Deep Domain-Adversarial Adaptation for Automatic Modulation Classification under Channel Variability

Abstract

Automatic Modulation Classification (AMC) plays a significant role in modern cognitive and intelligent radio systems, where accurate identification of modulation is crucial for adaptive communication. The presence of heterogeneous wireless channel conditions, such as Rayleigh and Rician fading, poses significant challenges to the generalization ability of conventional AMC models. In this work, a domain-adversarial neural network (DANN) based deep learning framework is proposed that explicitly mitigates channel-induced distribution shifts between source and target domains. The approach is evaluated using a comprehensive simulated dataset containing five modulation schemes (BPSK, QPSK, 16QAM, 64QAM, 256QAM) across Rayleigh and Rician fading channels at five frequency bands. Comparative experiments demonstrate that the DANN-based model achieves up to 14.93% absolute accuracy improvement in certain modulation cases compared to a baseline supervised model trained solely on the source domain. The findings establish the engineering feasibility of domain adversarial learning in AMC tasks under real-world channel variability and offer a robust direction for future research in adaptive spectrum intelligence