Second Order Blind Identification Algorithm And Its Application In Multi-fault Diagnosis Of Machinery

The fault feature of mechanical multi-fault are disturbed by each other,which makes it more difficult to be identified or analyzed than single fault.Thus,this thesis researches the blind source separation based on second-order statistics deeply,and applies it into mechanical multi-fault diagnosis.The possible problems that come up when separating the multi-fault source signals are analyzed,the solution is also proposed in this thesis.The specific contents are described as follows:(1)Mixing mechanism of several mechanical failure signals is studied in the thesis,then the characteristic of common mechanical components fault is introduced.Considering the mixing mechanism and hypothesis of second order blind identification,the separability of mechanical mixed signals is verified.(2)Aiming at the problem of unknown mixed model,a method of calculating the mixing matrix based on weight-adjusted variant second order blind identification is used in this thesis.The weight-adjusted variant second order blind identification is the improved algorithm of original second order blind identification algorithm,it could separate source signals better than original one.Moreover,the thesis analyze the algorithm's performance in mechanical multi-fault diagnosis field.(3)General blind source separation algorithm is unable to estimate source number,and the information of source number is much less when the observed signals' dimension is insufficient.To solve this problem,a new underdetermined source number estimation algorithm based multivariate wavelet packet decomposition is proposed.Applying the multivariate wavelet packet decomposition algorithm to obtain the vectors that includes the information of source number,then the vectors are processed by singular value decomposition to obtain feature distribution.Finally,source number is estimated by dominate eigenvalue criterion from feature distribution.(4)In underdetermined condition,not only the source number estimation is much difficult,but the weight-adjusted variant second order blind identification can't acquirethe mixed matrix of mechanical mixed signals.This thesis apply kernels to solve the problem of single channel blind source separation.Positive define matrix is reconstructed by kernels,and weight-adjusted variant second order blind identification algorithm was applied to separate the source signals.Thus,the fault feature can be identified.