Skip to content

[2024-1 DGU 종합설계 프로젝트 SF팀 & 2024 대한전자공학회 하계종합학술대회]

Notifications You must be signed in to change notification settings

CapstoneDesign-SF/TimesNet-Light

Repository files navigation

TimesNet-Light : 경량화된 TimesNet 이상 탐지 모델

TimesNet

TimesNet의 구조



TimesNet-Light

image
Bayesian-Optimization을 통해 적절한 컨볼루션 수를 training 동안 최적화하여 공정 모니터링 이상 탐지에 최적화.



How to

  • Bayesian-Optimization

    • Checkout timesnet_bayesian_optimization.ipynb.
    • Follow the instruction of the file
    • There's no need to change or modify.
  • Detect Anomalies

    • Checkout timesnet_tasks.py.
    • Required command line arguments
      • task: Use "train" for training TimesNet before detection or use "detect" for only detection. If you want to try simulating anomaly detection with test data, use "simulate".
      • data: Data for training TimesNet or detection should be within this directory.
      • model_name: Model name for saving trained model, or detecting with corresponding model.
      python timesnet_tasks.py --task detect --data PSM_simulation --model_name timesnet
      
    • The result is as follows.
      ---Start detecting anomalies---
      Anomaly occured at 2024-05-31 00:56:05
      Anomaly occured at 2024-05-31 00:56:05
      Anomaly occured at 2024-05-31 00:56:06
      Anomaly occured at 2024-05-31 00:56:06
      Anomaly occured at 2024-05-31 00:56:08
      Anomaly occured at 2024-05-31 00:56:08
      Anomaly occured at 2024-05-31 00:56:08
      Anomaly occured at 2024-05-31 00:56:09
      Anomaly occured at 2024-05-31 00:56:09
      Anomaly occured at 2024-05-31 00:56:09
      
      평가 지표 경량화 전 경량화 후
      모델 성능(AUC-score) 0.9976 0.9979
      모델 규모(MB) 18.79 0.1
      학습 시간(sec) 418.8087 15.0395
      추론 시간(sec) 260.2487 14.6195



Citation

If you found this code helpful, please consider citing:

Semin Kim and Soohyun Oh, Minje Park, Jiho Lee & Moongi Seock (2024). Efficient Time-Series Data Anomaly Detection
with a Lightweight TimesNet Model . 대한전자공학회 학술대회, 제주.



Reference

  • Wu, Haixu, et al. "Timesnet: Temporal 2d-variation modeling for general time series analysis." The eleventh international conference on learning representations. 2022.
  • Hongzuo Xu, Guansong Pang, Yijie Wang and Yongjun Wang, "Deep Isolation Forest for Anomaly Detection," in IEEE Transactions on Knowledge and Data Engineering, doi: 10.1109/TKDE.2023.3270293.
  • https://github.com/thuml/Time-Series-Library
  • https://github.com/xuhongzuo/DeepOD

About

[2024-1 DGU 종합설계 프로젝트 SF팀 & 2024 대한전자공학회 하계종합학술대회]

Topics

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published