Deep Learning for Joint Noncoherent Differential Detection and Channel Decoding

Abstract

We propose a deep-learning approach for the joint noncoherent differential detection and channel decoding problem. Conventional receivers adopt a model-based approach for noncoherent differential detection and channel decoding in linear manners. For the noncoherent differential detection, the conventional way is to use symbol-by-symbol differential detection (DD) and multiple symbol differential detection (MSDD). DD is simple to implement, but it suffers from performance degradation compared with coherent detection. The performance degradation can be narrowed by MSDD that detects a block of symbols jointly. However, its complexity increases exponentially with the block size. Furthermore, MSDD is not robust in that when the channel phase changes substantively within the block, its performance could be worse than that of DD. Channel decoding needs to be determined according to the channel coding. This work applies the advantages in deep learning for the design of receivers. In particular, we employ a Deep neural network (DNN) constructed by Long Short-Term Memory (LSTM) units to solve the joint noncoherent differential detection and channel decoding problem. Our simulations show that a DNN can outperform conventional model-based linear receivers. Furthermore, it does so with faster signal processing. This performance improvement points to a new direction for future receiver design.