Research On Gear Fault Diagnosis Based On Kernel Extreme Learning Machine

As a typical component that transfers kinetic energy and regulates rotational speed,gears play a crucial role in the safe operation of large-scale mechanical equipment.To satisfy the demands of industrial production,gears always need to be operated under harsh conditions such as overload,high temperature,high pressure,etc.,resulting in gears that are prone to failure,and in the event of a failure,it may cause the paralysis of the whole mechanical system.Most of the characteristic information of the gear failure is contained in the vibration signal generated when the gear failure occurs.Therefore,if as many fault feature information can be extracted from the vibration signals,and further be de-redundant and classified,then the gear fault can be accurately identified,which can ensure the safety and stability of operation of the equipment.In this paper,gear is taken as the research object.Based on analyzing the vibration characteristics and common types of faults of gears,the multi-domain feature extraction of gear vibration signals is studied.Then a cascading feature dimension reduction method based on Global Supervised Laplacian Score(GSLS)and Kernel Principal Component Analysis(KPCA),and a multiple fault recognition method based on Particle Swarm Optimization-Kernel Extreme Learning Machine-Binary Tree(PSO-KELM-BT)are proposed.Finally,the gear fault diagnosis system based on dimension reduction and fault recognition is implemented.The main contents of this paper are as follows:(1)Gear vibration characteristics analysis and multi-domain feature extraction.Based on the analysis of gear vibration characteristics and common fault types,the multi-domain feature extraction methods are used to extract the time domain,frequency domain,and time-frequency domain features,and a high-dimensional feature set is constructed to characterize gear faults,which serves as the data basis for feature dimension reduction.(2)Research on cascading feature dimension reduction based on GSLS and KPCA.To solve the problems of LS that the nearest neighbor graph parameter is difficult to set and over-rely on sample local structure information,a GSLS method fuses data category information and global structure information is proposed,which is used to obtain a small-scale and high-differentiated feature subset.Combing the KPCA method with the feature subset obtained from GSLS,the redundant information of the subset is further removed to reduce the feature set dimension.Experimental results show that the proposed cascading feature reduction method can combine the advantages of LS and KPCA and efficiently improve the discrimination of feature subsets.(3)Research on multiple fault recognition based on PSO-KELM-BT.Aiming at the problem of low accuracy of KELM,a KELM multi-classification algorithm where KELM works in single classification is proposed.The algorithm divides the training data by one-vs-all method based on the proposed mixed separability metric that fuses intra distance and inter distance into several training data sets,and the divided data are used to train the KELM binary classifiers separately,and then integrates the binary classifiers into an ensemble binary tree-structured classifier to improve the classification accuracy.Meanwhile,to solve the KELM parameter setting problem,the PSO algorithm is introduced to optimize the parameter of KELM and further develop the classification accuracy of the method.The experimental results show that compared with similar algorithms,the proposed method has higher classification accuracy.(4)Research on the application of feature dimension reduction and fault recognition in gear fault diagnosis.By researching the layered architecture technologies of system which fuse gear parameters monitoring,feature extraction,feature dimension reduction and fault recognition,a gear fault diagnosis system that includes parameters monitoring,signal analysis,feature extraction,feature dimension reduction and fault recognition modules and can display the processing results of each module in the form of visual interface is implemented.The operating results show that the proposed method has a good effect in the diagnosis of gear faults.