PaperlayerView exact revision on arXiv

Source-linked AI summary

Data Augmentation for Deep Learning-based Radio Modulation Classification

Liang Huang, Weijian Pan, You Zhang, LiPing Qian, Nan Gao, Yuan Wu

arXiv:1912.03026v2eess.SPcs.LG

TL;DR

Radio modulation classifiers need substantial training data, while the effects of different augmentation methods in this domain were not established. The paper evaluates rotation, flip, and Gaussian-noise augmentation with a deep-learning classifier, finding improved accuracy, especially for rotation and flip, plus gains from joint augmentation and shorter samples.

  • Problem

    The effects of data augmentation methods on radio modulation classification were not established, despite limited and costly collection of reliable training samples.

  • Method

    The study evaluates rotation, flip, and Gaussian-noise augmentation with a state-of-the-art modulation classifier across training and inference phases.

  • Results

    Rotation achieves the highest accuracy, flip also outperforms Gaussian noise, and joint rotation-and-flip augmentation with 12.5% of the initial training data exceeds the unaugmented 100% baseline.

  • Takeaways & Limitations

    Data augmentation supports accurate modulation classification with half the radio sampling points, enabling a simplified model and shorter classification response time.

Abstract

from arXiv · show

Deep learning has recently been applied to automatically classify the modulation categories of received radio signals without manual experience. However, training deep learning models requires massive volume of data. An insufficient training data will cause serious overfitting problem and degrade the classification accuracy. To cope with small dataset, data augmentation has been widely used in image processing to expand the dataset and improve the robustness of deep learning models. However, in wireless communication areas, the effect of different data augmentation methods on radio modulation classification has not been studied yet. In this paper, we evaluate different data augmentation methods via a state-of-the-art deep learning-based modulation classifier. Based on the characteristics of modulated signals, three augmentation methods are considered, i.e., rotation, flip, and Gaussian noise, which can be applied in both training phase and inference phase of the deep learning algorithm. Numerical results show that all three augmentation methods can improve the classification accuracy. Among which, the rotation augmentation method outperforms the flip method, both of which achieve higher classification accuracy than the Gaussian noise method. Given only 12.5% of training dataset, a joint rotation and flip augmentation policy can achieve even higher classification accuracy than the baseline with initial 100% training dataset without augmentation. Furthermore, with data augmentation, radio modulation categories can be successfully classified using shorter radio samples, leading to a simplified deep learning model and shorter the classification response time.

I. INTRODUCTION

Deep learning enables automatic radio modulation classification, but limited training data and unassessed augmentation choices motivate this study. The paper evaluates three signal-aware augmentation methods and reports accuracy gains, stronger performance from rotation and flip, and benefits from joint augmentation and shorter samples.

  • Deep learning automatically classifies received radio signals without prior knowledge, supporting applications including spectrum management and unauthorized signal detection.
  • Limited access to high-quality radio samples makes training-data scarcity a practical challenge for deep-learning modulation classifiers.Data augmentation expands datasets through label-preserving transformations.
  • The study evaluates rotation, flip, and Gaussian-noise augmentation for modulated signals using a state-of-the-art deep-learning modulation classifier.The methods are motivated by modulated-signal characteristics and evaluated on an open radio-signal dataset.
  • Around 2.5% improvement over the baseline in classification accuracy is achieved by augmenting radio signals during training and inference.
  • Rotation outperforms flip, while both achieve higher classification accuracy than Gaussian noise.
  • With only 12.5% of the training dataset, joint rotation-and-flip augmentation expands it to 75% and exceeds the unaugmented 100% baseline in classification accuracy.
  • Data augmentation enables successful classification from half the radio sampling points, simplifying the model and shortening classification response time.

A. Deep Learning in Radio Modulation Classification

Radio modulation classification has adopted deep-learning models and generic augmentation techniques, but augmentation remains sparsely studied for this task. Existing GAN-based augmentation has limitations, motivating evaluation of augmentation methods with an LSTM classifier.

  • CNNs such as GoogleNet and AlexNet, along with modified ResNet models, have been applied to radio modulation classification after transforming or directly processing radio signals.
  • Generic data augmentation increases training-data diversity, helps prevent overfitting, and improves model robustness across deep-learning tasks.
  • Few studies have examined data augmentation specifically for radio modulation classification.
  • GAN-based augmentation improves modulation classification accuracy by no more than 6%, requires sufficient samples for convergence, and can underperform equally sized real datasets.
  • The paper evaluates augmentation methods using a state-of-the-art LSTM model and a radio-signal dataset.

A. Radio Signal Dataset

The paper uses RadioML2016.10a and an LSTM-based classifier that transforms consecutive I/Q signals into amplitudes and phases before predicting one of 11 modulation categories.

  • A. Radio Signal Dataset: RadioML2016.10a contains 220,000 radio signal segments spanning 11 modulation categories and channel-related impairments.The dataset includes sample-rate offset, center-frequency offset, multipath fading, and additive white Gaussian noise.
  • A. Radio Signal Dataset: LSTM processes consecutive modulated in-phase and quadrature signals as inputs and maps them to a modulation category.LSTM is a recurrent architecture used for time-series data and selectively retains or discards input information through its gates.
  • A. Radio Signal Dataset: The classifier converts I/Q signals into amplitudes and phases, then feeds them into a two-layer LSTM with 128 cells per layer.A fully connected Softmax layer maps the extracted features to one of the 11 categories.
  • A. Radio Signal Dataset: Adam with a dynamic learning rate minimizes cross-entropy loss using class labels and predicted class probabilities.The loss is defined over K classes, with y_k as the ground-truth label and ŷ_k as the predicted probability for class k.

