Abstract: In this paper, an approach to select feature parameters in fault diagnosis of liquid rocket engine is presented. Based on Shannon's theory on information entropy, theoretical fault chassification information contained in the selected parameters is given, its corresponding engineering calculation formula is then presented as a criterion of feature parameter selection. The modified genetic algorithm is used to select the optimal parameters for the common faults of a liquid rocket engine. The numerical experiment shows that the selected parameters are reasonable and the fault classifier constructed with them is of much less computational complexity and more robust to noises and disturbances.