The Development of Advanced Deep Learning-Based EoR Signal Separation Techniques

The weak EoR signal is submerged in the strong foreground radiation interference. The classic the foreground removal methods assume that the foreground spectrum is smooth, but the complex instrumental effect will affect the spectral structure, resulting in the failure to accurately detect the signal. In this paper, a deep learning network called CNN-LSTM model is proposed to separate the EoR signal to improve computing resource utilization for massive observational data. Based on the simulation data of SKA1-LOW, a CNN-LSTM fusion model is constructed to reconstruct the EoR signal. Experimental results show that compared with the traditional methods including polynomial fitting and continuous wavelet transform, the EoR signals detected by the proposed deep learning model have better quantitative evaluation indexes of SNR and Pearson correlation coefficient. This property provides a new way to explore the research field of EoR.