Ultra-Wideband Channel Classification And Modulation Recognition Based On Compressed Sensing And Deep Learning

With the continuous development of artificial intelligence and wireless communication,intelligent communication is regarded as the development direction of wireless communication in the future.How to better integrate deep learning with wireless communication has attracted the attention of the communication field.As a new wireless communication technology,Ultra-Wide Band(UWB)technology has application prospects in many fields such as data transmission,range-based localization and radar imaging.In this thesis,UWB channel classification and modulation mode recognition are studied,and feasible deep learning algorithms are explored and experimentally verified.Data acquisition is the prerequisite for deep learning based UWB channel classification and modulation mode recognition.However,the ultra-high bandwidth of UWB signals poses a great challenge to data acquisition methods based on traditional sampling theory.Compressed sensing is an effective method to realize low rate acquisition.This thesis studies solutions combining compressed sensing and deep learning,and the specific research content and achievements are reflected in the following aspects:1.In order to improve the accuracy of UWB channel classification in low Signal-Noise Ratio(SNR)region,the channel classifier architecture based on the combination of compressed sensing and deep learning is deeply studied.Comprehensive simulation experiments are carried out to analyze the effects of observation matrix,deep network model and SNR distribution of training data on classification accuracy.Experiment results suggest that:(1)Random Bernoulli matrix has the advantages of simplicity in calculation and small storage requirement,which is beneficial to reduce the implementation complexity;(2)Compared with convolutional networks,fully connected networks are more advantageous to extract channel details and achieve better classification accuracy;(3)The SNR distribution in the process of training data augmentation is the key factor affecting the channel classification accuracy of low SNR region.The channel classification accuracy of low SNR region can be greatly improved by reasonably controlling the SNR distribution range of training samples.2.In order to reduce the computational complexity of channel classifier and improve the classification accuracy of low SNR region,an improved channel classifier architecture is proposed.The new classifier architecture can significantly reduce the forward calculation times of deep network model while improving SNR by introducing average denoising module to suppress the inherent noise folding effect of compressed sensing.The average denoising module introduces a new degree of freedom for the classifier,which can be adjusted jointly with the decision level fusion module according to the environmental conditions to optimize the classification accuracy and computational complexity.Simulation results show that the improved channel classifier architecture can effectively reduce the time complexity and significantly improve the channel classification accuracy of low SNR region.3.Aiming at recognition of UWB modulation mode,an fusion architecture combining multi-symbol period recognizer and decision layer is firstly designed based on traditional data acquisition method and convolutional neural network,and the feasibility of the design scheme is verified by simulation experiments.In order to further reduce the complexity and improve the recognition speed,a modulation recognizer scheme combining compressed sensing and deep learning is proposed.Comprehensive simulation experiments are carried out to compare the recognition accuracy,model complexity and calculation speed of the two schemes.The simulation results show that the modulation recognizer based on the combination of compressed sensing and deep learning has significant advantages in model complexity and computational speed,and can obtain satisfactory recognition accuracy...