Abstract
In increasingly complex electromagnetic environments, wireless communication systems face the severe challenge of non-Gaussian impulse noise. The moments of impulse noise tend toward infinity, reducing the distinguishability of signal features and thereby limiting improvements in signal modulation recognition rates. First, a time-frequency analysis method based on kernel space mapping is proposed to improve the distinguishability of time-frequency features in signals under impulse noise. On this basis, a multimodal kernel harmonic feature fusion network is constructed, combining convolutional neural networks and graph convolutional networks to extract and fuse kernel harmonic features from three modalities to achieve robust and accurate modulation recognition. The simulation results show a generalized signal-to-noise ratio of -2 dB, and the modulation recognition rate reaches 93.5%.