Abstract: Based on Manhattan distance as the information measure of similarity between features,a feature selection method based on Manhattan-weighted Discriminative Laplacian Score (M-WDLS) was proposed.It can reduce the difficulty in fault classification identification caused by overlapping interactions of different fault characteristics. In this method, Manhattan distance was used to measure the similarity between high-dimensional feature vectors,and the data sample marker information was integrated into the weight calculation to enhance the discrimination of weight,so as to improve the sensitive feature screening performance of LS algorithm. A rotor fault diagnosis method based on M-WDLS and PCA was proposed by combining M-WDLS with Principal component analysis (PCA). Firstly, the time domain,frequency domain and time frequency domain of the original vibration signal were extracted to construct the feature set of the mixed domain. Then, sensitive features were selected by M-WDLS to form sensitive feature matrix. Finally, the sensitive feature matrix was reduced by PCA and the result was input into k-nearest neighbor classifier (KNN) for pattern recognition. The experimental results showed that this method could effectively extract the vibration signal characteristics of the rotor system and improve the accuracy of fault identification.

Key words: laplacian score(LS), Manhattan distance, discriminant weight function, fault diagnosis