LMD Method of False Identification Component Based on Stepwise Regression Analysis

The actual project, many signals in the frequency components change over time, using the traditional Fourier transform for spectral analysis are very limited. Available for non-stationary signal analysis methods, such as the WVD, short-time Fourier transform, wavelet transform, there are also some problems. For example, the problems are cross-term in WVD method, the frequency resolution of short-time Fourier transform and select the appropriate wavelet in the wavelet transform. Jonathan S. Smith proposed a local mean decomposition (LMD) method on the empirical mode decomposition (EMD) algorithm, and first LMD applied in electroencephalogram (EEG) signal time-frequency analysis. LMD algorithm and EMD algorithm has the following differences: LMD uses the method of moving average instead of using cubic spline interpolation. As the final result of EMD decomposition is a series of IMF. And the end result of LMD decomposition is a series of AM-FM signal. Using inverse cosine function directly obtained the instantaneous frequency. Product function (PF) of LMD than IMF of EMD can save more frequency and envelope information. However, the signal change after LMD will produce some False PF component, especially has the low frequency of false weight. In this paper, based on the problem for False PF component in LMD, using of false component identification of the theory based on stepwise regression. PF component in the decomposition of the original signal a significant impact as a criterion, to identify and remove false weight. Through test the simulation signal and the actual voice signal, this improved method relative to the current method of the correlation coefficient by removing false weight has distinct advantages and reasonable.