DSCAN-Net: A Dual-Stream Network for Classifying Modulation Schemes in OTFS Systems

Sixth-generation (6G) wireless communication systems are required to ensure seamless connectivity among devices in high-mobility scenarios. In such scenarios, these systems must prevail over the challenges imposed by inter-carrier and inter-symbol interferences arising in doubly-spread channels (delay and Doppler shifts) with low-transmit power. Orthogonal time-frequency space (OTFS) has emerged as a promising solution in doubly-spread channels. Further, classifying the modulation schemes present in received signals accurately is very challenging under low signal-to-noise ratio (SNR) conditions. To address these challenges, this work proposes DSCAN-Net: a dual-stream cross-attention network that leverages the advantages of temporal in-phase/quadrature (I/Q) and spatial amplitude/phase (A/P) formats of the received OTFS signal in two separate streams. DSCAN-Net combines a convolutional neural network (CNN) and residual channel attention network (RCAN) to extract meaningful representations from these two formats separately. These representations are mapped with each other using multi-headed cross-attention to enhance the cross-context representation learning that improved the overall classification accuracy to 72.92%. Experimental results demonstrate the effectiveness of the proposed DSCAN-Net under low SNR conditions and claim superior accuracy over existing approaches utilized in OTFS systems. Additionally, the results also signify the effectiveness of A/P formats over I/Q in modulation classification.