A Novel Demodulation Network for Binary Partial Response CPM Signals

Abstract

Continuous phase modulation (CPM) is a promising modulation scheme, due to its constant envelope and compact spectrum. However, the application of CPM is limited by the demodulation and the strict requirements of synchronization. A novel method based on the convolution neural network (CNN) is proposed for binary partial response CPM signals, where the structure of the neural network is designed according to the traditional demodulation processing. Specifically, the convolution kernels are applied to extract the high-dimensional features, which are different from the branch metrics calculated by the matched filters and phase rotation. And then the extracted features are mapped in the fully-connected layers, which plays the same role as the Viterbi decoder. Besides, the moving step of the convolution kernels is small, so that the extracted features can obtain more information than the branch metrics, even though there are some timing errors. Our numerical evaluations demonstrate that the performance of the proposed method approaches that of the theoretical optimal method. Moreover, the designed network is robust to normalized timing variance with no extra training.