Abstract: Aiming at the requirement of high speed and precision in blast furnace fault diagnosis systems, a new strategy based on global optimization least-squares support vector machines (LS-SVM) was proposed to solve this problem. Firstly, the variable metric discrete particle swarm optimization algorithm was employed to optimize the feature selection and LS-SVM parameters. Secondly, the feature vector was compressed by kernel principal component analysis. Finally, the heuristic error correcting output codes were constructed on the basis of Fisher linear discriminate rate. In the fault diagnosis scheme, fewer LS-SVM classifiers were applied through meaningful partitions and recombination of fault training samples. Simulation results show that the proposed fault diagnosis method can not only improve the fault detection accurate rate, but also enhance the timeliness of the entire system.