Much inportance has been attachad to the selection of optimal feature parameters subset in the fault diagnosis fields such as liquid rocket propulsion system. This paper presents an effective method of selection for the optimal feature parameters subset using Genetic Algorithms and based on the maximum and minimum clustering criterion for samples, so that the selected feature parameters subset can be used to compose a simplified real-time fault classifier with high robustness to various sorts of noises and distur-bances. First, a composite directional divergence index for samples is proposed as an evaluation criterion for the selected feature parameters subset for fault diagnosis 'purpose; then, Genetic Algorithm has been modified in parts for this specific permutation problem, the dynamic fitness adaptation technique and all-sharing function are introduced in order to avoid the population's premature convergence. An ad-hoc genetic operator is specially designed to improve the feature selection efficiency. In an addition, all the selection procedures for the optimal feature parameters subset are based on the data set for 16 sorts of common faults simulated for a type of liquid rocket engine system. The numerical experiments show that this selection algorithm is highly effective and the constructed fault classifier with the selected feature parameters possesses morerobustness.