This repository supplements our paper "TranAD: Deep Transformer Networks for Anomaly Detection in Multivariate Time Series Data" accepted in VLDB 2022. This is a refactored version of the code used for results in the paper for ease of use. Follow the below steps to replicate each cell in the results table.
This code needs Python-3.7 or higher.
pip3 install -r requirements.txt
Creation a virtual environment:
python3 -m venv --prompt TranAD .venv
source activate-venv.sh
pip install --upgrade pip
Preprocess all datasets using the command
python3 preprocess.py SMAP MSL SWaT SMD UCR MBA NAB
Distribution rights to some datasets may not be available. Check the readme files in the ./data/
folder for more details. If you want to ignore a dataset, remove it from the above command to ensure that the preprocessing does not fail.
To run a model on a dataset, run the following command:
python3 main.py --model <model> --dataset <dataset> --retrain
where <model>
can be either of 'TranAD', 'GDN', 'MAD_GAN', 'MTAD_GAT', 'MSCRED', 'USAD', 'OmniAnomaly', 'LSTM_AD', and dataset can be one of 'SMAP', 'MSL', 'SWaT', 'WADI', 'SMD', 'MSDS', 'MBA', 'UCR' and 'NAB. To train with 20% data, use the following command
python3 main.py --model <model> --dataset <dataset> --retrain --less
For ablation studies, use the following models: 'TranAD_SelfConditioning', 'TranAD_Adversarial', 'TranAD_Transformer', 'TranAD_Basic'.
The output will provide anomaly detection and diagnosis scores and training time. For example:
$ python3 main.py --model TranAD --dataset SMAP --retrain
Using backend: pytorch
Creating new model: TranAD
Training TranAD on SMAP
Epoch 0, L1 = 0.09839354782306504
Epoch 1, L1 = 0.039524692888342115
Epoch 2, L1 = 0.022258711623482686
Epoch 3, L1 = 0.01833707226553135
Epoch 4, L1 = 0.016330517334598792
100%|███████████████████████████████████████████████████████████████████| 5/5 [00:03<00:00, 1.57it/s]
Training time: 3.1920 s
Testing TranAD on SMAP
{'FN': 0,
'FP': 182,
'Hit@100%': 1.0,
'Hit@150%': 1.0,
'NDCG@100%': 0.9999999999999999,
'NDCG@150%': 0.9999999999999999,
'TN': 7575,
'TP': 748,
'f1': 0.8915325929177795,
'precision': 0.8043010666204187,
'recall': 0.9999999866310163,
'threshold': 0.16133320075167037}
All outputs can be run multiple times to ensure statistical significance.
Our paper is available in the Proceedings of VLDB: http://vldb.org/pvldb/vol15/p1201-tuli.pdf. If you use this work, please cite using the following bibtex entry.
@article{tuli2022tranad,
title={{TranAD: Deep Transformer Networks for Anomaly Detection in Multivariate Time Series Data}},
author={Tuli, Shreshth and Casale, Giuliano and Jennings, Nicholas R},
journal={Proceedings of VLDB},
volume={15},
number={6},
pages={1201-1214},
year={2022}
}
BSD-3-Clause. Copyright (c) 2022, Shreshth Tuli. All rights reserved.
See License file for more details.