Abstract: A fault diagnosis method based on adaptive Locality Preserving Projection (LPP) feature dimensionality reduction and improved Variable Predictive Model based Class Discriminate (VPMCD) was proposed to solve the issues of high fault feature dimensionality and high time consumption caused by pattern recognition in mechanical system condition monitoring and fault diagnosis. Firstly,time-frequency domain features,energy features,and complexity features were extracted from rolling bearing vibration signals to form a high-dimensional fault feature dataset;secondly,the high-dimensional fault feature set was downgraded by using the adaptive LPP method to obtain the low-dimensional sensitive fault features;lastly,the low-dimensional features were classified and recognized using the improved VPMCD method,and then the type of faults was judged. The analysis of rolling bearing fault diagnosis experiments showed that the adaptive LPP method overcomes the defect of manual pardameter selection in the LPP method,and has less computational time on the basis of obtaining low dimensional sensitive fault features.Compared with methods such as PCA,LTSA,LLTSA,Isomap,LLE,etc,it had obvious advantages;the improved VPMCD method can overcome the contingency and one-sided nature of the artificial selection model,and achieve 99.4 % diagnostic accuracy in the identification of 10 fault states of rolling bearings. Compared with the optimization parameter support vector machine method,it reduced the identification time and improved the efficiency of fault diagnosis,which has certain advantages.

Key words: rolling bearing, fault diagnosis, feature dimension reduction, pattern recognition, Locality Preserving Projection (LPP), Variable Predictive Model based Class Discriminate (VPMCD)