De-SaTE: Denoising Self-attention Transformer Encoders for Li-ion Battery Health Prognostics

This repository contains the code for De-SaTE: Denoising Self-attention Transformer Encoders for Li-ion Battery Health Prognostics

Citation

Bibtex

@inproceedings{desate,
  author={Shinde, Gaurav and Mohapatra, Rohan and Krishan, Pooja and Sengupta, Saptarshi},
  booktitle={2023 IEEE International Conference on Big Data (BigData)}, 
  title={De-SaTE: Denoising Self-attention Transformer Encoders for Li-ion Battery Health Prognostics}, 
  year={2023},
  volume={},
  number={},
  pages={2221-2228},
  doi={10.1109/BigData59044.2023.10386134}}

In text

G. Shinde, R. Mohapatra, P. Krishan and S. Sengupta, "De-SaTE: Denoising Self-attention Transformer Encoders for Li-ion Battery Health Prognostics," 2023 IEEE International Conference on Big Data (BigData), Sorrento, Italy, 2023, pp. 2221-2228.

De-SaTE Architecture

Steps to run

1. Change the mask function to incorporate various types of noise

   • Gaussian noise: corrupted_x = x + self.noise_level * torch.randn_like(x)

   • Poisson noise: rate = torch.abs(x) / (self.noise_level + 1e-6)

     poisson_noise = torch.poisson(rate) * (self.noise_level + 1e-6)

     corrupted_x = x + poisson_noise

   • Speckle noise:

     corrupted_x = x * (1 + self.noise_level * torch.randn_like(x))

   • Uniform noise:

     corrupted_x = x + self.noise_level * (torch.rand_like(x) - 0.5)

   Change the wavelet_denoise function according to soft , hard and garrote mode: Wavelet decomposition Threshold for denoising Reconstruct the denoised signal

2. Run the training loop for ## number of epochs

Experiments

NASA data

CALCE data

Feel free to reach out to the authors with any questions about the experiments.