Fault diagnosis of blast furnace is a hot topic and has a very important practical significance and value. At the same time, rapid diagnosis of blast furnace fault is a difficult problem. In this paper, a novel strategy based on CLS-SVM is proposed to solve this problem. A modified discrete particle swarm optimization is applied to optimize the feature selection and the LS-SVM parameters. Fitness function considers in detail the training time and the recognition accuracy and the feature selection. The CLS-SVM algorithm is presented to increase the performance of the LS-SVM classifier. The new method can select the best fault features in much shorter time and have fewer support vectors and better generalization performance in the application of fault diagnosis of the blast furnace.