MC-CPC

Multi-Channel Lightweight Contrast Prediction Coding for Feature Extraction of Radar Emitter Signals

Overview

This project implemented Multi-Channel CPC for Multi-Channel Lightweight Contrast Prediction Coding for Feature Extraction of Radar Emitter Signals.

Specifically, to extract low pulse width radar signal features and meet the real-time requirements of signal identification, we design a lightweight encoder to realize the CPCfeature encoding function. Then, we construct a multi-channel CPC feature decoder to mine and extract subtle individual signal feature from the perspective of multi-domain and multi-channel information input. Simulation results verify the effectiveness of our proposed method, which can achieve state-of-the-art results in both accuracy and running time compared to the existing optimal methods.

Training datasets and test datasets

Relevant training data sets and Validation data sets see in https://pan.baidu.com/s/1phD1gqDxqQCYLHclP8DZ_g, and the data extraction code is "zjnc"

This repository was developed and tested in PyTorch 1.0.

How to run

• The installation dependency is very simple. The quick installation method is to install Anaconda3 + pytorch>1.0
• Relevant training data sets and Validation data sets see in https://pan.baidu.com/s/1phD1gqDxqQCYLHclP8DZ_g, and the data extraction code is "zjnc", When the dataset is downloaded, it is placed in ./datasets/
• Relevant training dataset lists and Validation dataset lists see in ./test/
• Modify signal dataset directory in ./run.sh and ./run_train_spk.sh
• Run ./run.sh for CPC feature training and Validation
• Run ./run_train_spk.sh for resnet training and Validation

Results


the encoder genc is a strided convolutional neural network, which has three convolutional layers with kernel sizes [8,6,4], strides [4,2,2], paddings [2,2,1] and 512 hidden units with ReLU activations.

An example image : The network loss function with epochs.


The Fig shows the architecture of CPC model based on multi-channel decoder.

*Comparison of CPC training effects with different CPC network structures

*Comparison of identification effects of Resnet with different CPC