Fault feature extraction based on improved locally linear embedding

The performance of locally linear embedding (LLE) for fault feature extraction is influenced by noise, embedding dimension and neighborhood size. Here, it was improved with a new estimation model of weight coefficients and a new estimation model of neighborhood sizes and embedding dimension. Cross-correntropy was used to replace Euclidean distance to measure similarity of vectors. An estimation model of weight coefficients was created based on cross-correntropy. At the same time, the model was simplified with Lagrange method to overcome computation difficulties. The model of weight coefficients based on cross-correntropy improved the performance of LLE and reduced the influence of noise on fault feature extraction. Ncut criterion was employed to choose neighborhood sizes and embedding dimension. A model for choosing their parameters in an automatic way was created. The improved LLE was employed in fault feature extraction of rolling bearings. The test results for fault diagnosis of rolling ball bearings showed that compared with other approaches, the proposed approach is more effective to extract fault features from vibration signals of rolling bearings and to enhance the classification of failure patterns. ©, 2015, Chinese Vibration Engineering Society. All right reserved.