In light of neighborhood selection of a manifold learning algorithm, a novel adaptive neighborhood selection method was proposed. The adjacency matrix was calculated by Mahalanobis distance, then used to obtain the initial neighborhood. The final neighborhood was obtained through regulation of the initial neighborhood using Gauss kernel density estimation of all samples. The improved principal component analysis locality preserving projection (PCA-LPP) manifold learning algorithm using adaptive neighborhood selection was applied in the processing of gearbox fault signals. The results showed that the eigenvectors obtained by adaptive neighborhood selection had a more satisfying fault identification rate than that obtained by k neighborhood and original eigenvectors, reaching as high as 94.67%.