Aiming at how to effectively monitor the network abnormal events, find the bottleneck of network performance and potential point of failure in large-scale data center network, based on the deep analysis of the characteristics of InfiniBand (IB) network and introducing the feature selection strategy and incremental learning strategy, an incremental learning method of fault diagnosis for large-scale IB network (IL_Bayes) which based on the Bayes classification and added incremental learning mechanism was proposed. It could effectively improve the accuracy of fault classification. Through testing and verifying the diagnostic accuracy and the rate of misdiagnosis of this method in the Tianhe-2's real network environment, the result shows that the IL_Bayes method has higher classification accuracy and lower misdiagnosis rate.