IV. DATA AUGMENTATION METHODS

The paper expands modulation-signal datasets through rotation, flip, and Gaussian-noise augmentation, with rotation implemented by rotating I/Q samples around the origin.

  • IV. DATA AUGMENTATION METHODS: Data augmentation expands the dataset by a scale factor N while improving generalization and alleviating overfitting.The methods considered for modulation-signal recognition are rotation, flip, and Gaussian noise.
  • IV. DATA AUGMENTATION METHODS: Rotation maps an original modulated signal (I, Q) to an augmented sample (I′, Q′) by rotating it around the origin.The rotation is represented by a two-dimensional transformation parameterized by angle θ.
  • IV. DATA AUGMENTATION METHODS: The paper uses counter-clockwise rotation angles of 0, π/2, π, and 3π/2.For a QPSK sample, one raw-data set is augmented into four radio-signal samples.

B. Flip

The flip method changes the signs of I or Q in a modulated signal, while the referenced figures depict signal constellations and the LSTM architecture.

  • B. Flip: Horizontal flipping switches the I value of a given modulated radio signal to its opposite.The operation is defined for an input signal represented by (I, Q).
  • B. Flip: Vertical flipping switches the Q value to its opposite.Together, the horizontal and vertical definitions specify sign changes along the two signal components.
  • B. Flip: Figure 2 presents constellation diagrams for 11 modulated signals under different SNRs.The caption identifies the modulation-signal constellations and their varying SNR conditions.
  • B. Flip: Figure 3 presents the architecture of the LSTM network used by the classifier.The supplied caption identifies the network architecture but does not specify its internal layers.

C. Gaussian Noise

Gaussian noise augmentation adds noise to the in-phase and quadrature components of radio samples. The method expands the dataset but is not preferred for radio data augmentation.

  • C. Gaussian Noise: Gaussian noise augmentation adds N(0, σ2) to (I, Q), producing an augmented sample (I′, Q′).σ2 denotes the noise variance.
  • C. Gaussian Noise: Each augmentation method expands the original radio signal dataset by a default scale factor N = 4.The Gaussian-noise variants use different standard deviations σ.

A. Train-time augmentation

Train-time augmentation expands the training dataset while leaving the test dataset unchanged. With rotation, increasing the training set from 110,000 to 440,000 samples is associated with higher classification accuracy.

  • A. Train-time augmentation: Train-time augmentation expands the training dataset by scale factor N while leaving the test dataset unchanged.
  • A. Train-time augmentation: Rotation augmentation expands the training dataset from 110,000 to 440,000 radio signal samples.
  • A. Train-time augmentation: Larger training datasets generally yield higher modulation classification accuracy.

C. Train-test-time augmentation

Train-test-time augmentation applies augmentation during both training and inference. Across full and reduced datasets, augmentation improves classification accuracy, with rotation and flip generally outperforming Gaussian noise and their joint policy performing strongly with limited data.

  • C. Train-test-time augmentation: Train-test-time augmentation expands both training and test datasets by factor N.
  • C. Train-test-time augmentation: Train-test-time augmentation achieves the highest accuracy among the evaluated augmentation phases, improving over baseline by 8.87% at SNR -6dB and about 2.2% when SNR exceeds 4 dB.
  • C. Train-test-time augmentation: Rotation augmentation improves accuracy by 8% at SNR -6dB to -2dB and about 2% at SNR ≥4dB, while Gaussian noise performs better only at lower SNR.
  • C. Train-test-time augmentation: With 12.5% training data, joint rotation and flip augmentation expands the dataset to 75% and achieves similar accuracy to the unaugmented 100% baseline.The joint policy uses scale factor N = 6.
  • C. Train-test-time augmentation: With insufficient training data, joint augmentation improves classification of BPSK, WBFM, QAM16, and QAM64, outperforming individual rotation and flip for QAM16–QAM64 confusion.
  • C. Train-test-time augmentation: Adding Gaussian noise to joint rotation and flip slightly reduces classification accuracy.Rotation and flip are therefore preferred for radio data augmentation.

C. Augmentations On short Sample

The study also evaluates augmentation with shorter radio samples by splitting each 128-point sample into two 64-point samples.

  • C. Augmentations On short Sample: Each original 128-point radio signal sample is halved into two 64-point samples, producing 440,000 entries.

VII. CONCLUSION

The study evaluates rotation, flip, and Gaussian-noise augmentation for deep learning-based modulation classification, including training and inference-time use. Rotation performs best individually, while joint rotation–flip augmentation is especially effective with limited data and shorter radio samples.

  • Train-test-time augmentation achieves the highest classification accuracy among the evaluated training and inference strategies.
  • Rotation and flip augmentation achieve higher classification accuracy than Gaussian noise, with rotation achieving the highest accuracy.
  • Given only 12.5% of the initial training dataset, joint rotation–flip augmentation expands the dataset to 75% and exceeds the unaugmented 100%-dataset baseline.
  • Data augmentation allows classification from only one half of the radio sampling points, producing a simplified model and shorter classification response time.
Loading 1912.03026v2…