Research On Data Feature Learning-based Reference Signal Extraction Methods For Drm-based Passive Radar

Because the digital amplitude modulation broadcasting(DRM)passive radar uses the third-party signal as the detection source,it is necessary to first solve the problem of reference signal extraction when realizing target detection.To solve this problem,in order to make the reference signal extraction method more simple and efficient,and then provide a new intelligent means for the radar signal processing of this system,this paper uses the powerful data feature learning ability of deep neural network to study the reference signal extraction method based on learning the data features of the reference channel of DRM-based passive radar.The specific research contents are as follows:Generally,reference signal and target echo need to be matched and filtered when digital amplitude modulation broadcasting(DRM)passive radar realizes target detection.Reference signal extraction is one of the key technologies.However,on the one hand,the process of traditional extraction methods is complex,on the other hand,the existing channel estimation algorithms in reference signal extraction have some defects,such as insufficient estimation accuracy or the need for a priori channel statistics.Under this background,this paper uses the powerful data feature learning ability of deep neural network to study the reference signal extraction method based on learning the echo data features of DRM-based passive radar reference channel.This paper will provide a new intelligent method for the radar signal processing of this system.The specific research contents are as follows:(1)Based on the time correlation of the echo data of the reference channel,two end-to-end reference signal extraction methods based on the bidirectional long-short term memory network(Bi LSTM)are studied.The time characteristics between the symbols of the reference signal are automatically mined through the network,and the received signal of the reference channel is directly mapped to the pure reference signal.In this method,the frequency domain reference channel received signal is generated by simulation as the training set and the pollution-free direct wave signal is used as the label to complete the network training,and the bit error rate and mean square error are used as the performance indicators to verify the effectiveness of the network.(2)In order to make effective use of the pilot resources in the DRM signal,by learning the channel response characteristics contained in the pilot in the reference channel echo data,a pilot assisted super-resolution reconstruction channel estimation method is studied to replace the traditional channel estimation module in the reference signal extraction.On this basis,the pure reference signal is obtained through equalization.Combined with the pilot characteristics of DRM waveform,this method studies two super-resolution reconstruction networks based on super-resolution convolutional neural network(SRCNN)and very deep network(VDSR).By treating the channel response estimated by LS as a low-resolution picture,the super-resolution network is used to reconstruct a high-precision channel response,which improves the accuracy of channel response under the same pilot resources as LS method,and then improves the purity of reference signal.(3)Because the DRM-based passive radar has no transmitting signal,it is difficult to obtain enough channel response data for training network.To solve this problem,this paper proposes a training data generation method based on LS time domain channel response waveform feature rough estimation.Firstly,the multipath delay in the measured reference channel is roughly obtained by observing the time domain channel response waveform characteristics estimated by LS.On this basis,the channel parameters in the simulation are set,and then sufficient channel response training data samples are obtained.The simulation data is used to train the network,and finally the simulation and measured data are used to test the network.