A Contrastive Learner for Automatic Modulation Classification

Abstract

This is the official implementation for the paper: "A Contrastive Learner for Automatic Modulation Classification" (IEEE Trans. Wireless Commun., vol. 24, no. 4, 2025).

Data Preparation

• Segment by SNR: The RadioML 2018.01A dataset is split into individual files, each containing all modulation categories for a single SNR value.

• Generate Low-SNR Data: Synthetic low-SNR (5 dB) samples are created by adding random Gaussian noise to high-SNR (30 dB) data.

• Form Training Pairs: For contrastive learning, pairs can be created either from synthetically generated data or by directly pairing samples from existing SNR values (e.g., 30 dB and 4 dB).

Run

• Run run.py to generate a pretrained model through self-supervised learning on both noisy and clean data.
• Then, execute finetune.py to perform downstream classification, which learns a non-linear mapping from semantic features to category labels.

Citation

  @ARTICLE{10857965,
  author={Du, Mingyang and Pan, Jifei and Bi, Daping},
  journal={IEEE Transactions on Wireless Communications}, 
  title={A Contrastive Learner for Automatic Modulation Classification}, 
  year={2025},
  volume={24},
  number={4},
  pages={3575-3589},
  keywords={Noise measurement;Signal to noise ratio;Feature extraction;Modulation;Data models;Contrastive learning;Training;Time-frequency analysis;Robustness;Accuracy;Automatic modulation classification;contrastive learning;noise corruption;self-supervised},
  doi={10.1109/TWC.2025.3532438}}

Contact

If you have any question about our work or code, please email dumingyang17@nudt.edu.cn.