Get started
Work

Compositional Zero-Shot Recognition Based on Disentangled Semantic Space for Composite Modulation Signals

Year: 2026
Type: article
Abstract: Automatic composite modulation recognition (ACMR) is essential for enabling reliable signal awareness in integrated sensing and communication (ISAC) systems. However, existing ACMR approaches still face several fundamental challenges, including the intrinsic coupling between inner-layer and outer-layer modulations in composite modulation (CM) signals, performance degradation under joint hardware and channel impairments, and limited generalization in open-set scenarios. To this end, we construct a disentangled semantic space and propose a compositional zero-shot learning framework for recognizing unseen waveform combinations composed of known modulation types. Within this framework, a logarithmic projection first linearizes the multiplicative coupling between modulation layers, after which a learnable geometric transformation extracts layer-specific semantic features. To instantiate this framework, we develop a modulation component disentanglement network (MCDN). The MCDN employs a spatial transformer network to learn the geometric transformation and incorporates a multi-objective loss function to balance discriminative capability and cross-domain generalization. Comprehensive experiments demonstrate that the MCDN achieves over 93% accuracy in compositional zero-shot recognition and substantially outperforms both unified-semantic and multi-task baselines. Moreover, it remains robust to the combined effects of channel fading and hardware imperfections at SNRs as low as 4 dB.
DOI: 10.1109/tccn.2026.3737916
Language: English
Open Access status: hybrid
Estimated APC paid: $2,645
OpenAlex ID: W7214647883
FWCI:
Cites: 57
Cited by:
Related to: 10