PaperlayerView exact revision on arXiv

Source-linked AI summary

SigNet: A Novel Deep Learning Framework for Radio Signal Classification

Zhuangzhi Chen, Hui Cui, Jingyang Xiang, Kunfeng Qiu, Liang Huang, Shilian Zheng, Shichuan Chen, Qi Xuan, Xiaoniu Yang

arXiv:2011.03525v2eess.SPcs.LG

TL;DR

Radio signal classification methods can depend on expert-designed features, while deep models also face architectural and training challenges. SigNet converts signals into matrices with trainable S2M operators for CNN classification, and SigNet2.0 adds 1D convolutions. The authors report better baseline performance on two datasets and relatively high accuracy with very small training sets.

  • Problem

    Signal-classification methods may require extensive expertise and lack precision, while existing CNN structures may not adequately handle complex scenes and LSTMs can be time-consuming and difficult to train on long signals.

    Source: §I · p. 2

  • Method

    SigNet uses trainable S2M operators to convert signals into matrices for CNN classification, and SigNet2.0 integrates 1D and 2D convolution operators.

    Source: §I · p. 2

  • Results

    Across two signal datasets, the authors report that SigNet and SigNet2.0 outperform other models and retain relatively high accuracy with small training sets.

    Source: §I · p. 2, F. Learning with small samples · p. 9

  • Takeaways & Limitations

    The authors suggest the models could be useful when labeled signal data are difficult to obtain.

    Source: §V · p. 12

Abstract

from arXiv · show

Deep learning methods achieve great success in many areas due to their powerful feature extraction capabilities and end-to-end training mechanism, and recently they are also introduced for radio signal modulation classification. In this paper, we propose a novel deep learning framework called SigNet, where a signal-to-matrix (S2M) operator is adopted to convert the original signal into a square matrix first and is co-trained with a follow-up CNN architecture for classification. This model is further accelerated by integrating 1D convolution operators, leading to the upgraded model SigNet2.0. The simulations on two signal datasets show that both SigNet and SigNet2.0 outperform a number of well-known baselines. More interestingly, our proposed models behave extremely well in small-sample learning when only a small training dataset is provided. They can achieve a relatively high accuracy even when 1\% training data are kept, while other baseline models may lose their effectiveness much more quickly as the datasets get smaller. Such result suggests that SigNet/SigNet2.0 could be extremely useful in the situations where labeled signal data are difficult to obtain. The visualization of the output features of our models demonstrates that our model can well divide different modulation types of signals in the feature hyper-space.

I. INTRODUCTION

Radio signal classification is increasingly important as mobile-device use and 5G increase demand for limited spectrum resources. SigNet and SigNet2.0 address limitations in existing signal-classification approaches, with reported advantages in accuracy, efficiency, and small-sample learning.

I. INTRODUCTION

  • Radio classification methods based on hand-crafted features can require extensive expertise and may lack precision, limiting their use in complex scenes.Examples include features such as amplitude, frequency, phase, constellation diagrams, and time-frequency diagrams.Source: §I · p. 1
  • CNNs can automatically extract task-relevant features, but existing signal-classification CNNs may be too simple for complex scenes, while narrow 2D inputs can collapse feature maps to one dimension.LSTM models can capture time information but are described as time-consuming and difficult to train for long signals in complex tasks.Source: §I · p. 2
  • On RML2016.10a, SigNet scores 61.35% versus MCLDNN’s 61.16%; on Sig2019-12, 72.69% versus 67.52%, while SigNet2.0 reaches 62.30% and 72.87%, respectively, and both retain relatively high accuracy with much smaller training datasets.Source: §I · p. 2
  • SigNet transforms signals into matrices using trainable sliding operators before CNN classification, while SigNet2.0 integrates 1D and 2D convolutions to improve training efficiency and potentially accuracy.The paper describes S2M as a way to avoid tedious preprocessing and adaptively extract signal features.Source: §I · p. 2

II. METHOD

SigNet’s S2M operator converts the I and Q channels of an IQ signal into feature matrices using trainable filters. Sliding windows form signal-slice matrices, and the resulting feature matrices capture local signal relationships.

