lidan456 / gan-ad Goto Github PK

We used generative adversarial networks (GANs) to do anomaly detection for time series data.

Python 100.00%

gan-ad's Introduction

-- Multivariate Anomaly Detection for Time Series Data with GANs --

#GAN-AD

This repository contains code for the paper, Anomaly Detection with Generative Adversarial Networks for Multivariate Time Series, by Dan Li, Dacheng Chen, Jonathan Goh, and See-Kiong Ng.

Overview

We used generative adversarial networks (GANs) to do anomaly detection for time series data. The GAN framework was RGAN that taken from the paper, _Real-valued (Medical) Time Series Generation with Recurrent Conditional GANs. Please refer to https://github.com/ratschlab/RGAN for the original code.

Quickstart

Python3
Sample generation

"""python RGAN.py --settings_file gp_gen"""

"""python RGAN.py --settings_file sine_gen"""

"""python RGAN.py --settings_file mnistfull_gen"""

"""python RGAN.py --settings_file swat_gen"""

(Please unpack the mnist_train.7z file in the data folder before generate mnist)

To train the model for anomaly detection:

"""python RGAN.py --settings_file swat_train"""
To do anomaly detection:

"""python AD.py --settings_file swat_test"""

Data

In this repository, we applied GAN-AD on the SWaT dataset, please refer to https://itrust.sutd.edu.sg/ and send request to iTrust is you want to try the data.

gan-ad's People

Contributors

Stargazers

Watchers

Forkers

jiweitian chongyang915 jayju wallart dragon-dane zhouhaocomeon1 zhouzhou12 learning-sources asis012 anormaly-detection andyshields dhrubaadhikary tartaruszen hovercode jules-samaran sbhadade kellsky debaraj-barua alwaysproblem ngoclesydney zhujiahao111 shghahremani fish-kong wenjingsu abollo zhouchena1 vpredictor muleina imygirl kbishou blanik00 jmj8038 ptee mhahn0106 henrys-lab theoryoe rocky147258 zahrafarhadi ivetasarfyova fancy1573 shawnpi233 smile-zbj gyyixr kobeontheway zhanglining123 benjamin-ky engrtahar-noureddine jaychaontpuacc

gan-ad's Issues

could not broadcast input array from shape (7200) into shape (72) for swat data.

The figures are confusing me

Dear @LiDan456
I am so interested in your work on GAN-AD. However, i have some trouble to understand the figures created by running "phthon RGAN.py --settings_file swat_train" (see the attached images) . What do these curves represent？ I try to find some key informations in the paper, but iam failed. Thanks for your generosity.

Best wishes
KB.

Cannot achieve results as reported in the paper

Dear @LiDan456 and all,

I have some problem with testing part and would like to ask for help.

After trying to run the model follows your instructions and code, I managed to train RGAN.py model with a SWAT data set (Pure normal physical data), which working great so far.

However, I have some problem with test in AD.py part. I am trying to test the model (GAN-AD with 5 PCA components) with the SWAT data set (Abnormal physical data) with your original code and default setting but I could not get the results close to reported performance in your paper. (The paper reported that GAN-AD with 5 PCA components can outperform other methods. The Accuracy is 94.80, Precision is 93.33, Recall is 63.64, F-1 is 0.75 and FPR is 0.46.) The result that I get is pretty poor and not as good as it should. (Please see the attached images)

I also try training and testing again more than 10 times, but the best result I could achieve is Accuracy around 60-70% (which is far from the reported result)

As mention above, I would like to ask you questions as follows:

Why I could not achieve the same result as reported in your paper? Do you have any suggestion about my mistake I might make in code or implementation processes or some setting might be missing?
Did you do the cross-validation when you test? I am trying to read your paper and understand your code, but I still could not find any cross-validation method that you might implement. Please let me know if I misunderstood.

I would appreciate any comments, helps or suggestions. Thank you very much for your answer in advance.