Topology-Constrained Multi-Branch Neural Architecture Design for Robust Radio Signal Classification Under Distribution and Noise Shift

Abstract

A topology-constrained neural architecture is developed for radio signal classification from raw in-phase and quadrature sequences under simultaneous noise variation and waveform shift. The model is searched within a mixed operator space that combines short and long temporal convolutions, grouped cross-channel mixers, dilated residual transitions, and a covariance-preserving classification head. The resulting network, named TMR-Net, is trained with architecture sparsification, routing entropy control, spectral consistency regularization, and calibration-aware supervision, allowing the search procedure to trade classification accuracy against computational cost without collapsing to trivial wide branches. Experiments are conducted on two synthetic yet physically perturbed radio corpora and one cross-domain adaptation setting containing timing offset, carrier frequency offset, pulse-shape mismatch, nonlinearity, and symbol-rate variation. The discovered architecture reaches 95.2\% mean accuracy over the 0 to 20 dB interval on the primary 11-class task, 88.1\% over the full SNR range, and 82.9\% on the shifted 18-class setting after adaptation with 5\% labeled target data. Relative to strong residual, recurrent-convolutional, separable, and transformer-like baselines, the architecture improves low-SNR macro-F1 by 2.4 to 4.1 points while using 19\% fewer multiply-accumulate operations than the best convolutional competitor. Ablation analysis shows that heterogeneous receptive fields contribute most to in-domain accuracy, whereas covariance pooling and routing sparsity are more influential under distribution shift. Search outcomes are stable across five retraining seeds and repeatedly favor moderate depth, nonuniform branch widths, and early-stage large receptive fields. These findings indicate that radio signal classification benefits more from topology allocation and second-order feature preservation than from indiscriminately increasing depth or channel count.