.5
• scipy==1.6.1
• scikit-learn==0.24.1
• scikit-image==0.18.1
• matplotlib== 3.3.4
• pillow == 7.2.0

pip install -r requirements.txt

Data

We test our method for five public datasets, e.g., SWaT, WADI, PSM, MSL, and SMD.

SWaT WADI

mkdir Dataset
cd Dataset
mkdir input

Download the dataset in Data/input.

Train

• train for MITGFlow For example, training for WADI

sh runners/run_WADI.sh

• train for DeepSVDD, DeepSAD, DROCC, and ALOCC.

python3 train_other_model.py --name SWaT --model DeepSVDD

• train for USAD and DAGMM We report the results by the implementations in the following links:

USAD and DAGMM

Test

We provide the pretained model of MTGFlow.

For example, testing for WADI

sh runners/run_WADI_test.sh

BibTex Citation

If you find this paper and repository useful, please cite our paper.

@inproceedings{zhou2023detecting,
  title={Detecting Multivariate Time Series Anomalies with Zero Known Label},
  author={Zhou, Qihang and Chen, Jiming and Liu, Haoyu and He, Shibo and Meng, Wenchao},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  volume={37},
  number={4},
  pages={4963--4971},
  year={2023}
}