II. METHOD

  • SigNet applies separate S2M operators to the I and Q channels, producing two matrices that can be used as CNN inputs.The overall method is presented as parameter-trainable matrix operations followed by a typical deep-learning classifier.Source: A. S2M operator · p. 3, §II · p. 3

A. S2M operator

  • A sliding window of adjustable length k samples the I-channel signal, with default stride 1, to form N − k + 1 slices of length k.The signal slices are arranged into a matrix of size (N − k + 1) × k.Source: A. S2M operator · p. 3
  • A trainable k × k filter matrix is applied to the signal-slice matrix to form a feature matrix.The filter is initialized from a standard Gaussian distribution and updated during training.Source: A. S2M operator · p. 3
  • Each feature-matrix element is a transformed dot product of signal slices at different positions, allowing the matrix to capture local signal autocorrelation to some extent.The same operation is used to obtain a feature matrix for the Q channel.Source: A. S2M operator · p. 3

B. Work-flow of SigNet

SigNet concatenates the I- and Q-channel feature matrices into a two-channel image for CNN classification. Its filters and CNN are trained end-to-end, together or alternately, with this paper using joint training.

B. Work-flow of SigNet

  • The I- and Q-channel feature matrices are concatenated into a two-channel image and passed to a CNN for training and classification.The framework can use a typical CNN such as AlexNet, VGG, or ResNet.Source: B. Work-flow of SigNet · p. 3, Figure 1 · p. 4
  • The S2M filters learn matrix representations while the CNN learns higher-level features, and the complete framework is trained end-to-end.The filters and CNN may be trained together or alternately; this paper trains them together using the same training parameters.Source: B. Work-flow of SigNet · p. 4
  • The authors state that the framework uses CNN image-processing capacity without destroying the signals’ inherent structure.The passage says simulation results confirm the method’s performance.Source: B. Work-flow of SigNet · p. 4

C. Principal and advantages

SigNet learns a signal-to-matrix representation by mapping signal slices and computing their pairwise correlations. Its trainable representation generalizes the Gram matrix, while flexible filter and stride settings let the framework adapt to signal length.

C. Principal and advantages

  • A sliding window extracts continuous signal slices, and a linear mapping transforms each slice before pairwise correlations form the feature matrix.Source: C. Principal and advantages · p. 4
  • SigNet’s S2M feature matrix is a transformed correlation matrix of signal slices, generalizing the Gram matrix through a trainable filter.Unlike angular-field encoding, it transforms the slices’ representational space before computing their inner products.Source: C. Principal and advantages · p. 4
  • End-to-end training lets the CNN learn a more suitable representation, while breaking Gram-matrix symmetry can use information that differs when slice order is exchanged.Source: C. Principal and advantages · p. 4
  • The illustrated SigNet2.0 configuration replaces 2D convolutions with 1D convolutions in stages 1 and 2, selected for its highest classification accuracy.Source: Figure 2 · p. 5
  • SigNet can adjust filter size to target longer or shorter information units and change stride to reduce input size for long signals.Source: C. Principal and advantages · p. 5

D. Acceleration and upgrade by integrating 1D convolution

SigNet’s matrix conversion increases input size and computing demands, especially for long sequences. SigNet2.0 applies 1D convolutions before S2M to shorten the sequence, then uses the remaining CNN structure for classification.

D. Acceleration and upgrade by integrating 1D convolution

  • SigNet’s matrix conversion increases input size and computing demands, particularly for long sequences, motivating integration of 1D convolutions.Source: D. Acceleration and upgrade by… · p. 5
  • SigNet2.0 replaces early 2D convolution layers with 1D layers that shorten the input sequence before S2M converts it to a matrix for final classification.In the ResNet50 example, stages 1 and 2 use 1D convolutions; the remaining structure performs classification.Source: D. Acceleration and upgrade by… · p. 5
  • The shortened feature sequence could contain most useful information, and simulations report improved training efficiency without loss of accuracy for SigNet2.0.With an appropriate 1D/2D ratio, classification accuracy can also improve to some extent.Source: D. Acceleration and upgrade by… · p. 5

III. SIMULATIONS

The simulations were conducted to validate the models’ effectiveness, with deep learning models trained on an NVIDIA Tesla V100-PCIE.

