Abstract
Automatic modulation recognition (AMR) is an important research topic in the area of non-cooperative communication and cognitive radio. Existing AMR studies have combined data-driven preprocessing modules with feature extraction networks to construct effective AMR architectures. However, most preprocessing modules are not adaptable to low SNR scenarios, and the feature extraction mechanisms are generally homogeneous for different modulation types, leading to unsatisfied generalization. To address these limitations, this paper proposes an adaptive entropy adjustment (AEAD) algorithm for noise suppression and signal purification, together with a hierarchical AMR scheme that leverages both modulation-shared and modulation-specific characteristics. Specifically, the AEAD algorithm processes in-phase/quadrature (I/Q) signals weighted by the quadratic Rényi entropy metric, while the hierarchical AMR scheme employs a main network for shared features and two sub-networks for modulation-specific features. Evaluations on the Rician datasets RML2016.10A/10B and a Rayleigh dataset show that the proposed method consistently outperforms baseline approaches, and exhibits strong adaptability under diverse channel conditions.