Multiple-phase Shift Keying Recognition Based On Convolutional Neural Network

Multiple phase modulation is a common modulation method in wireless communication.It has the advantages of strong anti-interference,high spectrum utilization and simple circuit implementation.It is widely used in satellite communication and civil mobile communication.The recognition of its subclass signals,such as binary phase shift keying signals,quaternary phase shift keying signals and octal phase shift keying signals has always been a hot issue in the field of modulation recognition.Since traditional modulation recognition methods require strong professional knowledge and engineering skills to design the feature extractor,it is difficult to realize automatic extraction of signal features.Therefore,this thesis applies the convolutional neural network in the field of modulation signal recognition,and the characteristics of the multiple phase modulation signals are extracted automatically to achieve the recognition.Firstly,the multiple phase modulation signals required for the experiment were collected using instruments such as signal generators and receivers.Three different datasets suitable for convolutional neural network training are formed by using the original in-phase orthogonal data,time-frequency information and time-phase information of the acquired signals.Secondly,for the original data of multiple phase modulation signals,the in-phase orthogonal data recognition based on convolutional neural network is studied.In order to apply the convolutional neural network to the time series signal,the convolutional neural network structure suitable for image is adapted,and the 4 layers convolutional neural network model CNN-IQ is applied to the in-phase orthogonal data set.CNN-IQ achieves a recognition accuracy of 92.8%.The recognition performance and feature extraction ability of convolutional neural networks are verified by comparison with traditional machine learning algorithms.Then,in order to further improve the recognition accuracy of the multiple phase modulation signal,the spectrogram recognition based on convolutional neural network is studied.The spectrograms of multiple phase modulated signals contain relatively complete key information,but at the same time it has the problem of high similarity and is difficult to be identified by eyes.Therefore,based on the state-of-the-art models of convolutional neural network,a deep network ReSENet is proposed.ReSENet combines the characteristics of ResNeXt and SENet,it can learn the complex abstract features in the data through deep learning and feature redirection,and a recognition accuracy of 99.9% is attained on the spectrogram dataset.Finally,the recognition of multiple phase modulation signals based on convolutional neural networks is studied from the perspective of constellations.The constellation of the actual signal is often distorted due to the inter-symbol interference and the difference of carrier frequencies at the receiving end and the signal transmitting end.The adjacent points clustering and rotation method is proposed to recover the distorted constellation,and the recovered constellation is used as input to the proposed convolutional neural network model CNN-4 for training.The recognition accuracy reaches 99.9%.In addition,the complexity of the ReSENet used in spectrogram recognition and the CNN-4 used in the constellation recognition were calculated,and thus ReSENet and CNN-4 were compared and analyzed.