PaperlayerView exact revision on arXiv

Source-linked AI summary

Deep Neural Network Architectures for Modulation Classification

Xiaoyu Liu, Diyu Yang, Aly El Gamal

arXiv:1712.00443v3cs.LGstat.ML

TL;DR

The paper addresses increasingly challenging wireless modulation recognition, where traditional methods depend on prior signal and channel knowledge. It evaluates CNN, ResNet, DenseNet, and CLDNN architectures on a realistic 10-modulation dataset, with CLDNN achieving 88.5% high-SNR accuracy.

  • Problem

    Wireless modulation recognition is becoming more challenging, while traditional methods require prior signal and channel knowledge that can be inaccurate and need a separate control channel.

  • Method

    The study evaluates CNN, ResNet, DenseNet, and CLDNN architectures using the GNU Radio-generated RadioML2016.10b dataset.

  • Results

    88.5% high-SNR accuracy is achieved by CLDNN, the highest among all tested neural network architectures.

  • Takeaways & Limitations

    CLDNN delivers the best classification architecture in this evaluation, improving accuracy by approximately 13.5% over the original CNN model.

Abstract

from arXiv · show

In this work, we investigate the value of employing deep learning for the task of wireless signal modulation recognition. Recently in [1], a framework has been introduced by generating a dataset using GNU radio that mimics the imperfections in a real wireless channel, and uses 10 different modulation types. Further, a convolutional neural network (CNN) architecture was developed and shown to deliver performance that exceeds that of expert-based approaches. Here, we follow the framework of [1] and find deep neural network architectures that deliver higher accuracy than the state of the art. We tested the architecture of [1] and found it to achieve an accuracy of approximately 75% of correctly recognizing the modulation type. We first tune the CNN architecture of [1] and find a design with four convolutional layers and two dense layers that gives an accuracy of approximately 83.8% at high SNR. We then develop architectures based on the recently introduced ideas of Residual Networks (ResNet [2]) and Densely Connected Networks (DenseNet [3]) to achieve high SNR accuracies of approximately 83.5% and 86.6%, respectively. Finally, we introduce a Convolutional Long Short-term Deep Neural Network (CLDNN [4]) to achieve an accuracy of approximately 88.5% at high SNR.

I. INTRODUCTION

Wireless modulation recognition is increasingly challenging as communication systems use more modulation methods and parameters. The paper evaluates deep architectures to improve autonomous recognition under realistic channel imperfections.

  • Accurate modulation recognition is essential because receivers must demodulate signals correctly for smooth transmission.
  • Traditional methods require prior signal and channel knowledge, which can be inaccurate and require a separate control channel.
  • CNNs were previously applied to distinguish 10 modulation types and outperformed expert-based approaches while providing greater flexibility.
  • ResNet and DenseNet strengthen feature propagation through shortcut paths between layers to address gradient and depth-related problems.
  • The study develops ResNet, DenseNet, and CLDNN architectures on the RadioML2016.10b dataset, achieving roughly 13.5% higher high-SNR accuracy than the prior architecture.The dataset was generated with GNU Radio to capture channel imperfections present in real systems.

II. SIMULATION SETUP

The experiments use RadioML2016.10b, a GNU Radio-generated dataset representing 10 digital and analog modulation types as complex time-domain samples.

  • The dataset contains 10 modulation types: eight digital schemes and two analog schemes.The digital set includes BPSK, QPSK, 8PSK, QAM16, QAM64, BFSK, CPFSK, and PAM4; the analog set includes WBFM and AM-DSB.
  • Each example is a 128-sample complex time-domain vector generated in GNU Radio.
  • The 160,000 samples are divided into training and testing data using 128-sample rectangular windowing.
  • Neural-network inputs represent each complex sample with separate real and imaginary components in 2*128 vectors.

A. Evaluation Network

The evaluation compares CNN, ResNet, DenseNet, and CLDNN architectures for modulation recognition. Deeper CNN variants and shortcut or recurrent connections improve high-SNR accuracy, with CLDNN performing best.

  • CNN: 75% high-SNR accuracy is achieved by the two-convolutional-layer CNN baseline.
  • ResNet: 83.5% high-SNR accuracy is achieved by the four-convolutional-layer ResNet, while deeper combined architectures reduce recognition accuracy.The residual connection forwards an earlier layer's output two levels deeper to alleviate gradient vanishing.
  • DenseNet: 86.6% high-SNR accuracy is achieved by the four-convolutional-layer DenseNet with one densely connected block.Each layer receives inputs from preceding layers and passes feature maps to subsequent layers.

