A method of fault diagnosis by using orthogonal iterative local fisher discriminant was proposed to better recognize faults of rotor system. Divergence s within and between classes were both redefined on base of local weighted adjac ency matrix, and local fisher discriminant function was established. Then optima l projection vector was found by iterative orthogonal approach and fault diagnos is model was achieved which can be directly used to recognize patterns of increm ental data. The method guarantees minimum reconstruction errors during dimension ality reduction and be free from model reconstruction on the dynamic incremental data in general manifold learning methods. The experimental result shows that t he orthogonal local fisher discriminant (OLFD) algorithm is superior to other ma nifold learning algorithms in rotor fault diagnoses.