III. SIMULATIONS

  • The simulations were conducted to validate the effectiveness of the proposed models.Source: §III · p. 5
  • Deep learning models were trained on an NVIDIA Tesla V100-PCIE.Source: §III · p. 5
  • The stated training setup specifies the GPU used for the simulations.Source: §III · p. 5

A. Modulation dataset

The study evaluates modulation classification on RML2016.10a and Sig2019-12. The datasets differ in signal length and modulation categories, and Sig2019-12 includes simulated communication-system nonidealities.

A. Modulation dataset

  • The methods are evaluated on two datasets for signal modulation classification; the cited passage establishes the dataset count and task but does not specify the dataset names.Source: A. Modulation dataset · p. 5
  • RML2016.10a contains 220,000 samples across 11 modulation categories, with 128-point signals and SNRs from -20dB to 18dB.Its data are split 6:2:2 into training, validation, and test sets.Source: A. Modulation dataset · p. 5
  • Sig2019-12 contains 468,000 samples across 12 categories, with 512 sampling points per sample and SNRs from -20dB to 30dB.Its simulations include carrier phase, pulse shaping, frequency offsets, and noise; data are split 6:2:2.Source: A. Modulation dataset · pp. 5–6

B. Models’ setup and baselines

The setup converts signals into matrix or sequence inputs for ResNet50 and compares SigNet with signal-to-matrix and deep-learning baselines. Models are trained and selected using validation accuracy before testing.

B. Models’ setup and baselines

  • SigNet converts each IQ signal into a two-channel matrix and feeds it to a ResNet50 classifier.The described inputs are 126 × 126 × 2 for 128-length signals and 510 × 510 × 2 for 512-length signals.Source: B. Models’ setup and baselines · p. 6
  • Baselines include reshaped signals, mathematical signal-to-matrix conversions, constellation diagrams, and existing CNN and LSTM classifiers.The CNN and LSTM baselines are narrow 2D-CNN, 1D-ResNet, MCLDNN, and a 2-layer LSTM.Source: B. Models’ setup and baselines · p. 6
  • Training uses Adam with warm-up cosine annealing, and the checkpoint with the highest validation accuracy is selected for testing.Batch sizes are 128 for RML2016.10a and 32 for Sig2019-12; Table I describes model architectures for RML2016.10a.Source: B. Models’ setup and baselines · p. 6, Table I · p. 6

C. Effectiveness of S2M operator

SigNet’s trainable S2M operator outperforms fixed signal-to-matrix conversions on both datasets. Against deep-learning baselines, its results vary by dataset, with a close comparison to MCLDNN on RML2016.10a and stronger performance on Sig2019-12.

C. Effectiveness of S2M operator

  • On both datasets, SigNet achieves higher accuracy than Reshape, GAF, MTF, and constellation-diagram conversions.The comparison trains on the full training set and reports test accuracy at each SNR; the authors suggest GAF and MTF vary across datasets, indicating weaker generalization.Source: C. Effectiveness of S2M operator · pp. 6–7

D. Comparison with other deep learning models

  • On RML2016.10a, SigNet reaches 61.35% test accuracy versus MCLDNN’s 61.16%; on Sig2019-12, it reaches 72.69% versus MCLDNN’s 67.52%.The paper reports SigNet performs significantly better than the other models on Sig2019-12, especially below 0 dB.Source: D. Comparison with other deep learning… · p. 7
  • On RML2016.10a, SigNet and LSTM reduce QAM64/QAM16 confusion seen with 1D-ResNet and 2D-CNN, while all models still confuse some AM categories.On Sig2019-12, SigNet and MCLDNN have more distinct diagonals, especially for OQPSK, 2FSK, 4FSK, and 8FSK.Source: D. Comparison with other deep learning… · p. 7
  • Compared with LSTM, SigNet saves 49.58% and 32.40% training time on RML2016.10a and Sig2019-12, respectively.The authors describe LSTM’s classification behavior as close to SigNet’s.Source: D. Comparison with other deep learning… · p. 7

