Machine Learning-based Automatic Modulation Classification for 5G-Advanced and 6G Waveforms: Robust Identification Under Realistic Channel Impairments

Abstract

Automatic modulation classification (AMC) for 5G-Advanced and 6G networks must blindly identify waveforms from received signals under realistic channel impairments, enabling cognitive radio dynamic spectrum access and interference avoidance. No prior work has simultaneously applied machine learning to classify all eight leading waveforms (UFMC, GFDM, FBMC, NOMA, OFDM-IM, OTFS, ODDM, and AFDM) under realistic channel impairments, nor quantified the minimum feature set for resource-constrained deployment.
We present a framework that (i) extracts a 38-dimensional feature vector that includes three novel channel-aware characteristics (amplitude fading variance, phase discontinuity, and frequency drift); (ii) benchmarks nine machine learning classifiers, including an FC-MLP deep learning baseline and five feature selection methods, on 201600 signals across twelve channel conditions (nine custom plus three 3GPP TDL profiles) and seven SNR levels, with leakage-free feature selection; and (iii) identifies a compact 10-feature subset validated with Bonferroni-corrected McNemar tests and Wilson confidence intervals.
FC-MLP achieves 99.09% accuracy; ensemble-boost (99.04%) and random forest (99.02%) are statistically equivalent. The 10-feature random forest reaches a score of 98.90% within 0.12 pp of the full feature baseline at a cost that is 74% lower and with a 0.071 ms inference per block. The five-fold cross-validation confirms stability (98.54%, Wilson 95% CI: 98.49%, 98.59%). Per channel accuracy ranges from 98.87% (Rayleigh) to 99.98% (AWGN/Rician); 3GPP TDL-A/B/C profiles confirm transferability to 5G NR. The three channel-aware features yield up to 3.1% gain under double-selective fading and an average overall improvement.