Research On Deep Learning-based Recognition Algorithm For Digital Modulated Signal

The increasing requirements for massive,high-speed information transmission has greatly promoted the development of cognitive radio domains where efficient channel equalization and modulation recognition in complex electromagnetic environments play critical roles.To recover the signals distorted by rugged communication channel,the traditional equalization methods generally estimate the channel characteristic vector by gradient descent method,conversely convolving this vector with the received signal.The traditional modulation recognition methods extract the expert features of the received signals and utilize a reasonable machine learning model to derive the classification bounds in feature space.In the past few years,an increasing number of advanced convolutional neural network architectures and optimization algorithms have been successively presented,which brings great achievements of deep learning in various fields.Based on the powerful ability of convolutional neural networks to represent the features from original inputs,this dissertation deeply studies its applications to signal recovery and modulation recognition.The main contributions are as follows.1)Aimed at the problem that the current deep learning architectures and optimization algorithms can hardly learn the structural features of complex base-band signal,this dissertation proposes a multi-path convolutional neural network architecture to satisfy the requirements of digital signal processing in complex base band.Each path of this architecture can respectively learn the structural features of the real and imaginary part of the input signal,and fuse the outputs of paths in designed order to obtain the final output,which realizes fitting the complex-valued mapping.2)To address signal distortion caused by communication channel,this dissertation discusses the traditional methods of denoising and channel equalization,proposing a novel end-to-end method based on multi-path convolutional neural network.Moreover,for the validation of feasibility and performance,we conduct the simulated experiments:Firstly add white Gaussian noise and channel impairment to the original signals by MATLAB.Secondly,respectively recover the distorted signals by the traditional methods and the proposed method,and demodulate them.Finally,compute and compare their performances by symbol error rate.The results show that the proposed method has stronger and more comprehensive abilities of signal recovery and achieves the best performance.3)This dissertation proposes a scheme of "signal recovery preprocessing-modulation recognition".At the modulation recognition stage,both multi-path and sequential architectures are exploited to design the convolution neural network.Additionally,four-type features are discussed and analyzed in term of accuracy,including the abstract features obtained by unsupervised convolutional auto-encoder and supervised classification network,the instantaneous feature and high-order statistics of modulated signal.Moreover,the performances of commonly used machine learning methods,e.g.,softmax regression,support vector machine and ensemble learning,on modulation recognition are compared.4)In order to verify the feasibility of the proposed scheme in real environment,this dissertation adopts NI-USRP 2920 as signal transmitter and receiver,establishing the real communication environment to implement the experiments.As the result,the proposed scheme can meet the requirements of modulation recognition.On grounds of several unknown interference(e.g.,the clocks of the transmitter and receiver are not synchronous),the recognition accuracy is slightly lower than simulated results.