Research On Processing Continuous Phase Modulation Based On Deep Learning

Continuous phase modulation(CPM)is a modulation with high spectral efficiency,constant envelope,and fast outside attenuation,which can effectively improve the spectrum utilization and power utilization of the communication system.In recent years,CPM becomes a research hotspot in academia,due to its wide applications,such as satellite communications,telemetry and remote sensing,satellite digital video broadcasting,and so on.However,the high spectral efficiency of the CPM is obtained at the cost of complex reception,and the traditional algorithms with poor adaptability are susceptible to errors.Therefore,in this dissertation,the improved CPM receiving algorithm with better adaptability and higher efficiency is focused,mainly including demodulation,timing recovery,and modulation index estimation algorithms.The main contents are as follows:(1)A low-complexity non-coherent demodulation algorithm is proposed to reduce the complexity of non-coherent demodulation.The frequency truncated pulse and decision feedback are integrated to reduce the number of matched filters and grid states with the aim of realizing low-complexity non-coherent demodulation.Analysis and simulation results show that when the bit error rate is 10^-5,the complexity of the proposed demodulation network is reduced to 1/16 of the traditional multi-symbol differential detection algorithm,at the cost of 0.7d B performance loss.(2)A convolutional demodulation network is designed to tackle the problem that maximum likelihood sequence detection is susceptible to timing errors,where extracted features by sampling points is employed,instead of the branch metrics by symbols,thereby improving the resistance to bit timing errors of demodulation.Analysis and simulation results show that the training network not only achieves the optimal demodulation performance under ideal timing,but also has better performance than that of the traditional method,when the normalized timing variance is greater than 0.0625.(3)A timing estimation algorithm based on neural network is proposed to shorten the synchronization time and to improve timing performance.The maximum branch metrics are adapted to directly estimate time errors,and thus synchronization time is shortened.Analysis and simulation results show that the proposed scheme can not only effectively avoid the accumulation of errors caused by feedback architecture,but also reach the theoretical performance of timing recovery.More importantly,the synchronous symbols required by the proposed timing estimation are less than that of the traditional early-late loop and decision-directed timing recovery,which can achieve faster and more accurate bit synchronization.(4)A blind modulation index estimation scheme that integrating traditional features and high-dimensional features is proposed,to solve the problems of limit application sceneries and poor performance of traditional methods.The proposed scheme further extracts common features from the traditional autocorrelation function,and thus improving the estimation performance of CPM signals shaped by different filters.Analysis and simulation results show that the proposed estimation scheme based on high-dimensional features not only can be applied to CPM signals shaped by multiple non-linear filters,but also has better estimation performance than traditional methods,thereby improving the adaptability of the proposed scheme in practice.