Diagnosis Of Mechanical Failure Local Mean Decomposition

Machinery fault diagnosis is designed to find the potential fault andmechanical vibration will happen in the work, and a large amount of vibrationsignals are non-stationary, extracting the various characteristics of the actualmeasured vibration signals, and applying characteristic parameters for conditionmonitoring and fault diagnosis, so the digital signal analysis method of thevibration signal is the main research direction. When machinery in the event ofa fault, the vibration signal will exhibit nonlinear and non-stationary, thecharacteristic parameters reflected by various signal are different, and they arecorrelation and redundancy, this will reduce the generalization ability and theprecision of recognition classifier. So the reasonable choice of characteristicparameters and classifier is of great significance to detect the fault.In this paper, according to the above problem, do the following research:1. Local mean decomposition is applied to analyze mechanical faultvibration signal. Local mean decomposition has the performance of the adaptivetime frequency distribution characteristics of the signal, however, in the localmean decomposition of the signal, the trend of the endpoint can not be predictedthat cause contaminate the entire signal sequence, the original moving averageof the signal cause over-smoothing treatment, fault characteristics can not beaccurately extracted. The article put forward improvement program on the basisof introducing local mean decomposition. It is verified by simulation signal inMATLAB and the PF component decomposed by improved algorithm have highcorrelation coefficient with the original signal. The energy of PF component isimported BP network, the output show that the optimized PF component hashigher classification accuracy.2. When a device fault occurs, a variety of statistical parameters of vibrationsignals contain a wealth of state characteristics that is a relevant and redundant,which will reduce the generalization ability and the recognition accuracy of theclassifier. A series of production function PF component of the signal are obtained by LMD, using KPCA to remove redundant features of sample datawhich include the vibration signal sequence domain characteristic parametersand timing AR parameters, as well as energy entropy PF component. To extractthe nonlinear principle component in the input data space, and then use LSSVMfor fault classification. Experimental results show that, PF energy entropycharacteristics are better than temporal characteristics and timing parameters ofAR parameters, and KPCA-LSSVM classification model with respect to thedirect use of LSSVM has better classification accuracy.