Robust Deep Architectures for Low-SNR Automatic Modulation Recognition in Nonstationary Wireless Channels Using Residual Channel-Attentive Feature Fusion

Abstract

Wireless receivers increasingly operate in congested, fading, and interference-rich environments in which the received signal energy may approach or even fall below the effective noise floor. In such settings, automatic modulation recognition is no longer a simple pattern classification problem, because the observable waveform is jointly distorted by channel filtering, carrier frequency offset, phase uncertainty, timing mismatch, and nonstationary noise. Classical likelihood-based and handcrafted-feature pipelines remain informative, but their performance often degrades when the assumed signal model is violated or when discriminative structures are weakly expressed in short observation windows. This paper develops a technical framework for low-SNR automatic modulation recognition that combines signal-domain modeling with a residual channel-attentive deep architecture designed to preserve shallow waveform evidence while progressively refining class-discriminative latent features. The proposed formulation treats modulation recognition as inference over short in-phase and quadrature sequences and emphasizes the preservation of local envelope, phase-transition, and short-memory temporal signatures that are frequently lost in purely deep hierarchical compression. A grouped residual backbone is therefore augmented with cross-stage feature fusion and channel-reweighting operations that selectively amplify informative latent responses under adverse noise conditions. The paper further analyzes optimization under SNR uncertainty, calibration of posterior outputs, and the interaction between architectural depth, receptive field growth, and feature separability. The resulting perspective is not restricted to a single dataset or benchmark protocol; instead, it provides a general, mathematically grounded view of why residual aggregation, feature reuse, and adaptive channel emphasis are particularly relevant for low-SNR recognition, where class manifolds overlap and inter-class confusion is driven more by nuisance variability than by nominal constellation geometry.