MSNCIL: A Domain-Agnostic Class-Incremental Learning Method Tailored for Automatic Modulation Recognition
Zhiliang Deng, Chunbo Luo, Zixi Tang, Yang Luo · IEEE Communications Letters · 2025
The emergence of new modulation types in 6G challenges the adaptability of deep learning-based automatic modulation recognition (DL-AMR) models. This letter presents multi-state neuron class-incremental learning (MSNCIL), the first domain-agnostic class-incremental learning (CIL) method for AMR. Leveraging the sparsity of wireless signal features, MSNCIL dynamically partitions a DL-AMR model into specialized sub-models, each dedicated to different modulation types. In each session, neurons are selected based on activation values, trained, frozen, and assigned state values. During inference, the session ID of a test sample is identified, which directs the corresponding neurons for recognition. Extensive experiments confirm MSNCIL’s effectiveness.