As an improved algorithm of LS-TSVM, RELS-TSVM has a high training speed, good robustness and strong generalization ability. This paper introduces it to the fault diagnosis of transformer and proposes a fault diagnosis model based on RELS-TSVM optimized by CSO algorithm. In the model, it constructs a multi-class classifier for transformer fault type identification by combining binary tree and RELS-TSVM, and uses CSO that has strong search performance to optimize the parameters of RELS-TSVM, which would upgrade the performance of the model to the best state. The example of transformer fault diagnosis based on DGA shows that the fault diagnosis model in this paper is of high accuracy and works better than PSO-SVM model. This work is supported by National Natural Science Foundation of China (No. 51477121).