1 of 25

Data Augmentation for Deep Learning-Based

Radio Modulation Classification

INC Lab.

| 김범기

2 of 25

​

    • Introduction
    • Related work
    • Preliminaries
    • Data augmentation methods
    • Data augmentation time
    • Augmentation performance
    • Conclusion

CONTENTS

3 of 25

Introduction

| Deep Learning

​

  • Speech and audio processing
  • Natural language processing
  • Object detection
  • etc..

​

-> Also, apply on field of wireless communications

Deep Learning

4 of 25

Data Augmentation

Introduction

| Data Augmentation

​

  • Requires a large volume of training radio samples
  • But difficult to collect

​

-> Data augmentation is mandatory

​

220,000 dataset

1/8 dataset

5 of 25

Data Augmentation

Introduction

| Data Augmentation Methods

​

  • Image recognition

Random cropping

Rotation

Mirroring

​

  • Speech recognition

Pitch shifting

Time stretching

Random frequency filtering

6 of 25

Data augmentation in deep learning

Related Work

| Data augmentation for

image classification

​

  • Flip
  • Rotation
  • Cropping
  • Color jittering
  • Edge enhancement
  • Fancy PCA
  • Generative Adversarial Nets (GAN)

​

-> But, not many for radio modulation classification

​

7 of 25

Data Augmentation

Introduction

| Data augmentation

​

Radio modification is similar as image recognition

​

In this research

  • Rotation
  • Flip
  • Gaussian noise

​

8 of 25

Deep learning in radio modulation classification

Related Work

| CNN-based deep learning

​

  • GoogleNet
  • AlexNet
  • ResNet
  • CNN-based LSTM (CLDNN)

​

  • Reducing the dimensions of input signals for decrease training time

​

9 of 25

LSTM vs Other Models

Related Work

| LSTM vs Other Models

​

LSTM shows very high accuracy.

  • LSTM is used in this research

​

S. Ramjee, S. Ju, D. Yang, X. Liu, A. E. Gamal, and Y. C. Eldar,

‘‘Fast deep learning for automatic modulation classification,’’

10 of 25

Preliminaries

| Radio signal dataset

​

  • RadioML2016.10a
  • 220,000 modulated radio segments
  • 11 different modulated catagories
  • 20 different SNRs from −20dB to 18dB

​

BPSK, QPSK, 8PSK, CPFSK, GFSK, PAM4, QAM16, QAM64, AM-DSB, AM-SSB, WB-FM

Radio Signal Dataset

11 of 25

LSTM Network Architecture

Preliminaries

| LSTM Network Architecture

​

LSTM is a special category of Recurrent neural network (RNN)

  • Widely used to process time series data

​

I/Q signals are first converted into amplitudes and phases

​

12 of 25

LSTM Network Architecture

Preliminaries

| LSTM Network Architecture

​

  • In order to avoid overfitting, set dropout rate to be 0.5 at both two LSTM layers.
  • The number of training epoch is 80 and the mini-batch size is 128.
  • Learning rate is initially set as 0.001

​

​

​

​

Implemented based on PyTorch

​

13 of 25

Time

| Data Augmentation Time

​

  • Train-time Augmentation
  • Data augmentation during the training stage of the model.

​

  • Test-time Augmentation
  • Fuses features of all augmented radio signal samples in inference phase.

​

  • Train-Test-time Augmentation
  • Conducts both train-time augmentation and test-time augmentation.

​

Data Augmentation Time

14 of 25

Methods

| Accuracy under different augmentation times

​

Train-Test-Time method is most accurate.

  • The train-test-time augmentation improves the modulation classification accuracy by 8.87% when SNR is −6dB and by about 2.2% when SNR is greater than 4dB.

​

-> Train-Test-Time is used in this research

​

Data Augmentation Time

15 of 25

Methods

| Rotation

​

By rotating a modulated radio signal (I , Q) around its origin

Rotation

16 of 25

Methods

| Flip

​

By flipping a modulated radio signal (I , Q)

Flip

Horizontal

Vertical

17 of 25

Methods

| Gaussian noise

​

By adding a Gaussian noise N (0, σ 2 ) to the modulated radio signal (I , Q)

Gaussian Noise

18 of 25

Augmentation Performance

Performance

| Augmentation Performance

​

All augmentation methods improve the accuracy when the SNR is greater than −10dB

​

Rotation and flip data augmentation are more preferred for radio signals in modulation classification

​

19 of 25

Performance

| Confusion when SNR is −2dB

  • At low SNR, difficult to classify 8PSK and QPSK

​

Gaussian Noise

20 of 25

Performance

| Confusion when SNR is 18dB

  • At high SNR, the accuracy of the LSTM model is mainly limited by the confusion between AM-DSB and WBFM
  • In general, rotation and flip achieve better classification accuracy than Gaussian noise

​

Gaussian Noise

21 of 25

Augmentation on Partial Dataset

Performance

| Augmentation on Partial Dataset

​

Resulting a low modulation classification accuracy around 45%

  • After deploying different radio data augmentation methods, the classification accuracy is improved.

​

​

22 of 25

Performance

| Confusion with 12.5% training dataset when SNR is 18dB

  • When training dataset is insufficient, it is difficult to classify BPSK, WBFM, QAM16 and QAM64
  • Rotation and flip methods are preferred for radio data augmentation

Gaussian Noise

23 of 25

Augmentation on Short Sample

Performance

| Augmentation on Short Sample

​

Evaluate data augmentation methods for modulated radio signals with fewer sampling points

​

It can reduce number of parameters and inference complexity

​

  • 201.1K Params -> 54.1K Params
  • 2.8K FLOPS -> 1.4K FLOPS

​

​

24 of 25

Conclusion of Research

Conclusion

| We can studied radio data augmentation methods

  1. Train-test-time augmentation achieves the highest accuracy.
  2. A joint augmentation policy with both rotation and flip methods can further improve the classification accuracy.
  3. Given only 12.5% of initial training dataset, the joint augmentation method obtains even higher than the 100% training datasets without augmentation.
  4. After deploying data augmentation, resulting in a simplified deep learning model and a shorter classification response time.

25 of 25

Thank You

| 김범기