Holistic Fusion of Fragmented Signal Features for Automatic Modulation Recognition via an Adaptive Topological Network

Abstract

Automatic modulation recognition (AMR) remains challenging under low-signal-to-noise ratio (SNR) conditions, where severe noise can obscure weak modulation-specific waveform patterns. To address this issue, this paper proposes an Adaptive Holistic Fusion Network (AHFN) for robust low-SNR modulation recognition. AHFN first employs a multi-resolution nonlinear fusion module composed of parallel KAN branches with different spline-grid resolutions to extract complementary waveform dynamics from standardized I/Q samples and temporal positional information. An adaptive soft-threshold denoising module then generates node- and channel-specific thresholds to suppress noise-sensitive responses while preserving discriminative modulation cues. Subsequently, a topology-aware multi-scale fusion network performs feature- and structure-adaptive message processing over a fixed temporal graph and combines multi-scale graph representations with Mamba-based long-range sequence modeling. Experiments on RML2016.10A and RML2016.10B show that AHFN consistently improves recognition accuracy over representative baseline methods from −20 dB to 0 dB, with gains ranging from 3.03% to 14.34%. These results demonstrate the effectiveness of multi-resolution nonlinear representation, adaptive denoising, and topology-aware local–global fusion for low-SNR AMR.