To early discover the critical degradation status of rolling bearing and predict its degradation trend, Tdistributed stochastic neighbor embedding (TSNE) sample entropy state degradation feature index and the prediction method based on time convolutional network (TCN) are proposed. Firstly, the lowdimensional manifold features of original vibration signal are extracted by TSNE algorithm. Then, sample entropy of lowdimensional manifold features are calculated as the status degradation. Finally, the degradation trend of the bearing is predicted by the TCN algorithm based on features of historical status degradation. Compared with traditional feature index, experimental results show that the TSNE sample entropy feature index can detect the critical status of the bearing degradation significantly at least 50 minutes ahead of time and the prediction error of the TCN algorithm is only 045%. These results have high application values in engineering.