Automatic Modulation Recognition of Radio Frequency Proximity Sensor Signals Based on Adaptive Relational Graph Attention Network

Automatic modulation recognition (AMR) plays a critical role in signal reconnaissance. However, due to the high-frequency, low-power characteristics, and significant propagation loss of radio frequency proximity sensor (RFPS) signals, accurately achieving AMR at low signal-to-noise ratios (SNRs) remains a substantial challenge. To address this, we propose an adaptive relational graph attention network (ARGAT) for AMR in low-SNR conditions. In the preprocessing phase, time-series data from multiple raw signals are rearranged into a 2-D feature matrix to preserve the temporal continuity and local correlations of the original signals. In addition, we propose an accelerated synergistic correlation coefficient (ASCC) to adaptively assign connection weights in the graph, with the primary goal of robustly capturing signal correlations in the presence of noise and outliers, particularly in low-SNR environments. ASCC combines the strengths of cross correlation and the Pearson correlation coefficient (PCC) to enhance noise resilience while maintaining the ability to capture both temporal shifts and linear dependencies. The ARGAT framework incorporates improved graph convolution, graph attention, and transposed graph convolution layers to capture both local structural information and global feature relationships. Experimental results demonstrate that ARGAT achieves over 91% classification accuracy at an SNR of −14 dB, significantly outperforming state-of-the-art models such as ResNet and DenseNet. Moreover, ASCC consistently outperforms PCC, particularly in severely low-SNR environments, maintaining higher accuracy across all tested SNR levels.