Abstract: In order to extract fault features of rolling bearing precisely and steadily, a method which is based on locality and globality preserving embedding was proposed for fault diagnosis using K-nearest neighbor algorithm(KNN). First of all, constructing the original feature space with the time domain indexes and the frequency domain indexes of vibration signal. By using this method to select features, fault sensitive feature vectors are obtained. The method comprehensively considers the local structure and global structure of the data, so as to avoid the loss of data and information in the dimension reduction process, and the orthogonal constraint is introduced to reduce the redundancy of information to enhance the fault feature. Then, the K-nearest neighbor (KNN) method is used as a fault feature classifier to recognize different fault types of a rolling bearing. By comparing with other typical feature extraction methods, this approach can more effectively extract the sensitive characteristic vector of four state of rolling bearing, and exhibits better classification performance in the fault diagnosis, the overall classification accuracy is still maintained 100%. Therefore, the method proposed can extract the fault features accurately and stably, providing a good reference for the actual rolling bearing intelligent fault diagnosis.

Key words: fault diagnosis, feature extraction, orthogonal constraint, structure keeping, rolling bearing