LCP-RepNet: A Dual-Domain Reparameterized Network for Robust Automatic Modulation Recognition Under Low SNR Conditions

Automatic modulation recognition (AMR) is an important technology in intelligent wireless communication systems. However, under low signal-to-noise ratio (SNR) conditions, strong noise reduces the separability among modulation classes, making accurate recognition particularly challenging. To address this challenge, this paper proposes LCP-RepNet, an end-to-end network for AMR under low SNR conditions. The Spectral–Temporal Feature Integration (STFI) module applies sample-dependent complex-valued filters generated using low-rank factors to the spectral representations of learned features. It then combines the filtered spectral response with short-term interaction responses constructed between adjacent temporal positions. The Structurally Reparameterized Multi-Scale Aggregation (SRMA) module employs multi-scale temporal convolution branches to aggregate feature responses across different receptive fields. After training, its auxiliary convolution paths are equivalently fused into the corresponding main convolutions. The Adaptive Context-Gated Downsampling (ACGD) module generates input-dependent channel and soft-gating weights to combine paired adjacent temporal positions. Experiments were conducted on three synthetic datasets: RadioML2016.10A, RadioML2016.10B, and RadioML2018.01A. LCP-RepNet achieves overall accuracies of 68.48%, 69.84%, and 64.41%, respectively, with corresponding low SNR accuracies of 47.95%, 50.42%, and 22.82%. Among the models included in the comparisons, LCP-RepNet obtains the highest overall and low SNR accuracies on all three datasets.