Decoupled Spatiotemporal Convolutional Network with Uncertainty-Aware Learning for Non-Cooperative Incremental Modulation Recognition

Automatic Modulation Recognition (AMR) in open wireless environments is challenging due to emerging modulation types and evolving data distributions. Class-incremental learning addresses this by enabling models to progressively learn new modulation classes without retraining from scratch, thus improving scalability and adaptability. To this end, we propose DUCNet, a Decoupled and Uncertainty-aware Convolutional Network tailored for class-incremental AMR. DUCNet employs a specialized convolutional module to separately model temporal and cross-component dependencies in I/Q signals, enhancing robustness to distribution shifts. Then, an uncertainty-aware mechanism further improves the detection and adaptation to new modulation types. Experiments demonstrate DUCNet's superior performance and generalization in class-incremental settings.