Deep Learning Based Communication Signal Recognition Technologies

With the development of the next generation mobile communication network and mobile Internet technology,the future wireless communication network will be confronted with the problem of the dynamic coexistence of heterogeneous network and complex wireless signal on the limited spectrum resources,the development of new spectrum sensing technologies that using signal detection and processing to obtain spectrum usage information in wireless networks is particularly important.Traditional communication signal recognition technologies that rely on complex manual analysis for feature extraction cannot meet the requirements above,therefore,this paper proposes to construct a feature-learning network and signal recognition/detection network based on deep learning algorithms,and improves the abilities of autonomous learning,decision-making and renovate of cognitive communication terminals in complex environments of future wireless communication.Inspired by the original intention of attention based model in the field of image processing,tart from time-frequency analysis methods,this paper proposed some deep feature-learning networks and communication signal recognition networks by introducing the mechanisms of short time Fourier transform(STFT)and discrete Wavelet transform(DWT),in order to improve the adaptabilities of the feature extraction algorithms and the accuracies of the signal recognition networks.The main works of this paper are as follows.Aiming at the limitations of traditional signal recognition algorithms,such as the limited application scenarios and the dependency of artificial analysis,a self-learning algorithm based on deep learning network for communication signal features is proposed.By introducing the two different computation mechanisms of STFT,this paper proposed two unsupervised feature-learning networks based on convolutional Restrict Boltzmann Machine(CRBM)and Restrict Boltzmann Machine(RBM)respectively for communication signal,in which the network based on RBM is greatly improved in computing performance compared to the network based on CRBM,and greatly reduces the requirement for high-performance hardware in deep learning networks.To fix the problem of fixed resolutions in learning-to-STFT networks,this paper also proposed a discrete wavelet transform mechanism based communication signal deep feature learning network that detects mutations and suppresses noise.Two different deep learning networks-direct and indirect,have been constructed based on the differences in feature calculation methods.For signal modulation recognition and signal detection problems in communications,the paper constructed two modulation recognition networks using the learning-to-STFT networks and the Back Propagation Neural Network(BPNN)classifier,and a signal detection network using the learning-to-DWT networks and BPNN.Compared with the traditional algorithms,the algorithms proposed in the paper all obtained better performance in the recognition accuracy,especially under the condition of low SNR.