A triple-path hybrid deep learning architecture for robust analog modulation classification via EMD, Gramian angular summation field, and bidirectional GRU

This paper proposes a novel Triple-Path Hybrid Deep Learning Architecture for automatic analog modulation classification (AMC) that simultaneously processes three complementary signal representations: (i) Empirical Mode Decomposition (EMD) statistical features, (ii) Gramian Angular Summation Field (GASF) images processed by a dual-kernel Squeeze-and-Excitation Convolutional Neural Network (SE-CNN), and (iii) raw temporal sequences processed by stacked Bidirectional Gated Recurrent Units (BiGRU). A Cross-Path Attention Gating mechanism learns a static, data-driven combination weight for each path’s contribution; per-SNR analysis (Section “Cross-Path Attention Weight Analysis”) shows this weighting is fixed rather than SNR-adaptive, and Section “Gate Regularisation Comparison” confirms it is not restored by entropy or L2 regularisation. An ensemble of five independently seeded models with zone-adaptive temperature-scaled voting further stabilizes classification. Experiments are carried out for four analog modulation types (AM, PM, FM and SSB) across six SNR levels from − 10 dB to 20 dB. The proposed method achieves a mean classification accuracy of 94.5% across all SNRs, outperforming the EMD-SVM baseline (83.4% mean). Critically, performance at low SNR is substantially improved: 73.7% at − 10 dB (+ 33.7 pp), 93.8% at − 5 dB (+ 25.3 pp), and 99.4% at 0 dB (+ 7.3 pp). At 20 dB, the system achieves 100% classification accuracy. An in-depth ablation study confirms the architecture components contribute additively to performance, with the BiGRU path contributing the single largest gain (− 31.2 pp at − 10 dB when removed) and the ensemble contributing + 22.9 pp over mean single-model accuracy at − 10 dB. Statistical significance testingWilcoxon signed-rank test: p = 0.0000 at all SNR ≤ 0 dB; McNemar statistic = 450.33, p = 0.0000 at − 10 dB (paired correct/incorrect indicator across the four-class task). Under pedestrian and vehicular Doppler-shifted fading, accuracy remains at 76.29–80.38%, with the larger vehicular Doppler spread (70 Hz) producing a 2.80–4.42 pp additional drop relative to the pedestrian case (5 Hz), plausibly attributable to increased sensitivity of the BiGRU temporal path to Doppler-induced phase rotation. Single-sample GPU inference latency is 88.2 ms (unbatched); batching improves this to 0.75 ms/sample at batch = 256, while end-to-end CPU preprocessing (EMD, GASF, and sequence preparation) adds a further 25.97 ms per sample.