A method to integrate KSSOMFA and WKNN together on faults identification of rotating machinery

Abstract

Aiming at the problems of the dimension of fault feature data sets of the rotating machinery are too high and labeled information of the sample is insufficient, which lead to fault recognition too difficult. A kind of combinational method of fault identification of rotating machinery based on Kernel Semi-supervised Orthogonal Marginal Fisher Analysis (Kernel Semi - Supervised Orthogonal Marginal Fisher Analysis, KSSOMFA) and Weighted K-Nearest Neighbor (Weighted K-Nearest Neighbor, WKNN) classifier is proposed. Firstly, this method needs to establish multi-domain multi-channel high-dimensional fault feature dataset which can comprehensively reflect the characteristics of different fault information. Then, KSSOMFA algorithm is used to reduce the dimension of the dataset, which extracted the low dimensional essential and good discriminant fault feature subset. After the dimension reduction, making the distance between the same kinds of samples was closer and the distance between heterogeneous was pushed away. At the same time, making the output of based vectors are orthogonal to each other. Finally, the extracted feature set of the lower dimensional nature was fed into WKNN classifier to recognize fault pattern. The advantage of this method is that it can combine the excellent properties of dimension reduction of the KSSOMFA algorithm and high classification accuracy and stable classification decision of the WKNN algorithm, which can improve the accuracy of fault identification. The method was applied in fault feature set of a double span rotor system, and the conditions show that the data driven fault identification technology was able to improve the accuracy of fault identification through data mining, which provides a certain of theoretical reference for fault identification.