Research On Acquisition Method Of OFDM Waveform Passive Radar Reference Signal Based On End-to-end Deep Network

With the widespread deployment of Orthogonal Frequency Division Multiplexing(OFDM)waveform signals worldwide,passive radar based on OFDM waveforms has gradually become a research hotspot in recent years.However,unlike the active radar,the OFDM waveform passive radar needs to extract the direct wave contaminated by multipath and noise in the reference channel at the receiving end as the reference signal.Combining the waveform characteristics of the OFDM signal,the reference signal extraction method based on "demodulation-remodulation" reconstruction is a common way to solve this problem.Traditional reconstruction methods mainly use least squares(LS)and minimum mean square error(MMSE)channel estimation methods that rely on the channel information at the pilot,but the LS method ignores the influence of channel noise,the MMSE method is complicated to calculate and needs to know the channel.Prior Information.With the wide application of deep learning in the field of wireless communication,its strong feature extraction ability,high parameter flexibility,and good generalization performance have led researchers to develop many methods based on neural network models to solve OFDM channels estimate problem in wireless transmission.In this context,based on the end-to-end idea,this paper uses the multi-parameter cascade structure and nonlinear characteristics of the deep network model to establish a deep model that can deeply mine the complex mapping relationship between the reference channel time-domain received symbol and the transmitted symbol.Perform OFDM demodulation,channel estimation,channel equalization and constellation point inverse mapping in the traditional method to directly restore the transmission data and reduce the computational complexity.Taking the bit error rate as a criterion,this paper focuses on the reconstruction performance of the reference signal reconstruction method based on the deep network model under ideal conditions and under the influence of different waveform parameters,and compares it with the traditional reference signal reconstruction method based on LS and MMSE on reference signal reconstruction performance.The specific research content is as follows:(1)The problem of acquiring the sample set of network training and testing required by the deep network model method is studied,including the acquisition of time-domain receiving symbols for network input and transmission symbols as network training tags.A channel model is established to obtain channel state information,and a binary symbol sequence is randomly generated as a label.Combined with OFDM modulation technology,an input sample model from the transmission symbol to the receiving end of the time domain signal is built under ideal conditions.On this basis,the time-domain received data generation method under different conditions considering the influence of actual system parameters is analyzed,and the sample set under the influence of specific waveform parameters is obtained.(2)The deep neural network(DNN)based on the multi-layer cascade structure is studied,and a DNN reference signal reconstruction method is established that uses network learning to adaptively mine the mapping relationship between received symbols in the time domain and transmitted symbols.Simulation verified its feasibility;then,based on the bit error rate,the reference signal acquisition accuracy of the proposed method and the traditional method with the influence of ideal conditions and different waveform parameters are quantitatively compared.(3)The Wasserstein Generative Adversarial Network(WGAN)based on the fully connected structure neural network is studied,and a WGAN reconstruction scheme combining OFDM demodulation,channel estimation,channel equalization and constellation point inverse mapping is established,and the simulation verifies its feasibility.Then the reference signal reconstruction method based on WGAN and the reference signal reconstruction performance of the above three methods with the influence of ideal conditions and different waveform parameters are analyzed.