Graph Learning-Based Method for Automatic Modulation Identification in Unknown HF Environments

Automatic modulation identification (AMI) is crucial in non-cooperative high frequency (HF) communication systems, especially for spectrum monitoring and management. However, the dynamics and open nature of HF environments pose many challenges for AMI, including the presence of unknown modulation types and shifts in signal data distributions caused by environmental changes. To address these issues, this paper presents a graph learning-based open-set domain adaptation (GLOSDA) method using multi-modal features for AMI in unknown HF environments. The method employs a progressive learning strategy to iteratively assign pseudo-labels to signals from unknown HF environments and optimize model parameters for improved performance. Through graph learning, GLOSDA propagates modulation information between signals from different environments, aligning feature distributions of known and unknown environments to enhance performance in unknown HF environments. A shared-dummy classifier is used to differentiate known modulation types and separate known and unknown types by employing multiple prototypes. In addition, conditional adversarial learning and prototype learning are incorporated to reduce distribution shifts and refine decision boundaries for known types. Experimental results indicate that GLOSDA significantly outperforms baseline methods in various scenarios, demonstrating its effectiveness and robustness in complex HF environments.