本项目面向X射线脉冲星新星搜索与计时观测前沿科学领域，开创性地基于高阶循环统计方法进行特征提取，结合在线深度神经网络学习，解决有限时间下极低信噪比X射线脉冲星观测信号的特征表示与特征感知关键科学问题。（1）提出基于循环双谱的脉冲星信号固有特征表示方法和基于二维循环轮廓图谱的脉冲星信号泛化特征表示方法，分别从时域和频域对脉冲星信号进行特征提取，该特征具有明显的抗噪抗干扰能力及良好的相位信息保持能力，从本质上降低了在轨应用时对信号观测时长的要求。（2）提出将时域二维循环轮廓图谱与循环神经网络相结合，频域循环双谱与卷积神经网络相结合，研究网络模型的设计及超参数优化方法，解决脉冲星在轨观测信号的快速识别与相位估计关键技术问题，提高处理的速度和精度。（3）提出基于对冲反向传播的在线深度学习策略，解决大空间范围、长时间跨度下的脉冲星在轨观测信号识别与相位估计动态更新不确定问题，提高在轨应用的稳定性。

This project is aimed at the frontier scientific fields of pulsar timing observation and new pulsar searching, used creatively the high-order cyclic statistical methods with online deep neural network learning to solve the key scientific problems of feature representation and feature perception of very low S/N X-ray pulsar signals under finite time. (1) It is proposed the intrinsic frequency domain feature based on cyclic bispectrum and the generalized time domain feature based on two-dimensional cyclic profile of pulsar signals, which have obvious anti-noise and anti-interference ability and good phase information preservation capability to essentially reduce the requirement for signal observation time in orbit application.(2) It is proposed the methods of time domain two-dimensional cyclic profile with RNN, frequency domain bispectrum combined with convolutional neural network to study the network model design, hyper-parameter optimization, in order to solve the key technical issues of fast recognition and phase estimation of pulsars, and improve the processing speed and accuracy.(3) It is proposed an online deep learning strategy for pulsar signal with hedging back propagation, which can solve dynamically updating uncertainty problem of in-orbit observation signal recognition and phase estimation under large spatial range and long-term span, and improve the stability of on-orbit applications.