Modulation Mode Recognition Method with Multi-feature Fusion under a Complex Noise Floor

Increasing the accuracy of modulation signal recognition in a generalized noise environment is an important problem nowadays. To address this problem, a deep neural network-based modulation mode recognition method FF-net (Feature Fusion net) is proposed in a generalized noise environment. The method firstly extracts and fuses the constellation map features and time domain features of the signal, which increases the diversity of feature extraction; then the Res2Net module and SENet module are used to realize the extraction of channel attention and spatial multi-scale features, which improves the validity and versatility of the feature extraction, and thus improves the recognition accuracy of the modulation mode recognition system in different noise environments. The experimental results show that its average recognition rate reaches 99.5% under the condition of α=0 signal-to-noise ratio of 0~10dB, and 86.1% under the condition of α=2 signal-to-noise ratio of 0~10dB, which is better than other existing algorithms under the same conditions.