本文运用分形理论中的关联维数计算方法，对齿轮正常、齿根裂纹和断齿三种不同工况的振动信号进行识别。为了降低噪声对关联维数计算结果的影响，提出EMD滤波方法对采集的信号进行预处理。通过对仿真信号和齿轮箱实测信号的诊断，证明了该方法的可行性和有效性。

Three different conditions of a gear, normal condition, crack in root of tooth, broken-tooth, are diagnosed by calculating the correlation dimension in the fractal theory. In order to weaken the effect of the noise on the precision of calculating , a new approach to filter based on EMD is put forward to pre-conduct the signal sampled. Through experiment with a gear box, this method is proved to be feasible and valid.