DKDNet: Dual Knowledge and Data-Driven Network for Cross-Domain Automatic Modulation Classification

This paper addresses distribution shifts in automatic modulation classification caused by changing communication environments. It proposes DKDNet, which combines signal-prior representations with data-driven domain adaptation. After analyzing five signal representations, the authors select in-phase/quadrature (IQ), amplitude-phase (AP), and autocorrelation function (ACF) inputs. A multi-representation feature encoder and dynamic lightweight fusion unit learn and adaptively combine features, while modulation-classification and adversarial domain-alignment objectives are optimized jointly. Experiments on simulated and public datasets are reported to validate the design and performance.