基于稀疏自编码器的混合信号符号检测研究

基于深度神经网络(DNN)的符号检测器(SD)的结构直接影响检测精度和计算复杂度，然而，已有的工作中并未对DNN符号检测器的结构选择方法开展研究。此外，已知的基于DNN的符号检测器复杂度较高且仅能完成单一调制信号的检测。针对以上问题，该文提出基于误符号率(SER)度量的低复杂度稀疏自编码器符号检测器(SAED)结构选择策略，同时，利用提出的累积量和矩特征向量(CMFV)实现了对混合信号的检测。所设计的符号检测器不依赖信道模型和噪声假设，对不同调制方式的信号具有较好的检测性能。仿真结果表明，该文设计的SAE符号检测器的SER性能接近最大似然(ML)检测理论值，且在频偏、相偏和有限训练样本等非理想条件下具有较强的鲁棒性。

Research on Symbol Detection of Mixed Signals Based on Sparse AutoEncoder Detector

The architecture of Deep Neural Network (DNN) based detectors can affect the Symbol Detection (SD) accuracy and computational complexity. However, most of the works ignore the architecture selection method when establishing a DNN-based symbol detector. Moreover, the existing DNN detectors use complex architectures and only perform single-type modulated symbols detection. The Symbol Error Rate (SER) based strategy is proposed to design a low complexity Sparse AutoEncoder Detector (SAED) to tackle this problem. Furthermore, a Cumulant and Moment Feature Vector (CMFV)-based method is introduced for mixed symbols detection. Also, the designed symbol detector does not rely on a comprehensive knowledge of channel models and parameters but has the capability to detect various modulation signals. Simulation results show that the SER performance of the SAE symbol detector is close to the values of the Maximum Likelihood (ML) detection approach and provides a stable performance against phase offsets, frequency offsets, and under a limited training dataset.