B. Training Complexity

Training cost varies substantially with architecture depth and connectivity. More complex models require longer epochs and substantially longer total training times.

  • 15 seconds per epoch is required by the two-layer CNN, compared with approximately 400 seconds for the four-layer CNN.
  • The three more complex models require approximately 1000 seconds per epoch.
  • Total training time is approximately 70 hours for DenseNet, 20 hours for ResNet, and 50 hours for CLDNN.
  • A higher dropout rate can slow training while reducing overfitting, and increasing patience can produce smaller validation loss for deeper networks.The dropout rate is set to 0.6; patience is set to 20 for the remaining models.

A. Convolutional Neural Network

The CNN experiments vary filter settings and network depth to improve modulation classification. A four-convolutional-layer design achieves the strongest reported CNN accuracy at high SNR.

  • Larger filters in earlier convolutional layers and smaller filters in deeper layers optimize high-SNR accuracy.
  • 83.8% high-SNR accuracy is achieved by the four-convolutional-layer CNN.This improves on the two-convolutional-layer model by 8.8%.
  • Deeper CNNs become harder to train because validation loss is less likely to converge.Five- and six-convolutional-layer models show early loss vibrations and poorer classification performance.

B. Residual Network

Residual-network combinations perform similarly to the pure CNN, with the best result at four convolutional layers. DenseNet reaches higher high-SNR accuracy at the same depth, while deeper designs degrade performance.

  • ResNet: 83.5% is the best accuracy for the ResNet combination with a four-convolutional-layer network.Recognition accuracy decreases when the combined architecture has more than four convolutional layers.
  • DenseNet: 86.6% is the best high-SNR accuracy achieved by the four-convolutional-layer DenseNet.The DenseNet design uses one densely connected block, with convolutional layers added until accuracy begins to descend.
  • DenseNet: Four convolutional layers produce the best high-SNR performance for the DenseNet architecture.This result is reported in Figure 7.

D. CLDNN

The CLDNN adds recurrent processing after convolutional feature extraction for modulation recognition. It achieves the highest tested accuracy at high SNR, while remaining errors primarily involve similar modulation types.

  • Architecture and performance: 88.5% is the CLDNN accuracy at high SNR, the highest among all tested neural network architectures.The CLDNN adds an LSTM unit after the convolutional part, using cyclic connections to extract temporal features.
  • Architecture and performance: CLDNN and DenseNet outperform the other candidate architectures, with best accuracies of 88.5% and 86.6%, respectively.Figure 8 compares classification performance across candidate architectures.
  • Error analysis: At SNR=18dB, the CLDNN confusion matrix shows WBFM misclassified as AM-DSB and QAM16 misclassified as QAM64.These are identified as the two main discrepancies besides the clean diagonal.
  • Error analysis: QAM16 and QAM64 are likely to be misclassified as each other because similarities in their constellation diagrams make differentiation vulnerable to small feature differences.Small portions of 8PSK are misclassified as QPSK, and small portions of WBFM as GFSK.
  • Error analysis: Further architecture optimization and possibly increasing depth may help capture subtle feature differences underlying residual misclassifications.This expectation is stated in connection with the observed 8PSK/QPSK and WBFM/GFSK errors.

IV. DISCUSSION

Shortcut-based architectures improve training behavior and feature propagation, while CLDNN achieves the best performance among the tested models, attributed to long-term memory for causal radio signals.

  • DenseNet outperforms ResNet by including more shortcut connections that strengthen feature propagation.
  • DenseNet and ResNet reach validation-loss plateaus earlier than CNN.
  • At equal depth, DenseNet and ResNet begin with lower validation errors and remain lower throughout training than plain CNN.
  • CLDNN delivers the best performance among all tested architectures, possibly because its long-term memory suits causal time-domain radio signals.

V. CONCLUSION

The paper evaluates deep neural network architectures for radio modulation recognition and identifies CLDNN as the best-performing architecture. CLDNN improves accuracy by approximately 13.5% over the original CNN model.

  • Conclusion: 13.5% is the approximate accuracy improvement of CLDNN over the original CNN model introduced in.The study compares convolutional, residual, densely connected, and recurrent-layer architectures for radio modulation recognition.
  • Conclusion: CLDNN delivers the best classification architecture among the tested deep neural networks.The architecture combines convolutional feature extraction with recurrent processing for the modulation-recognition task.
Loading 1712.00443v3…