Research On Deep Learning Based Single-channel Blind Separation Of Co-frequency Modulated Signals

Single-channel blind separation of co-frequency mixed signals is widely used in Paired Carrier Multiple Access(PCMA) non-cooperative communication.For overlapped signals received by a single channel,it is difficult for non-cooperative third parties to construct positive definite conditions for signal separation,and it is impossible to carry out effective information demodulation.Studying how to achieve single-channel blind separation of signals is of great significance for improving the ability of reconnaissance of communication signals.In this paper,a deep learning method is introduced in the study of blind separation,which reduces the complexity and has stronger adaptability.It can get better performance than traditional schemes in multiple time-varying channels.Aiming at the time-invariant channel scenario,this paper proposes a blind separation scheme for co-frequency overlapped signals based on bidirectional recurrent neural network.This scheme can recover two original information bits from the received signal after network training,and considering the inter-symbol correlation caused by the channel memory,it avoids the traversal search for the symbols,and achieves a compromise between performance and complexity.For continuous received signals,a block processing strategy is proposed to solve the problem of high error rate at the beginning and end of each data block.When the equivalent channel memory length is large,compared with the traditional Per-Survivor Processing(PSP)scheme,the proposed scheme achieves better demodulation performance and lower computational complexity in both distortion-free and nonlinear distortion channels.In addition,when the signal amplitude does not match in the training phase and the test phase,the scheme also has a certain generalization ability,so it can adapt to the scenario where the channel response changes.For the time-varying channel scenario,this paper proposes a single-channel blind separation scheme for co-frequency overlapped signals combined with deep learning channel interpolation estimation.In order to estimate the time-varying channel,a neural network-assisted interpolation estimation algorithm is proposed,which can learn and mine the time-varying characteristics of the channel.After training,accurate channel interpolation estimates are obtained from the initial channel estimates at the pilot.At the same time,the algorithm is integrated into the blind separation algorithm,which achieves better estimation and demodulation performance than the traditional linear interpolation algorithm.In addition,the scheme can adapt to changes in channe l statistical characteristics and has strong robustness.