在应用神经网络来判断透平在线状态的工作中，面对众多的相关过程参数及环境参数，如何从中为神经网络选择适当的输入参数是非常重要的。根据透平热力性能在线诊断的需要，本文研究了如何用遗传算法为神经网络寻找优化的输入组合以期达到少输入、快训练和准确回忆的目的。结果表明，由遗传算法选定的较少的参数作为诊断网络的输入即可判断透平性能状态的好坏。 In the application of neural network to determine the state of the turbine online, the face of a large number of related process parameters and environmental parameters, how to select the appropriate input parameters for the neural network is very important. According to the needs of on-line diagnosis of thermal performance of turbine, this paper studies how to find the optimal input combination for neural network by using genetic algorithm in order to achieve the goal of less input, fast training and accurate recall. The results show that fewer parameters selected by genetic algorithm can be used as the input to the diagnosis network to judge the state of turbine performance.