DA Outlier Detection

"Unsupervised Anomaly Detection in Energy Time Series Data Using Variational Recurrent Autoencoders with Attention" 논문 기반 이상치 탐지 프로젝트.

Authors

• YBIGTA 22기 홍세아
• YBIGTA 23기 정회수
• YBIGTA 23기 김소민
• YBIGTA 24기 김종진
• YBIGTA 25기 문찬우
• YBIGTA 25기 한예지

Environment

Create conda virtual environment

  conda create -n outlier python=3.11
  conda activate outlier

Github pull request

  git init
  git remote add origin https://github.com/YBIGTA/25th-da-outlier-detection.git
  git branch -m main
  git pull origin main

Install requirements

  pip install -r requirements.txt

File Structure

WORKING DIRECTORY
├── data                    # Outlier dataset (confidential)
├── output                  # Model checkpoints & outlier visualizations
├── utils
│   ├── config.yaml         # Configurations
│   ├── model.py            # Model initialization
│   ├── dataset.py          # Dataset initialization
│   └── main.py             # Main method
└── requirements.txt

Data Pipeline

Configurations

# dataset.py
data_path: "PATH_TO_CSV"
interval_path: "PATH_TO_JSON"
step_size: 50
split_ratio: 0.8

# model.py
lstm_size: 128
latent_size: 20
input_size: 1
seq_size: 150
num_lyears: 1
batch_size: 16
attention_size: 2
sample_reps: 20
directions: 2

# main.py
train: True
recon_prob_threshold: 0.20
optimizer_choice: 'AdamW'
learning_rate: 0.02
epochs: 10
lambda_kl: 0
eta: 0.01

Deployment

  python main.py -c 'CONFIG_PATH'

Acknowledgements

• "Unsupervised Anomaly Detection in Energy Time Series Data Using Variational Recurrent Autoencoders with Attention"
• https://github.com/LauJohansson/AnomalyDetection_VAE_LSTM.git