DeepAnT: A Deep Learning Approach for Unsupervised Anomaly Detection in Time Series

This repository contains an implementation of the paper "DeepAnT: A Deep Learning Approach for Unsupervised Anomaly Detection in Time Series." The original paper can be found here.

About the Paper

The paper, authored by Mohsin Munir, Shoaib Ahmed Siddiqui, Andreas Dengel, and Sheraz Ahmed, presents DeepAnT, a novel deep learning model designed for unsupervised anomaly detection in time series data.

• Model Structure: The model consists of a sequence of convolutional layers followed by fully connected layers to capture local patterns and long-term dependencies. Below is the model structure from the paper:
• Unsupervised Learning: Suitable for applications with scarce labeled anomalies.
• Anomaly Detection: Detects anomalies by comparing predicted values with actual values and calculating anomaly scores.

Implementation Details

This implementation follows the architecture and methodology described in the paper using PyTorch and PyTorch Lightning.

• Sliding Window: Preprocesses the time series data using a sliding window approach, converting raw sequences into overlapping windows (input) and the subsequent step (target).

• Forecasting-Based Model: Predicts the next value in the sequence. Anomalies are inferred by large deviations of the predicted value from the actual value.

• Training and Validation:

  • Initial training without considering the validation loss.
  • After the initial training, the best model is selected based on the validation loss, using early stopping to prevent overfitting.

• Dynamic Threshold Calculation: Thresholds are computed per feature dimension based on the anomaly score distribution (mean ± N standard deviations).

• Visualization: Provides visualizations for predicted sequences as well as detected anomalies.

Multivariate Support

This implementation now supports both univariate and multivariate time series datasets. When feature_dim > 1, it automatically:

• Calculates per-feature thresholds.
• Identifies anomalies in each feature separately.
• Creates subplots for each feature to show anomalies per feature dimension.

Results

Univariate NAB Dataset

Below is an example run on the NAB (univariate) dataset:

• Final Training Loss: 0.0022
• Final Validation Loss: 0.007
• Dynamic Threshold: 0.031
• Detected Anomalies: [54, 55, 84, 134, 135, 139, 142, 144]

Multivariate Air Quality Dataset

The model also was tested on the Air Quality dataset from the UCI Repository:

• Number of Features: 15 (Date, Time, CO, PT08.S1, PT08.S2, etc.)

• Window Size: 24 (one day of hourly data)

• Final Training Loss: 0.043

• Final Validation Loss: 0.066

• Per-Feature Thresholds: [0.191, 0.003, 0.175, ...]

• Detected Anomalies (per feature):

  {
    "Feature_1": [16, 17, 54, 55, 56, 147, ...],
    "Feature_2": [427, 595, 643, 667, 673, ...],
    ...
  }
  

  An example visualization of one of its features:

Usage

1. Clone the repository:

   git clone https://github.com/EnsiyeTahaei/DeepAnT.git
   cd DeepAnT

2. Install the required packages:

   pip install -r requirements.txt

3. Configure Dataset:

   • Edit config.yaml to set dataset_name, window_size, and other hyperparameters. Currently, two datasets are configured:
     • NAB (for univariate)
     • Air Quality (for multivariate)

4. Run the main script:

   python main.py --dataset_name <dataset_name>

Note: dataset_name is optional. If not provided, it defaults to "Air Quality" (as it is specified in config.yaml).

License

This project is licensed under the MIT License.

Citation

If you use this code for your research, please cite the original paper:

@ARTICLE{8581424, author={Munir, Mohsin and Siddiqui, Shoaib Ahmed and Dengel, Andreas and Ahmed, Sheraz}, journal={IEEE Access}, title={DeepAnT: A Deep Learning Approach for Unsupervised Anomaly Detection in Time Series}, year={2019}, volume={7}, number={}, pages={1991-2005}, keywords={Anomaly detection;Time series analysis;Clustering algorithms;Data models;Benchmark testing;Heuristic algorithms;Anomaly detection;artificial intelligence;convolutional neural network;deep neural networks;recurrent neural networks;time series analysis}, doi={10.1109/ACCESS.2018.2886457} }

Questions or Contributions

Feel free to open an issue if you want to contribute further improvements.