Awesome

Adversarial Autoencoder for Unsupervised Time Series Anomaly Detection and Interpretation

This is the implementation for the DAEMON model architecture described in the paper: "Adversarial Autoencoder for Unsupervised Time Series Anomaly Detection and Interpretation".

DAEMON is a robust time series anomaly detection model which uses two discriminators to learn the normal pattern of multivariate time series and uses the reconstruction error to determine anomalies and interpret the root cause. DAEMON has the following advantages: 1) Unsupervised: It can detect anomalies without labels; 2) Robustness: It achieves high performance in different real scenarios. 3) Effectiveness: It outperforms baselines in both accuracy and efficiency, achieving an overall F1-score of 0.94, and very fast training (169 s per epoch) and inference (0.1 ms per entity); 4) Explainability: It pinpoints the dimensions caused the anomalies.

Requirements

• Python 3.6
• Numpy (1.91.1)
• xlrd (1.2.0) # For data preprocess
• PyTorch (1.0.0)
• Scikit-learn (0.19.1)

Dataset

• SMAP and MSL:

wget https://s3-us-west-2.amazonaws.com/telemanom/data.zip && unzip data.zip && rm data.zip

cd data && wget https://raw.githubusercontent.com/khundman/telemanom/master/labeled_anomalies.csv

• SMD:

https://github.com/NetManAIOps/OmniAnomaly

• SWaT:

http://itrust.sutd.edu.sg/research/dataset

Usage

• Preprocess the data

python data_preprocess.py

• Run the code

python main.py <dataset>

where <dataset> is one of SMAP, MSL, SMD, SWAT.