Modulation Recognition Based on Denoising Bidirectional Recurrent Neural Network

Modulation recognition is an important research area in wireless communication. It is commonly used in both military and civilian domains, such as spectrum detection and interference identification. Most existing modulation recognition algorithms have a better recognition performance at high signal noise ratio (SNR). However, when SNR decreases to - 10 dB or even lower, such as the battlefield and disaster areas and other harsh environment, the recognition rate may decrease dramatically. In order to solve this problem, a modulation recognition algorithm based on denoising bidirectional recurrent neural network is proposed. Firstly, the state memory ability of the signal reconstruction layer in the network is utilized to learn the temporal correlation of the modulated signal, the reconstruction of the received signal is completed and the noise in the modulated signal is suppressed. Then, the reconstructed signal is encoded and decoded by the feature reconstruction layer, in which the feature of reconstructed signal is compressed and reconstructed, thereby the influence of noise can be further reduced. Finally, the reconstructed features are identified and classified by the fully connected layer. Simulation results demonstrate that the proposed network can effectively suppress the noise in the signal. Compared with other existing algorithms, the proposed method has higher recognition accuracy in the low SNR environment.