Improved CLDNN Signal Modulation Recognition Based on Feature Fusion

Aiming at the shortcomings of the existing communication signal modulation identification technology methods such as few identification types and low overall identification rate, this paper proposes an improved Convolutional Long and Short-term Deep Neural Network (CLDNN) signal modulation identification based on feature fusion. First, feature extraction is performed by Convolutional Neural Netwoek (CNN), and an attention module is introduced to obtain low-level features of the effective part. Then, after Long Short Term Memory (LSTM) feature extraction, it enters Atrous Apatial Pyramid Pooling (ASPP) for deep feature extraction. Finally, the superposition and fusion of low-level features and deep features are performed to complete feature extraction and mapping. The network can realize automatic extraction of signal features and classification of modulated signal types. Experiments show that the method in this paper can also effectively identify the communication signal category under low signal-to-noise ratio, and the overall recognition rate is over 94%.