Toward Robust Generalization in Automatic Modulation Recognition Under Channel, Parameter, and Distributional Shift Conditions

• Rafael Mendes

  Departamento de Engenharia Elétrica, Universidade Federal do Pampa, Avenida Pedro Anunciação 111, Vila Batista, Bagé, Brazil

  Author

• Thiago Ribeiro

  Departamento de Engenharia Elétrica e Eletrônica, Universidade Federal Rural do Semi-Árido, Avenida Francisco Mota 572, Costa e Silva, Mossoró, Brazil

  Author

• Lucas Carvalho

  Departamento de Engenharia Elétrica, Universidade Estadual do Oeste do Paraná, Rua Universitária 2069, Jardim Universitário, Foz do Iguaçu, Brazil

  Author

Background wireless receivers increasingly operate in settings where the statistical properties of signals vary across hardware platforms, propagation conditions, waveform parameterizations, and interference structures. Background constraints in these environments make automatic modulation recognition a generalization problem rather than a narrow classification exercise. A model that performs well only on a single curated dataset has limited operational value, because deployment commonly introduces carrier offsets, timing errors, pulse-shape mismatches, front-end distortion, unequal sampling assumptions, and previously unseen mixtures of nuisance factors. This paper develops a technical treatment of generalization in automatic modulation recognition by integrating statistical learning concepts with communication-theoretic signal modeling. The discussion formulates modulation recognition as supervised inference under latent environment variables and studies how inductive bias, representation geometry, optimization, and evaluation protocol shape the gap between in-domain accuracy and out-of-domain reliability. The analysis emphasizes invariance to nuisance transformations, sensitivity to hidden coupling between modulation identity and dataset construction, and the importance of environment-aware splitting strategies. It further examines how channel models, augmentation policies, transfer learning, calibration, uncertainty estimation, and deployment-time adaptation affect robustness. The paper also proposes mathematical formulations for invariant feature extraction, risk under environment shift, and representation regularization, while discussing their implications for complex baseband inputs. The central argument is that generalization in automatic modulation recognition emerges from joint design across data generation, architecture, training objective, and validation regime. Progress therefore depends less on maximizing a single benchmark score and more on constructing receivers whose internal representations remain stable when the physical realization of the same modulation class changes.

• PDF

Copyright (c) 2023 International Journal of Scientific and Applied Computing Research

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.