Artificial neural networks have been widely used in nonlinear analysis in various fields due to their excellent self-regulation and learning ability. Low background liquid scintillator neutron detector in China JinPing underground laboratory (CJPL) have been recording neutron background data, The energy spectrum of detector output is actually the nuclear recoil energy spectrum, which can be in one-to-one correspondence with the input spectrum, and changes as the parameters of the input change. Therefore, the detector output signal can be input into the trained neural network to determine the emission spectrum of the external radiation source. The neural network used in this paper is the Elman neural network, and the data used in the training neural network is simulated by Geant4. The nuclear recoil energy spectrum obtained from the experiment is input into the trained neural network for decomposition, Finally, the Elman network has a spectral error of 0.1%~11.8% for the Am-Be neutron source and 0.1%~8.9% for the 252Cf neutron source.
