To study and predict the variation on the nonlinear temperature gradient of CRTSII track slab with time, for achieving early warning and reducing the occurrence of track slab diseases, this paper adopted the C-C method to reconstruct the optimal phase space of nonlinear temperature gradient time series of CRTSII track slab. On the basis of phase space reconstruction, NARX dynamic neural network with feedback and memory function was applied to predict and analyze the nonlinear temperature gradient time series. The results showed that according to the optimal phase space reconstruction of the temperature gradient time series of CRTSII track slab, the NARX dynamic neural network method was used to predict the temperature gradient inside the CRTSII track slab in the future time T=30 酌*tau，(min)，酌=1，2，3，…， T=30 酌*tau，(min)，酌=1，2，3，…. The temperature gradient prediction results of 11 sampling time nodes from 2016/11/17/8:00 to 2016/11/17/13:00 (酌=1) showed that the part of the predicted values was in good agreement with the real values. The study provides certain scientific and practical values in the prediction of CRTSII track slab nonlinear time series system.