finloop / usad-on-ucr-data

USAD model on UCR Time Series Anomaly Archive

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Testing USAD on UCR dataset

The aim of this repository is to test how USAD (UnSupervised Anomaly Detection on multivariate time series) behaves on UCR Time Series Anomaly Archive dataset.

Model

The model was created based on this paper: Audibert et al. USAD : UnSupervised Anomaly Detection on multivariate time series. 2020 and its implementation by Francesco Galati - Github .

Full implementation can be found in src/model.py.

Examples on how the model behaves on different dataset can be found in: finloop/usad-torchlightning.

Data

The data comes from UCR Time Series Anomaly Archive . Download link: https://www.cs.ucr.edu/~eamonn/time_series_data_2018/UCR_TimeSeriesAnomalyDatasets2021.zip

To run this put contents of ZIPFILE/UCR_TimeSeriesAnomalyDatasets2021/FilesAreInHere/UCR_Anomaly_FullData to datasets/data .

How to run

Install requirements:

  • pytorch
  • pytorch-lightning
  • jupyterlab

To run open and run: notebooks/UCR_Anomaly.ipynb.

Results

The model (for now) was tested on dataset 003. This is how it looks: Dataset 003 The model seems to perform poorly. It is unable to find the desired anomaly( the bigger the reconstruction error - the stronger the anomaly). The biggest reconstruction error is not near the red area. Dataset 003

About

USAD model on UCR Time Series Anomaly Archive


Languages

Language:Jupyter Notebook 98.7%Language:Python 1.3%