Automatic modulation recognition in OTFS systems using optimized SE-CNN architecture

Abstract

Automatic modulation recognition (AMR) plays an important role in wireless communications due to its ability to enable adaptive modulation to cope with dynamic nature of the wireless channels. However, the implementation of AMR in delay-Doppler channels presents significantly greater challenges due to the complex time-variant and frequency-selective characteristics of such environments. Very little work is done to date, most of which suffers from the limited accuracy in low signal-to-noise ratio (SNR) cases. %, yet classical likelihood- and feature-based methods either impose prohibitive complexity or degrade sharply at low SNR - challenges amplified for delay-Doppler waveforms like OTFS. Prior SE-CNN designs reach near-perfect accuracy only above 8 dB SNR. Here, In this paper, we propose an improved machine learning model exploiting squeeze-and-excitation CNN (SE-CNN) that fuses, and task-specific hyper-parameter optimization. Trained on 90,000 IQ samples covering OTFS, OFDM, 2-ASK, PAM-4, and BPSK over -10 to 12 dB, the model attains 100% accuracy at 6 dB and 98.7% at 0 dB, a 6 dB gain over the baseline. ROC analysis yields macro-AUC ? 1.000 with 0.99 TPR at FPR ?10-3, extending reliable AMR 4 dB deeper into noise for real-time, high-mobility applications.