The relationship between reliability and original random variables is discussed for the design of the reliability sensitivity of a complex compressor rotor system. First, by introducing the basic concepts of isolative intervals and amplification factors, the actual and permitted isolative intervals can be calculated by the dynamic response of the rotor system. Then, the vibration reliability model for the compressor system is established according to the rule that permitted isolative intervals cannot exceed the actual ones. Second, the strong non-linear relationship between the reliability of a complex structural model and original random parameters is obtained by an artificial neural network instead of the finite element (FE) model. Finally, a geared rotor system is taken as an example. The reliability at a dangerous position is calculated based on the reliability design theory, and the reliability sensitivity of the mean value and variance of parameters are also obtained. The results show that the system’ s reliability is mainly affected by the working, manufacturing, installation and meshing parameters. In practical designs, the working parameters are dominant, followed by the manufacturing, installation and meshing parameters.