finloop / usad-torchlightning

Implementation of USAD (UnSupervised Anomaly Detection on multivariate time series) in PyTorch Lightning

usad-torchlightning

Implementation of USAD (UnSupervised Anomaly Detection on multivariate time series) in PyTorch Lightning.

Original implementation by: Francesco Galati. Original code can be found in: USAD.

Getting started

To start, first download the data.

Data

Data can be found in:

• Normal data: SWaT Dataset Normal
• Attack data: SWaT Dataset Attack

After downloading them put them in data/raw.

Running the model

dvc exp run

Changing the parameters

All the parameters (for example epoch size) can be found in params.yaml.

Requirements

• pytorch 1.9
• dvc
• pytorch-lighting
• python 3.8

How to cite

If you use this software, please cite the following paper as appropriate:

Audibert, J., Michiardi, P., Guyard, F., Marti, S., Zuluaga, M. A. (2020).
USAD : UnSupervised Anomaly Detection on multivariate time series.
Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, August 23-27, 2020

usad-torchlightning's Issues

Threshold Definition

Hi @finloop I am just wondering how did you define the threshold? I am looking at the paper but they didn't explain how did they set it up? Because you define it as "thresholds = np.arange(0.0, np.max(y_pred), np.max(y_pred)/50)", and in the loop when it searches for the best f1 score index, if it gives 0, then all the test dataset will be labelled as an anomaly.