For spiral bevel gears widely used in various fields of industrial engineering, the vibration signal is greatly disturbed by environmental noise. When the fault occurs, the signal exhibits nonlinear, nonstationary characteristics, the fault feature information is weak, the fault feature extraction is difficult, and the diagnostic efficiency is low. Therefore, a spiral bevel gear state recognition method based on MPELPP and ELM is proposed. Firstly, construct multiscale entropy values as the original highdimensional feature vectors, then use LPP to obtain the optimal lowdimensional sensitive feature vectors by reducing the original highdimensional feature vectors, which can mine and preserve the nonlinear structural features of the original highdimensional features. The obtained sensitive feature quantity is input into the ELM for recognition diagnosis. The method is applied to the diagnosis of four kinds of fault state spiral bevel gears under three kinds of speeds, and compared with MPEPCAELM and MPEELM. The results prove the accuracy and superiority of the proposed method.