Source code for the paper "SigDA: A Superimposed Domain Adaptation Framework for Automatic Modulation Classification", which is accepted by IEEE Transactions on Wireless Communications.

Abstract

Due to the uncertainty of non-cooperative communication channels, the received signals often contain various impairment factors, leading to a significant decline in the performance of existing deep learning (DL)-based automatic modulation classification (AMC) models. Several preliminary works utilize domain adaptation (DA) to alleviate this issue, however, they are constrained by singular domain difference factor, whereas in practice, these factors often manifest cumulatively. Therefore, this paper introduce a more realistic task named superimposed DA, where multiple domain difference factors are overlaid, reflecting the cumulative nature of them. We propose the SigDA as a solution framework, which adopts adversarial training to align the data distribution in different domains. Two technical modules, Multi-task based Masked Signal Feature Extractor (M2SFE) and Signal Feature Pyramid Aggregation (SFPA), are innovatively designed in SigDA. M2SFE utilizes mask and reconstruction task to enhance feature extraction and achieves discriminative feature selection through the design of feature mapping layers, while SFPA can solve the problem of inconsistent signal length in superimposed DA and can aggregate the features of signals into the same dimension. We consider and superimpose various typical signal domain difference factors, comprehensive experiments demonstrate that the proposed framework can achieve significant performance improvement in various communication channels.

Dataset

For publicly available datasets, we use RML2016.10a and RML2016.04c,you can download the dataset here.

For the generated typical dataset, we have considered AWGN, Rician, and Rayleigh channels, and will open source the dataset to the current repository in the future.

Please download and unzip the dataset and place it in the .Datasets folder. After successful extraction, the .Datasets folder should contain the following files: 2016.04C.multisnr.pkl,RML2016.10a_dict.pkl.

Both of them contain data of 11 modulation types (BPSK, QPSK, 8PSK, 16QAM, 64QAM, BFSK, CPFSK, PAM4, WB-FM, AM-SSB, AM-DSB) under 20 different SNRs range from -20dB to 18dB. Each sample contains a complex number of 128 points. As a more standard dataset, for each SNR in each category of modulated signals, RML2016.10a contains 1000 samples, which means that it contains 220000 samples in total. The number of samples in each category of RML2016.04c is not consistent, it contains 162060 samples in total.

Environment Requirements

These models are implemented in Pytorch, and the environment setting is:

Python 3.7
pytorch 1.7.0

Script Description

Script and Sample Code


├── checkpoints
│   ├── a2c_backbone.pth
│   ├── a2c_discriminator.pth
│   └── m2sfe_10A.pth
├── comparative_experiment
│   ├── checkpoints                        #comparative experiment checkpoints
│   ├── cldnn                              #code for CLDNN model 
│   ├── cnn2                               #code for CNN2 model 
│   ├── daelstm                            #code for DAELSTM model 
│   ├── lstm                               #code for LSTM model 
│   ├── mcldnn                             #code for MCLDNN model 
│   ├── resnet                             #code for ResNet model 
│   └── vgg
├── Datasets                               #Dataset file 
│   ├── 2016.04C.multisnr.pkl              
│   └── RML2016.10a_dict.pkl
├── model
│   ├── config.py                          # Porcess config file
│   ├── domainclassifer.py                 # model for domainclassifer
│   ├── __init__.py                   
│   └── m2sfe.py                           # model for M2SFE
├── scripts
│   ├── run_eval_da.sh                     # shell script for domain adaptation evalation
│   ├── run_eval.sh                        # shell script for M2SFE evalation
│   ├── run_train_da.sh                    # shell script for domain adaptation training
│   └── run_train.sh                       # shell script for M2SFE eval
├── utils
│   ├── da_loader.py                       # process dataset, create dataloader for domain adaptation
│   ├── dataloader.py                      # process dataset, create dataloader for supervised training
│   └── utils.py                           # tools file
├── default_config.yaml                    # parameter configuration file
├── domain_adaptation.py                   # training Script for domain adaptation
├── train_m2sfe.py                         # training Script for supervised training of M2SFE
├── eval_domain_adaptation.py              # evalation Script for domain adaptation
├── eval_m2sfe.py                          # evalation Script for supervised training of M2SFE
└── README.md

Script Parameters

Parameters for both training and evaluation can be set in config.py

Config for SigDA

data_root: "Datasets"
data_name_10A: "RML2016.10a_dict.pkl"
data_name_04C: "2016.04C.multisnr.pkl"
model_root: "checkpoints"                        # weight file
checkpoints_10A: "m2sfe_10A.pth"                 # weight file for supervised training of RML2016.10A dataset
checkpoints_04C: "m2sfe_04C.pth"                 # weight file for supervised training of RML2016.04C dataset
checkpoints_DA: "a2c_backbone.pth"               # weight file for domain adaptation backbone 
checkpoints_DA_disc: "a2c_discriminator.pth"     # weight file for domain adaptation discriminator 
show_confusion_metrix: False

Epoch_num: 300                                   # number of Epochs
lr: 0.001                                        # learning rate of supervised training
Batch_size: 64                                   # Training batch size
da_lr_g: 0.00001                                 # learning rate of generator in domain adaptation process
da_lr_d: 0.00001                                  # learning rate of discriminator domain adaptation porcess

For more configuration details, please refer the script config.py.

Training Process

Training

-train for M2SFE

export CUDA_VISIBLE_DEVICES=0
python train_m2sfe.py > train_m2sfe.log 2>&1 &

or
cd scripts
bash run_train.sh

-train for Domain Adaptation

export CUDA_VISIBLE_DEVICES=0
python domain_adaptation.py > domain_adaptation.log 2>&1 &

or
cd scripts
bash run_train_da.sh

Evaluation Process

Evaluation

Evaluation for M2SFE
```
python eval_m2sfe.py > eval_m2sfe.log 2>&1 &
或
bash run_eval.sh
```
The above shell command will run in the background. You can view the results through the file eval/eval_m2sfe.log.
Evaluation for Domain Adaptation
```
python eval_domain_adaptation.py > eval_domain_adaptation.log 2>&1 &
或
bash run_eval_da.sh
```
The above shell command will run in the background. You can view the results through the file eval/eval_domain_adaptation.log.

M2SFE Series

The current code is set for M2SFE large. If you want to train or test other models in the M2SFE series, such as M2SFE-tiny or M2SFE-medium, Please check the partial code of the reference model in train_m2sfe.py

   from model.m2sfe import M2SFE as M2SFE

and modify it as follows:

   from model.m2sfe import M2SFE_tiny as M2SFE    # M2SFE-tiny model
   from model.m2sfe import M2SFE_medium as M2SFE  # M2SFE-medium model

Citation

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

@ARTICLE{10557536,
  author={Wang, Shuang and Xing, Hantong and Wang, Chenxu and Zhou, Huaji and Hou, Biao and Jiao, Licheng},
  journal={IEEE Transactions on Wireless Communications}, 
  title={SigDA: A Superimposed Domain Adaptation Framework for Automatic Modulation Classification}, 
  year={2024},
  volume={},
  number={},
  pages={1-1},
  keywords={Feature extraction;Task analysis;Modulation;Adaptation models;Convolution;Data models;Wireless communication;Automatic Modulation Classification;Domain Adaptation;Multi-task learning;Adversarial training},
  doi={10.1109/TWC.2024.3399067}}

HantongXING / SigDA

readme

Contents