Synergistic Dual-Stream Neural Network Based on Adaptive Variational Mode Decomposition Denoising for Automatic Modulation Classification

Automatic Modulation Classification (AMC) is a key technology for electromagnetic signal cognition and plays a pivotal role in intelligent spectrum sensing and management. Although deep learning-based AMC has become the mainstream paradigm, its potential is constrained by insufficient internal synergy in hybrid network architectures and the degradation of feature extraction capability at low signal-to-noise ratios (SNRs). To address these limitations, this paper proposes an adaptive variational mode decomposition (VMD) denoising-based synergistic dual-stream network (AVDSDN) for modulation classification. The proposed network decouples the traditional hybrid network structure, enabling it to mitigate the impact of noise and achieve full synergistic utilization of different network feature extraction capabilities. AVDSDN comprises three core components: an adaptive VMD denoising module, a dual-stream feature extraction module, and a fusion module. Specifically, we first propose an adaptive VMD denoising mechanism that selectively enhances key signal components while suppressing noise components. Subsequently, we develop a synergistic dual-stream module comprising parallel CNN and Transformer streams. In the CNN stream, a multi-feature interaction extraction block and a lightweight backbone are designed to capture robust local and spatial features, while the Transformer stream innovatively integrates deformable attention and convolutional feed-forward networks to efficiently model long-range temporal dependencies. Finally, a cross-stream attention fusion module is designed to synthesize features from both streams. Simulations on three public datasets demonstrate that AVDSDN achieves superior performance compared to current state-of-the-art (SOTA) algorithms, especially under low SNR conditions. Furthermore, by employing post-training quantization (PTQ), AVDSDN achieves a 5- to 9-fold speedup in inference while maintaining high accuracy, demonstrating its suitability for deployment in real-time systems.