TFMix: A Robust Time-Frequency Mixing Approach for Domain Generalization in Specific Emitter Identification

This repository contains the official implementation of the paper "TFMix: A Robust Time-Frequency Mixing Approach for Domain Generalization in Specific Emitter Identification".

Introduction

Specific Emitter Identification (SEI) faces significant challenges when dealing with signal variations caused by changing environmental conditions and time-varying hardware characteristics. TFMix is a novel domain generalization method designed to address these challenges. It leverages a time-frequency mixing strategy to generate diverse training samples, thereby improving the robustness of SEI models against domain shifts.

Features

• Time-Frequency Mixing: A novel data augmentation technique that mixes signals in both time and frequency domains to simulate realistic channel variations.
• Domain Generalization: Designed to train models that generalize well to unseen domains (e.g., data collected on different days or under different conditions).
• Complex-Valued Neural Networks: Utilizes complex-valued CNNs to effectively process IQ signal data.

Project Structure

• TFMix.py: The main entry point for training and evaluation. It handles the training loop, validation, and testing across different domains.
• fe.py: Defines the Feature Extractor (FE) module.
• cls.py: Defines the Classifier (FC) module.
• complexcnn.py: Implements complex-valued convolutional layers and operations.
• get_ManySig_CR_unequal.py: Data loader script responsible for loading and preprocessing signal data from different domains.

Requirements

The code is implemented in Python using PyTorch. The main dependencies are:

• Python 3.x
• PyTorch
• NumPy
• TensorBoard

You can install the necessary packages using pip:

pip install torch numpy tensorboard

Usage

To train and evaluate the model, simply run the TFMix.py script:

python TFMix.py

The script will:

1. Iterate through different dates (domains) defined in the code (1-1, 1-19, 14-7, 18-2).
2. For each iteration, it uses one date as the target (test) domain and the others as source (training) domains.
3. Train the model using the TFMix strategy.
4. Evaluate the model on the held-out test domain.
5. Log training progress and results to TensorBoard and text files.

Results

Training logs and checkpoints are saved in the logs/ and model_weight/ directories, respectively. Final accuracy results are appended to result/TFMix_CR_Acc.txt.

License

This project is distributed under a Custom Non-Commercial License. See the LICENSE file for more details. Any form of commercial use is prohibited.

Citation

If you find this work useful in your research, please consider citing our paper.