Study On Multiple Kernel Learning SVM Based Fault Identification Method Of Rotating Machinery

With the rapid development of modern science and technology, rotating machinerycontinues melting towards maximization, complex, high speed, continuous process andautomation directions. Not only do these developments lead to higher productivity but alsohigher requirements are put forward for the safe operation of the rotating machinery.Monitoring and identifying the state of the equipment are important measures to ensure itssafe and reliable operation. Based on multiple kernel learning support vector machinetheory, this thesis studies solutions to some key problems in the process of rotatingmachinery fault identification.Support vector machine with a single kernel function can not meet some practicalrequirements such as heterogeneous information or unnormalized data, non-flatdistribution of samples, etc. Therefore, it is an inevitable choice to consider thecombination of kernel functions for better results. In this thesis, multiple kernel learningsupport vector machine is applied to the rotating machinery fault identification, and theeffectiveness of the method is verified by experiment.As a data processing method, ensemble empirical mode decomposition (EEMD) cannot only deal with the non-stationary and non-linear problem preferably but also cansuppress the phenomenon of mode mixing. This thesis combines EEMD with multiplekernel learning support vector machine. Before the fault classification, EEMD is used forsignal processing in order to get better classification results.In order to effectively extract the fault feature, it is necessary to select fault-sensitiveintrinsic mode function (IMF) from the decomposition results of EEMD. A sensitive IMFselection algorithm which based on multiple kernel learning is proposed. The algorithm isable to give each IMF quantitative contribution to the final classification result, which ismore persuasive and reasonable.According to the great limitations of individual feature, and for the sake of mixingthe features from different platforms together, a multiple kernel learning framework whichmerges several features together is performed to identify hydraulic pump fault. The algorithm is able to train different features by choosing different kernel function.Experiment results show that the proposed multiple kernel multiple feature methodsignificantly improves the classification performance.