Research On Fault Diagnosis Of Rotating Machinery Based On Local Linear Embedding Algorithm

Rotating machinery whose the key part is rotor-bearing system has been the coreequipment in modern engineering industries. When the rotating machinery in such ashigh speed, over-load working environment, it is easy to appear a series of failureslike the rotor imbalance or misalignment and so on. If not timely diagnosis andmaintenance, it will cause certain economic losses and social influence. Faced withhigh-dimensional nonlinear characteristics of reality fault data, the traditional linearmethod has certain limitation. However, the nonlinear manifold learning method canaccurately extract the effective information of fault data, and avoid "dimensiondisaster". Therefore, with the rotor-bearing system as the research object, in this paperit designs fault simulation experiment scheme of rotor system. On the basis ofstudying on manifold learning method based on local linear embedding algorithm, itoptimizes the algorithm to get better effect of fault diagnosis. The main research workis as follows:(1) According to the research object to design fault simulation scheme of rotorsystem. Then simulate the normal state, rotor imbalance, rotor misalignment and thepedestal looseness fault of the rotor system in the integrated fault experimental bench.Collect corresponding vibration signal and analyze the time-frequency characteristicsof state signal.(2) Analyze the main ideas of the classical manifold learning methods as well aseach method’s advantages and disadvantages. In view of the problems of reality datadistribution, and local linear embedding algorithm (LLE) is easily affected by theneighboring point, an improved local linear embedding algorithm of homogenizationdistance (MLLE) is proposed. By changing the distance measurement of method andusing distance adjustment factor to make the samples well-distributed. Finally,comparing the fault recognition rate of two kinds of LLE and discussing the changesof recognition rate follow the changes of parameter k nearest neighbor points ord-dimensional embedding.(3) In order to increase fault recognition rate and stability of the algorithm, onthe basis of studying methods of kernel and supervised learning, a nuclear supervisedmethod based on locally linear embedding (KSLLE) is proposed. According to kernelfunction constructing kernel matrix space to map data from data space to a kernel feature space, then calculating distance of matrix data in the kernel space distance forSLLE dimensionality reduction and classification. On the other hand, furthercombining the thought of MLLE and supervised learning, a kind of supervision andhomogenization of distance local embedding algorithm (SMLLE) is put forward. Firstthe homogenization distance metric is used to make the samples well-distributed, thenthrough supervised learning to increase the sample points’ category information,which gather the same class and disperse different class in order to achieve effectiveclassification. Finally, analysis and comparison the fault recognition rate of LLE,SLLE, KSLLE, SMLLE.(4) Considering the different fault types in many different manifold subspace sothat the ideal of manifold learning extended from single manifold to multi-manifold, amulti-manifold fault diagnosis method based on locally linear embedding is putforward. First build the framework of multi-manifold LLE, then according to theintrinsic dimension of data sets to use LLE algorithm respectively extract data featureon the corresponding manifold, which is advantageous to classification recognition ofthe new sample. Finally, verify the feasibility of the algorithm by analyzing theexperimental data.