F-LLE algorithm and its application in fault feature extraction

The traditional methods to extract the faults features have two disadvantages. It is that the some important information may be lost in the feature selection. And to implement the effective faults identification is very difficult in a high dimensional space describing the faults status. This is because not only the high-dimensional data is generally linear inseparable, but a complex hyper surface or hyper curvature among them will be appeared. Aimed to the bad situation, a new method to combine the Floyd algorithm with the Local Linear Embedding (for short, F-LLE) was proposed in this paper. It was used to reduce the dimension of faults data to a special rotor system. Due to the reduced characteristics is the result of original data to be combined according to some certain rules, the method was used to reduce the dimensions of original feature space. The sensitive features with low dimension to improve the accuracy of faults classification were extracted out by the process. The original feature space of rotor faults is composed of time and frequency domains information. The Floyd algorithm has the ability to depict specifically the original feature space. And the LLE algorithm can keep the tendency invariant of original data set during the dimensionality reduction. So the sensitive characteristics with low dimensions can be extracted effectively out by the process. The results show that the final features extracted with the method are more suitable for the clustering and classification applications of faults data.