The Research On Demodulation For Digital Signal Based On Deep Learning

With the continuous development of digital communication technologies,the increase in channel services and transmission rates in emerging industries places higher demands on demodulation systems.Modern modulation technology and demodulation technology are the basis of communication systems.In order to cope with the increasingly serious nonlinear characteristic transmission channel,the demodulation system usually estimates and equalizes the channel in addition to signal detection.Under complex wireless channels,the multi-level neural network model exhibits excellent time-varying channel adaptive performance,effectively compensating for the distortion of digital modulated signals.In terms of decision performance,deep learning still has sufficient capacity to extract a large number of signals first.Check the information and set the threshold to a theoretical minimum loss value.Aiming at the problem that the existing digital demodulation system has difficulty in detecting complex background and fading channel and limited nonlinear fitting ability,this thesis introduces the signal processing technology of multi-level neural network structure into the field of digital demodulation,and designs corresponding corresponding Equilibrium system and decision system.In order to accelerate the convergence speed of multi-layer perceptron,an improved backpropagation algorithm is proposed,and the adaptive ability of the equalizer is tested by Stanford University temporary channel.The performance analysis of the established adaptive decision feedback equalizer is based on the Euclidean distance expansion analysis of the constellation diagram,and the constellation blur caused by the Doppler shift and the 2-path model is corrected.In the discrete-model multipath channel,the multilayer perceptron-based equalizer has better stability than the linear structure.In the simulation analysis,the digital modulated signal is transmitted through the modeled multipath channel,and the results are compared.It is found that under the nonlinear condition,the decision feedback equalizer is compared with the adaptive linear least mean square algorithm,and the multilayer perceptron structured decision feedback equalizer has better mean square error and robust performance.Finally,in order to improve the demodulation performance of the signal under the frequency selective fading channel,a stable convolutional neural network decision model is proposed.The input of the convolutional neural network is a one-dimensional sampling sequence in the time domain.By constructing two parallel convolutional layers,the independent feature extraction is performed on the information of the in-phase component and the orthogonal component respectively,and the fully connected layer of the latter layer is applied to the modulated signal.Features are detected.The simulation results show that the deconvolution scheme of convolutional neural network with dual symbol combined input decision effectively utilizes the big data feature of inter-symbol interference.Under low signal-to-noise ratio,the normalized channel estimation error is 5%.In demodulation,the 16 quadrature amplitude modulation signal's bit error rate is improved by nearly 6 decibel.