RLNAS for Autoencoder based Anomaly Detection in Time Series Data

This repository contains the official implementation of our research paper "Neural Architecture Search for Anomaly Detection in Time Series Data of Smart Buildings: A Reinforcement Learning Approach for Optimal Autoencoder Design"

Setup

python -m venv venv
source venv/Scripts/activate
pip install -r requirements.txt

Running

1. Generate the dataset:

   Running prepare_data.py will generate contextual and point anomalies (using the Gaussian Mixture Model and Multivariate Uniform Distribution methods, as described in the paper) and inject them into clean data.

   Path and contamination parameters need to be set first.

   Generated train and test data are saved to dataset/.

2. Set the Neural Architecture Search parameters:

   • config/autoencoder_params.py contains training parameters of the Autoencoder architecture. (not included in the search)
   • config/controller_params.py contains the parameters of the Controller network.
   • config/search_space.py defines the search space, specifying the size of the target Autoencoder, the possible parameters, and the number of episodes to run the NAS process.

3. Run the NAS framework: python main.py with the arguments --train_dataset_path, --test_dataset_path and --output_file:

   This will execute the RLNAS framework on the training data and subsequently evaluate the best discovered architecture on the test data. Performance metric and top architectures will be saved to the file specified by the --output_file argument.

Comments

• For collecting the weather dataset, we use code from the the-weather-scraper project.

Citation

If you find this repository useful for your research work, please consider citing it as follows:

@ARTICLE{10418166,
  author={Dissem, Maher and Amayri, Manar and Bouguila, Nizar},
  journal={IEEE Internet of Things Journal}, 
  title={Neural Architecture Search for Anomaly Detection in Time-Series Data of Smart Buildings: A Reinforcement Learning Approach for Optimal Autoencoder Design}, 
  year={2024},
  volume={11},
  number={10},
  pages={18059-18073},
  keywords={Time series analysis;Internet of Things;Anomaly detection;Intelligent sensors;Data models;Smart buildings;Optimization;Anomaly detection;autoencoder (AE);neural architecture search (NAS);reinforcement learning (RL);smart buildings;time series},
  doi={10.1109/JIOT.2024.3360882}
}