Feature selection is an important problem for fault diagnosis systems. This paper proposes a selection method of characteristic parameters based on mutual information theory and distance measurement between parameters. Its core goal is to search the parameter subset which has higher relevance between the selection and state class, the lower redundancy and the longer distance measurement between the parameters. The experiment compared the new algorithm with traditional methods by using Particle Swarm Optimization. The results prove that the new method leads to improvement on feature selection and classification accuracy.