Multi-Feature Multi-Channel Input-Level Time-Frequency Fusion for Integrated Sensing and Communication (ISAC) Signal Classification

Abstract

The increasing demand for efficient spectrum utilization in modern wireless and defense systems has accelerated the development of Integrated Sensing and Communication (ISAC) technologies. While extensive research has been conducted on the classification of radar and communication signals individually, and their coexistence in shared spectral environments, no prior study has specifically addressed the classification of ISAC waveforms designed to perform both functions simultaneously. In this study, we investigate the effectiveness of multi-channel input-level time-frequency representations for ISAC waveform recognition. Specifically, short-time Fourier transform (STFT), continuous wavelet transform (CWT) and Wigner-Ville distribution (WVD) images were generated and used as inputs to convolutional neural networks, ShuffleNet and ResNet-18. A multi-channel input-level fusion strategy was employed, where complementary time-frequency representations were mapped to separate RGB channels and jointly processed by a single CNN. Experiments on complex ISAC environments containing LFM, Frank, Legendre, ISAC-LFM, and ISAC-ZC waveforms reveal that ISAC-LFM signals can be reliably classified even under low SNR conditions, while the proposed multi-channel structure significantly improves ISAC-ZC detection performance.