Since fault diagnosis of blast furnace is very important in manufacturing, the prediction system is inefficient relatively. In this paper, a new strategy based on improved binary tree is proposed to solve diagnosis problem in blast furnace. According to the relations of categories in multi-class problem, it is needless to distinguish all the sorts. In order to improve classification efficiency, we take out the flimsy relatively support vectors in the proceeding of identifying, and then construct a new binary tree without flimsy branches by defining similarities between every two sorts. Compared with different multi-class classification algorithm, the simulation results show this algorithm keeps testing accuracy and proves better performance on identification efficiency and speed.