Abstract: Fault diagnosis method based on orthogonal semi-supervised local Fisher discriminant analysis(OSELF) is proposed, aiming to solve the problem of inadequate number of labeled fault samples and high dimensionality of the feature set. A new dimensionality reduction method named OSELF is proposed combining orthogonal iteration algorithm with semi-supervised local Fisher discriminant analysis(SELF), which can effectively utilize the fault information supported by the labeled and unlabeled fault samples to embed the fault samples into the low-dimensional subspace without the over-fitted problem. The basis vectors of the orthogonal projection matrix are statically uncorrelated, and the discriminations of the obtained low-dimensional fault feature vectors are improved. Then the low-dimensional fault samples are fed into coarse to fine k nearest neighbor classifier(CFKNNC) to recognize the fault type. The proposed method integrated the advantages of OSELF in dimension reduction and CFKNNC in pattern recognition and effectively improved the accuracy of fault diagnosis. The validity of the proposed method is verified by the instance of the fault diagnosis of a gearbox.

Key words: fault diagnosis;dimensionality reduction;orthogonal semi-supervised local Fisher discriminant analysis;coarse to fine- k nearest neighbor classifier