In the load change test, the vibration signals are measured on a two-level gearbox testbed. For coping with non-stationary signals effective, wavelet packet is used for obtaining a feature vector (values of condition attribute). For coping with noise data effective, variable precision rough set (VPRS) theory is introduced in the fault diagnosis of the gear, and a kind method of condition attribute reduction is presented. Firstly, continuous attributes are discretized.Then, the set of M is defined, and varying values of β(right ratio of classification) are selected according to the reality, and the quality of approximation classification of VPRS is utilized to carry through condition attribute reduction and contrasted its result with the reduction result coming from data (including noise data). Finally, a practical example in the gear shows the validity and practicability of the algorithm