GMFFR-Net: Geometry-Guided Manifold Feature Fusion Network for Automatic Modulation Classification

Automatic modulation classification (AMC) is a key enabling technology in wireless communication systems, where accurate recognition is crucial for enhancing signal processing efficiency and ensuring data security. To address AMC performance degradation under low signal-to-noise ratio (SNR) and data-limited conditions, this article proposes a geometry-guided manifold feature fusion model, termed the geometric manifold feature fusion recognition network (GMFFR-Net), based on nonlinear dynamical representations and multimetric geometric constraints on symmetric positive definite (SPD) matrix manifolds. The model constructs a multiperspective geometric analysis framework: the attractor trajectory manifold from phase-space reconstruction (PSR) characterizes the dynamical geometry; the root-mean-square (rms) energy trajectory characterizes the temporal geometry of energy evolution; and the SPD manifold feature space represents statistical geometry. Furthermore, by aligning and integrating across these views, the model achieves fusion modeling driven by geometric consistency, enabling the network to form a unified discriminative representation under multiple geometric prior constraints, thus obtaining a more stable and physically interpretable decision boundary. Experiments on four datasets, namely RadioML2016.10a, RadioML2016.10b, CustomMOD-2026.a, and RadioML2018.01A, demonstrate that GMFFR-Net consistently achieves performance improvements across different SNR ranges and observation scales, with average accuracies of approximately 64%, 66%, 72%, and 60%, respectively. In conclusion, GMFFR-Net, with its core of multiperspective geometric consistency fusion, incorporates dynamical, temporal, and statistical geometric priors into end-to-end discriminative modeling, achieving modulation recognition that balances robustness and physical interpretability.