E. The performance of SigNet2.0

With about 1% of training data, SigNet and SigNet2.0 retain accuracy while other deep-learning models decline more quickly. SigNet2.0 replaces selected 2D convolutions with 1D convolutions, improving accuracy and reducing training time relative to SigNet.

E. The performance of SigNet2.0

  • On RML2016.10a and Sig2019-12, SigNet2.0 improves accuracy by 0.95% and 0.18% while saving 30.65% and 12.46% training time, respectively, versus SigNet.The model replaces the ResNet50 stage1 and stage2 2D-convolution layers with 1D-convolution layers.Source: E. The performance of SigNet2.0 · p. 7
  • SigNet2.0 has the best classification performance among the evaluated methods, balancing accuracy and efficiency, while SigNet has more parameters than other models.The comparison notes that GAF and MTF training times include preprocessing and are listed only as a reference, not for comparison with other methods.Source: E. The performance of SigNet2.0 · p. 7

F. Learning with small samples

  • The small-sample evaluation varies the training fraction from 1/600 to 80/600 for RML2016.10a and from 1/900 to 100/900 for Sig2019-12.At the minimum fraction, each modulation category at each SNR level has one training sample; the paper motivates this setting by the cost of obtaining expert signal labels.Source: F. Learning with small samples · pp. 7, 9
  • Even with about 1% training data, SigNet and SigNet2.0 achieve 36.35% and 41.47% accuracy on RML2016.10a and 44.49% and 44.88% on Sig2019-12, respectively.With less than 15% training data, the respective accuracies are 56.94% and 56.63% on RML2016.10a, and 67.67% and 68.10% on Sig2019-12; other deep-learning models drop more quickly as training data shrink.Source: F. Learning with small samples · p. 9

G. A brief ablation study of SigNet

SigNet’s hyperparameter ablation found little accuracy sensitivity across tested filter sizes and, on Sig2019-12, batch sizes. Changing the ResNet structure generally had little effect, while SigNet with S2M performed better than other ResNet frameworks using GAF or MTF.

G. A brief ablation study of SigNet

IV. EXPLANATORY VISUALIZATION OF MODELS

The feature visualizations show clearer modulation-category boundaries for SigNet, SigNet2.0, and MCLDNN than for three other models, consistent with their classification results. The paper concludes that SigNet and SigNet2.0 retain comparable accuracy with very small training sets, and may be useful when labeled signals are difficult to obtain.

IV. EXPLANATORY VISUALIZATION OF MODELS

  • With test samples at maximum SNR—18 dB for RML2016.10a and 30 dB for Sig2019-12—t-SNE visualizes model output features in two dimensions after the softmax layer is removed.The visualization uses 300 iterations, learning rate 1000, perplexity 30, and momentum changing from 0.5 to 0.8; confusion matrices appear beside the plots.Source: §IV · p. 10, Figure 8 · p. 11
  • Misclassified categories overlap in t-SNE, while SigNet, SigNet2.0, and MCLDNN show clearer category boundaries than 1D-ResNet, 2D-CNN, and LSTM.The visualization is reported as consistent with the classification results in Table II.Source: §IV · p. 11

V. CONCLUSION

  • SigNet and SigNet2.0 give comparable accuracy with very small training sets, suggesting usefulness where labeled signals are difficult to obtain.The conclusion also proposes testing other CNN architectures and signal datasets, and adapting the model to samples containing multiple modulations.Source: §V · p. 12
  • The paper proposes testing other CNN architectures and signal datasets, and adapting SigNet to samples containing multiple modulations.Source: §V · p. 12

Related papers

Loading 2011.03525v2…