Research On Local Mean Decomposition And Its Application In Mechanical Fault Diagnosis

Vibration signals in the real-world are non-linear and non-stationary. Their frequencies change with the time. Conventional signal analysis method, Fourier transform (FT), can only achieve the range of frequency distribution,but not the saltation of the frequency in the whole time. Short time Fourier transform (STFT), Gabor transform (GT) and Wignar-Ville distribution (WVD) divide the vibration signals into several segments with a constant window functions. Besides, when there are several components in a signal, cross term follows when use WVD method. The window functions width of wavelet transform (WT) can be different, moving along the time axis. However, it divides the signal mechanically and is unable to decompose signals in a self-adaptive fashion. N. E. Huang proposed a new time-frequency analysis approach in 1998, and J. S. proposed another self-adaptive method, local mean decomposition (LMD), which is a great breakthrough of time-frequency method based on FT.LMD is an adaptive time-frequency analysis method, which can divide vibration signal into several product functions (PFs). And their frequencies range from high to low automatically. Each PF is multiplied by a pure frequency modulated function and an envelope function. The decomposition way, in which the LMD used, solves the mechanical division problem of the time-frequency analysis method before. And it makes the decomposition result be more accurate. In this paper, the theory of LMD algorithm and the problems in it are introduced in detail. Its end-point effect problem is improved in this paper. Meanwhile, introduce a method, which combine LMD with Teager energy operator and 1.5 dimensional spectrum. This method makes the fault characteristic frequency of roller bearing to be more explicit. Combine LMD method with enhanced envelope spectrum, and propose a new selection criterion of PF components, it makes the fault characteristic frequency of roller bearing to be more obvious. Combine LMD with support vector machine (SVM), meanwhile, propose a new feature extraction method to extract the fault feature, then use SVM to classify these fault features, the fault degree and rotor position of the fault bearing are identified by this method. The results show that the method is feasible.