A Parallel CNN-LSTM Automatic Modulation Recognition Network

Automatic modulation recognition (AMR) is crucial for signal interception and analysis in non-cooperative communication scenarios. To address the challenges of low signal-to-noise ratio (SNR) and model generalizability, this paper proposes a lightweight parallel network architecture that integrates convolutional layers, a channel attention mechanism, residual connections, and long short-term memory (LSTM) units. The model takes in-phase and quadrature (IQ) components of signals as inputs to jointly learn features for modulation scheme identification. Experiments are conducted on the expanded RML dataset to evaluate the model’s performance. Results indicate that the proposed network achieves recognition accuracy comparable to that of deep neural networks while requiring significantly fewer parameters. Furthermore, it demonstrates favorable generalization performance on other datasets, demonstrating its potential for efficient deployment under resource constraints.