Fault samples are of complex diversity, different fault types exist in multi-manifold subspace of different dimensions, which reduce dimension to single manifold of the same dimension can not be more efficient feature extraction. A fault diagnosis method based on multi-local linear embedding (Multi-LLE) was proposed, the single manifold fault diagnosis method was extended to the multi-manifold, Multi-LLE were extraited respectively the essential characteristics of each failure data sets which on its manifolds of the intrinsic dimension, to classify by comparing the distance of cluster centers of each feature vector and embedded vector of new failure samples on different dimensions. To verify the algorithm by making use of rotor test failure data, the results show that Multi-LLE method achieve more effective fault diagnosis from comparison Multi-LLE and LLE, HLLE.