Rolling Bearing Fault Diagnosis Based On Genetic Programming And Kernel Principal Component Analysis

The paper carried out method based on genetic programming (GP) and kernel principal component analysis (KPCA) for rolling bearing fault diagnosis. The vibration signal from healthy state and typical faults (inner-race flaw, outer-race flaw and rolling element flaw) of bearing were got through the test, the fault signal feature was extracted and the time domain and frequency domain features symptom parameters (SP) were calculated. Then the combination to the initial SPs was carried on to optimize and build composite characteristics by GP to improve the accuracy of identification bearing working conditions. Meanwhile, the kernel principal component analysis was combined with genetic programming to further enhance the ability of the composite index of identifying bearing working condition. Details were as follows:(1) The time domain and frequency features of the vibration signal from four working states were caculated. Discrimination Index (DI) was applied to select the SPs with high sensitivity. Using genetic programming (GP) algorithm to optimize the characteristic parameters and comparing with non-optimized characteristic parameters, which brought out the conclusion that the composite index’s ability to identify rolling working conditions have greatly improved after GP.(2) The kernel principal component analysis (KPCA) was respectively used for original SPs and composite SPs after GP. Analyzing and comparing the ability of identifying the operating states of the first principal component and the second principal component. We can draw that the ability of identifying the operating states will be stronger after doing KPCA on the composite SPs after GP comparing with the original SPs.(3) The GP, KPCA and BP neural network were combined for bearing fault intelligent indentification method study to verify the validity of the proposed method in this paper. There were three sets controlled trials:the original SPs were input to the BP neural network; the characteristics after KPCA were input to the BP neural network; the characteristics after GP plus KPCA were input to the BP neural network, The result showed that the characteristics after GP plus KPCA has much stronger indentification for bearing working states.