CMNet: Radar-Communication Waveform Recognition via Convolution and Mamba Networks

With the spectrum for modern wireless networks becoming increasingly limited, the integration of radar and communications through mechanisms that allow for spectrum sharing has shown significant potential. This research introduces a cutting–edge method for the classification of radar and communication waveforms, combining time–frequency analysis via smooth pseudo Wigner–Ville distribution (SPWVD) with deep learning. We propose an advanced neural network, the radar–communication waveform recognition convolution mamba network (CMNet), which demonstrates exceptional capabilities in feature extraction and capturing long–term information dependencies. CMNet implements hierarchical feature learning using advanced components: Dense Sparse Fusion Convolution, Visual State Space, and Attention Down Scale to efficiently process both local and global signal characteristics. Experimental results benchmarked on the dataset with 12 distinct signal types under realistic channel impairments indicate that CMNet attains high accuracy of 90.70%. Remarkably, CMNet achieves this accuracy with remarkable cost-efficiency, utilizing only 38K parameters, at least 3.6 times smaller than compared state-of-the-art models, and demonstrating superior computational speed with an average inference time of 0.059 ms.