Ultralight Signal Classification Model for Automatic Modulation Recognition

TL;DR

This work tackles automatic modulation recognition (AMR) under strict edge-resource constraints by introducing an ultralight XLW-CNN-LSTM hybrid that processes Cohen's class time-frequency representations to classify 11 intrapulse modulations. It achieves a worst-case accuracy of $93$.$8\%$ and an average accuracy of $96.3\%$ at $0$ dB SNR on a small synthetic dataset, aided by a sample-specific error-rate–driven augmentation strategy and a compact network (~12,059 parameters, ~47 kB). The approach outperforms larger architectures with only about 1% of their parameters and remains robust across $-20$ to $+20$ dB SNR, while enabling edge and photonic deployments for real-time AMR in defense and civilian wireless systems. The work highlights the value of targeted data augmentation, temporal modeling via LSTM, and efficient time-frequency preprocessing for practical, low-latency AMR on resource-constrained platforms.

Abstract

The growing complexity of radar signals demands responsive and accurate detection systems that can operate efficiently on resource-constrained edge devices. Existing models, while effective, often rely on substantial computational resources and large datasets, making them impractical for edge deployment. In this work, we propose an ultralight hybrid neural network optimized for edge applications, delivering robust performance across unfavorable signal-to-noise ratios (mean accuracy of 96.3% at 0 dB) using less than 100 samples per class, and significantly reducing computational overhead.

Figures (8)

• Figure 1: Examples of intrapulse signals: (a) BFSK signal, (b) P4 signal.
• Figure 2: Examples of the spectrograms of intrapulse modulation signals at an SNR of 20 dB.
• Figure 3: Examples of noisy modulation signals for varying SNR levels.
• Figure 4: Performance comparison of the four model configurations.
• Figure 5: Comparison of our model against state-of-the-art methods.