Multimodality-Aided Multicarrier Waveform Recognition in Low SNR Regimes Based on Denoised Cyclic Autocorrelation Transformation

Wireless signal recognition driven by artificial intelligence (AI) plays a pivotal role in 6G ultrareliable wireless communications, facilitating spectrum surveillance to impede illegal radio interference. As a promising wireless technique, multicarrier waveform recognition (MWR) has been explored to enhance the reliability of wireless data transmission. However, existing works fail to achieve reliable recognition accuracy under strong noise. It remains a daunting task for MWR in signal-to-noise ratio (SNR) regimes. To deal with this issue, we propose a denoised cyclic autocorrelation-based multimodality fusion network (DCA-MFNet). Specifically, we first leverage the cyclic autocorrelation (CA) transformation to convert intercepted signals into CA features in cyclic frequency domains, which have the robust property of being insensitive to low SNR. Next, the singular value decomposition method is employed to weaken the strong noise effects on useful CA peak values. Based on the denoised CA matrix (DCAM), the projection accumulation strategy is proposed to generate the time delay accumulation vector (TDV) and cyclic frequency accumulation vector (CFV), which can enlarge the discrimination among multicarrier signals. Finally, we fed the multimodality features of DCAM, TDV, and CFV into the developed DCA-MFNet to perform hierarchical learning, feature aggregation training, and multicarrier type prediction. Experimental results demonstrate that the proposed DCA-MFNet obtains better recognition performance than existing algorithms. Moreover, DCA-MFNet can effectively identify six multicarrier signals with a recognition accuracy of 100% at a low SNR of even −2 dB.