Open Set Modulation Recognition via Multimodal Contrastive Learning With Knowledge-Guided Diversity Negative Sampling

The presence of unknown signals in complex electromagnetic environments highlights the critical need for open set modulation recognition. This task remains challenging due to the scarcity of unknown signals for training and the inherent ambiguity in decision boundaries. To address these issues, this work proposes a novel knowledge-guided multimodal framework that jointly improves closed set recognition and open set detection. Specifically, we first construct multimodal signal representations by integrating time-frequency representations with semantic representations. The framework employs a dual-encoder architecture to align multimodal representations within a shared semantic space. Next, we introduce a novel regularization term that works synergistically with contrastive loss to not only ensure cross-modal semantic alignment but also actively separate non-matching signal pairs. This results in more compact decision boundaries and enhanced discrimination against unknown classes. Furthermore, to prevent over-confident misclassification of unknown signals, we devise a knowledge-guided diversity sampling strategy. This strategy first constructs a candidate negative prompt pool from generic communication-domain physical priors, encodes these prompts into candidate negative semantic features, and then employs a greedy diversity sampling algorithm to select a compact set of negative semantic anchors for robust open set inference. Extensive experiments on the RadioML2016.10a dataset demonstrate that the proposed framework achieves 95.94% closed set recognition accuracy and 0.9394 AUROC for unknown signal detection, validating its effectiveness in both known-class recognition and unknown-class rejection.