Deep Multi-Scale Representation Learning with Attention for Automatic Modulation Classification

Xiaowei Wu, Shengyun Wei, Yan Zhou

TL;DR

Researchers have developed SE-MSFN, a novel deep learning architecture that significantly improves Automatic Modulation Classification (AMC) for wireless signals. By combining large convolutional kernels, multi-scale feature fusion, and Squeeze-and-Excitation (SE) attention mechanisms, the model achieves state-of-the-art performance on the standard RADIOML 2018.01A dataset, particularly excelling in low Signal-to-Noise Ratio (SNR) environments.

What Problem Does This Paper Address?

In wireless communications, Automatic Modulation Classification (AMC) is the critical "middle step" between detecting a signal and demodulating it. It identifies how data is encoded into a radio wave (e.g., QAM, PSK, or AM). This is vital for cognitive radio, spectrum monitoring, and electronic warfare.

The challenge is that signals are often buried in noise. While deep learning has replaced traditional manual feature engineering, most current models use small convolutional filters (like $3 \times 3$) borrowed from computer vision. The authors argue that radio signals are more like text sequences than images; they possess long-term dependencies and forward/backward correlations that small filters miss. Furthermore, existing models often treat all feature channels as equally important, failing to prioritize the most informative parts of a noisy signal.

Methods

The authors propose the Squeeze-and-Excitation Multi-Scale Feature Network (SE-MSFN). Its architecture is built on three pillars:

  1. Large Kernel 1D Convolutions: Instead of stacking many small filters, the model uses $9 \times 1$ kernels. This larger "receptive field" allows the network to capture global features and long-term correlations in the raw In-phase and Quadrature (I/Q) sequence data more effectively.
  2. Multi-Scale Feature Fusion: Inspired by HRNet, the model extracts features at different resolutions simultaneously and fuses them. This ensures that the network doesn't just look at fine-grained details but also sees the "big picture" of the signal's structure.
  3. Squeeze-and-Excitation (SE) Blocks: This is an attention mechanism that adaptively recalibrates channel-wise feature responses. It "squeezes" global information into a descriptor and "excites" it to emphasize important feature channels while suppressing noise-heavy, redundant ones.
  4. Ensemble Learning: To squeeze out extra performance, the authors combined multiple versions of the model (trained with different kernel sizes) using a voting mechanism.

Datasets

The model was evaluated on the RADIOML 2018.01A dataset, a widely recognized benchmark in the AMC community. It contains:

  • 24 modulation types (both analog and digital).
  • SNR range from -20dB (extremely noisy) to +30dB (very clean).
  • 2.55 million samples, divided into training, verification, and test sets at an 8:1:1 ratio.

Key Insights and Ideas

  • Signals $\neq$ Images: The most significant "aha moment" is the realization that 1D signal sequences benefit from larger kernels. In computer vision, small kernels are preferred to build hierarchy, but in signal processing, capturing the temporal flow over a wider window is more efficient for identifying modulation patterns.
  • Attention is a Noise Filter: The SE blocks act as a dynamic filter. By learning which feature channels are most relevant for a specific modulation type, the model becomes much more robust against the random fluctuations of noise.
  • The "Analog Confusion" Gap: The authors noted that while digital modulations are classified with near-perfect accuracy at high SNR, certain analog modes (like AM-SSB-WC vs. AM-DSB-WC) remain difficult to distinguish because their features are fundamentally similar, suggesting a need for specialized sub-networks.

Key Results

The SE-MSFN model set a new bar for the RADIOML 2018.01A dataset:

  • Overall Average Accuracy: 64.50%, surpassing the popular CLDNN model by 1.42%.
  • Low SNR Performance (0dB to 10dB): Achieved 85.53% accuracy, a 2.85% improvement over CLDNN. This is the range most critical for real-world applications.
  • Maximum Accuracy: Reached a peak of 98.5% at high SNR.
  • Efficiency: Despite the "large kernel" name, the model uses 1D convolutions, resulting in only 332K parameters—significantly lighter than CLDNN (847K) or standard CNNs (466K).

Limitations and Future Work

While the model is highly accurate, the authors acknowledge a few hurdles:

  • Inference Latency: The SE-MSFN takes slightly longer to predict (0.137ms) compared to simpler models like ResNet (0.051ms) or VGG (0.043ms). While still fast, this might be a consideration for ultra-low-latency hardware.
  • Analog Similarity: The confusion matrix revealed that the model still struggles to differentiate between specific types of Amplitude Modulation (AM).
  • Future Direction: The authors suggest a two-stage classification strategy—using the primary network to identify the broad class and a specialized secondary network to resolve confusion between highly similar analog signals.

Summary updated

Abstract

Currently, deep learning methods with stacking small size convolutional filters are widely used for automatic modulation classification (AMC). In this report, we find some experienced improvements by using large kernel size for convolutional deep convolution neural network based AMC, which is more efficient in extracting multi-scale features of the raw signal I/Q sequence data. Also, Squeeze-and-Excitation (SE) mechanisms can significantly help AMC networks to focus on the more important features of the signal. As a result, we propose a multi-scale feature network with large kernel size and SE mechanism (SE-MSFN) in this paper. SE-MSFN achieves state-of-the-art classification performance on the public well-known RADIOML 2018.01A dataset, with average classification accuracy of 64.50%, surpassing CLDNN by 1.42%, maximum classification accuracy of 98.5%, and an average classification accuracy of 85.53% in the lower SNR range 0dB to 10dB, surpassing CLDNN by 2.85%. In addition, we also verified that ensemble learning can help further improve classification performance. We hope this report can provide some references for developers and researchers in practical scenes.
BibTeX
@misc{wu2022deep,
  author = {Xiaowei Wu and Shengyun Wei and Yan Zhou},
  title = {{Deep Multi-Scale Representation Learning with Attention for Automatic Modulation Classification}},
  year = {2022},
  eprint = {2209.03764},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2209.03764},
}
Figure