Experiments were performed on a high-speed online random neutron analyzing system (HORNA system) with a ²⁵²Cf neutron source (up to 1 GHz sampling rate and 3 input data channel), to obtain time- and frequency- dependent signatures which are sensitive to changes in the composition, fissile mass and configuration of the fissile assembly. The data were acquired by three high-speed synchronized acquisition cards at different detector angles, source-detector distances and block sizes. According to the relationship between ²⁵²Cf source and the ratio of power spectral density, R_psd, all the signatures were calculated and analyzed using correlation and periodogram methods. Based on the results, the simulated autocorrelation functions were utilized for identifying different fissile mass with Elman neural network. The experimental results show that the R_psd almost remains at constant amplitude in frequency range of 0–100 MHz, and is only related to the angle and source-detector distance. The trained Elman neural network is able to distinguish the characteristics of autocorrelation function and identify different fissile mass. The average identification rate reached 90% with high robustness.