Deep learning-based signal modulation recognition algorithm in complex electromagnetic environments

Abstract

Automatic modulation recognition (AMR) is a key technology in cognitive radio and intelligent communication systems. Although recurrent neural networks such as LSTM can effectively capture temporal dependencies of signals, they have limited capability in modeling long-range global features.[1]In contrast, Transformers can learn global contextual information through self-attention mechanisms but are less effective in extracting local temporal patterns. To address these limitations, this paper proposes a hybrid modulation recognition network, termed LSTM–Transformer Hybrid Network (LTH-Net).The proposed network adopts a dual-branch architecture. A stacked bidirectional LSTM module is employed to extract temporal evolution features from IQ signals, while a Transformer branch enhanced with positional encoding and multi-head self-attention is used to capture long-range dependencies. A cross-modal attention fusion module is further introduced to adaptively combine the complementary features extracted by the two branches, followed by a fully connected layer for modulation classification. Experiments conducted on the RadioML2016.10A dataset demonstrate that LTH-Net consistently outperforms conventional LSTM, Transformer, and CNN-LSTM models over a wide SNR range. At 0 dB, the proposed method achieves an average recognition accuracy of 86.5%, representing a 6.5% improvement over the baseline LSTM model. Moreover, when the SNR exceeds 10 dB, the recognition accuracy remains above 94%.[2]These results verify that LTH-Net effectively combines temporal feature learning and global dependency modeling, providing a robust and accurate solution for automatic modulation recognition.