Abstract: To improve the fault tolerance capability and decrease the fault tolerance overhead, a self-adaptive backup strategy suitable for cloud training system is proposed. The connotation and fault tolerance architecture of cloud training are analyzed. Three problems of the self-adaptive backup strategy are analyzed: node select, the number of backups and the location distribution. The node activity degree is introduced to evaluate the backup demands of nodes. The backup number is obtained combined with fault tolerance demands. By weighted ascending matching select, the location distribution is determined. Fault tolerance degree is introduced to the experiments. The results show that the proposed self-adaptive backup fault tolerance strategy can effectively improve the fault tolerance capability and decrease the overhead of fault tolerance.