Self-Supervised Time Series Representation Learning via Cross Reconstruction Transformer

The official implementation for our TNNLS paper Self-Supervised Time Series Representation Learning via Cross Reconstruction Transformer.

Overview of Cross Reconstruction Transformer

Getting Started

Installation

Git clone our repository, and install the required packages with the following command

git clone https://github.com/BobZwr/Cross-Reconstruction-Transformer.git
cd Cross-Reconstruction-Transformer
pip install -r requirements.txt

We use torch=1.13.0.

Processing Data (Optional)

We provide data_processing.py to generate phase and magnitude information based on the time-domain data. You can modify this file to adapt it to your own datasets.

Training and Evaluating

We provide the sample script for training and evaluating our CRT

# For Training:
python main.py --ssl True --sl True --load True --seq_len 256 --patch_len 8 --in_dim 9 --n_classes 6

# For Testing:
python main.py --ssl False --sl False --load False --seq_len 256 --patch_len 8 --in_dim 9 --n_classes 6

We also provide a subset of HAR dataset for training and testing.

If you found the codes and datasets are useful, please cite our paper

@article{zhang2023self,
  title={Self-Supervised Time Series Representation Learning via Cross Reconstruction Transformer},
  author={Zhang, Wenrui and Yang, Ling and Geng, Shijia and Hong, Shenda},
  journal={IEEE Transactions on Neural Networks and Learning Systems},
  year={2023},
  publisher={IEEE}
}