Research On Deep Feature Representation And Classification Of Modulated Signals

As an important module in the electromagnetic spectrum sensing system,automatic modulation classification plays a crucial role in the field of non-cooperative communication.With the increasing complexity of the communication environment,the more complex modulation and the increasing density of communication signals have brought challenges to automatic modulation classification.How to obtain the effective feature representation of the signal is of great significance to identify the modulation type quickly and effectively.The traditional methods mainly accomplish modulation classification by extracting hand-crafted features,which are cumbersome and require the prior knowledge.As a result,low recognition efficiency is formed.Deep learning can extract highly identifiable features from large-scale data,and its combination with automatic modulation classification is an effective means to improve the robustness of classification models.In this paper,based on deep learning theories and time-frequency feature representation,the automatic modulation classification models are constructed to improve the adaptability and classification accuracy of the algorithm model.The main innovation achievements of the paper include two aspects:(1)Aiming at the squeezed time-frequency representation of modulated signals under the condition of low signal-to-noise ratios and the extraction of time-frequency spatial features,a modulation classification algorithm based on two-dimensional squeezing transform and convolutional neural network is proposed.On the basis of wavelet transform and synchrosqueezing transform,the algorithm presents a new time-frequency feature representation—two-dimensional squeezing transform,which retains the instantaneous characteristics of the signal and suppresses the noise.Then,in order to obtain the spatial features of the squeezed time-frequency plane,a convolutional neural network is built based on the VGG-16 network,which is adapted to the feature representation of the twodimensional squeezing transform and fully exploits the hidden features of the signal.At the same time,the proposed algorithm effectively improves the recognition performance under low signal-to-noise ratios.(2)Aiming at the situation of unstable signal feature extraction under the condition of short data,the signal hopping features are modeled on the basis of deep feature analysis,and a deep hopping capture model for automatic modulation classification is constructed.The hopping transform unit of the model is built based on S-transform and differential operation,which effectively represents the transient variation of the signal.The deep feature classification module of the model builds a hopping feature perception model based on the Bi-LSTM,which is used to capture the deep hopping information in short data samples and recognizes the modulated signals.Under the condition of short data,the model represents and captures modulation types from the point of hopping feature contained in the signals,which improves the classification accuracy of short data signals and is more suitable for the engineering applications of the actual spectrum sensing system.