A novel method of wayside acoustic fault diagnosis for locomotive bearings is presented to evaluate the bearings status of running train, which mapped the original features extracted from corrected acoustic signals into a high-dimensional feature space through nonlinear mapping. Where, the original features are extracted from bearing′s acoustic signal acquired by the wayside microphone after Doppler distortion correction being conducted. In the high-dimension feature space, new features can be fused through the eigen-decomposition of the fourth-order cumulative kernel matrix. And the fusion features are extracted from acoustic signals of locomotive bearings with normal condition, inner race defect, outer race defect and ball defect respectively, and identified using the support vector machine classifier. The results show that the proposed method is efficient in nonlinear feature extraction from bearing′s acoustic signals and can obtain high accuracy in fault identification.