Deep Spatio-Temporal Graph Learning with Batch Normalization for Modulation Classification

Abstract

Automatic modulation classification (AMC) enables an intelligent receiver to identify an unknown modulation format before demodulation. This paper presents a spatio-temporal graph convolutional network (ST-GCN) with batch normalization for classifying raw in-phase and quadrature (I/Q) signal windows. The proposed method performs DC removal, power normalization, fixed-length framing, and graph construction in which I/Q observations form nodes and temporal or similarity relationships form edges. Spatial graph convolution captures relational constellation structure, while temporal convolution learns sample evolution. Global pooling and batch-normalized fully connected layers produce class probabilities. MATLAB R2024b results for a noisy 16-QAM test signal show a final 16-QAM probability of 0.85242 and a QPSK probability of 0.14758, corresponding to 85.24% confidence over seven windows of 256 samples. The constellation contains approximately sixteen grid-like clusters, and the estimated overall SNR is about -11.23 dB. The time-domain, envelope, phase, power-spectrum, SNR, and eye-diagram outputs demonstrate the difficult low-SNR operating condition. The results support graph-based relational learning for robust AMC, while also revealing that visualization labels must be generated from the same final decision variable to avoid inconsistent reporting.