Research On Bearing Fault Diagnosis System Based On Empirical Mode Decomposition

Bearing is one of the most widely used and also the most easily damaged mechanical parts, so bearing fault diagnosis is of great significance. This dissertation aims to process the collected vibration signal of rolling bearing so as to develop a diagnosis platform for rolling bearing fault.Firstly, signals need to be denoised. This dissertation proposes the wavelet entropy threshold denoising method based on EMD, as the traditional EMD algorithm directly ignores IMF component so that signals are distorted and because of the algorithm defect resulting from traditional wavelet entropy threshold taking the median values of high-frequent wavelet coefficient as noise variance. This algoritm combines EMD with wavelet entropy to denoise IMFcomponent with wavelet entropy after decomposing the target signals with EMD. In this way, the effective signals and noises could be distinguished clearly by using wavelet entropy and finer denoising could be achieved.Secondly, the feature parameters of the denoised signals should be extracted. For vibration signals, some feature parameters change with fault types and fault degrees. This dissertation proposes characteristic extraction method of partial mean of multi-scale entropy based on EMD, because the multi-scale entropy has advantages of analyzing degree of data complication in multiple scales and mean deviation can reflect the general tendency and reliance degree of a set of data. This method decomposes signals into several components by EMD to get the corresponding partial mean of multi-scale entropy of every IMF component and take them as the feature parameters of signals.Thirdly, it is necessary to use feature parameters on vibration signal for fault pattern classification. This dissertation proposes the existing method of generate initial population randomly of mind evolution algorithm with chaotic function, and use chaotic mind evolutionary algorithm to optimize the BP neural network’s initial weights and thresholds, as the traditional BP neural network could easily cause the network fall into local minimum,slow convergence speed and the result is unstable because of the traditional BP neural network produce initial weights and thresholds by using random way. Besides, as the mind evolutionary algorithm produce population randomly so that the algorithm convergence rate is low in numerical optimization and fall into local minimum easily because of the mind evolution algorithm produce population in a random way.Lastly, this dissertation develops a diagnosis platform of rolling bearing fault based on LabVIEW. The functions of denoising rolling bearing vibration signals, extractingcharacteristics and categorizing fault types can be realized. This platform categorizes the fault types of rolling bearing vibration signals collected by QPZZ-II experiment platform and the results are effective.