The Improved EMD Algorithm And Its Application In Non-stationary Signal Processing

The analysis and processing of non-stationary signal plays a crucial role in reallife, when dealing with this signal, the traditional time-frequency analysis method isshort-time Fourier transform, wavelet transform and Wigner-Ville distribution et al,but they all still can’t get away from the bondage of Fourier transform essentially.Hilbert-Huang transform is put forward by NASA scientists Norden E.Huang et al in1998, it is composed of empirical mode decomposition (EMD) and Hilbert transform.HHT is a new, self-adaptive time-frequency analysis method, it’s especially suitablefor nonlinear and non-stationary signal analysis and processing.This thesis mainly studies empirical mode decomposition method and improvedits end effect problems, and applied the improved EMD algorithm in non-stationarysignal processing. With the bearing fault signal feature extraction and sunspots timeseries prediction of these two non-stationary signals as an example to study. Detailsare as follows:(1) In view of the end effect problems in the EMD decomposition, on the basisof summarizing the existing signal continuation algorithm, this thesis proposed animproved method combining waveform feature matching with window function. Thismethod made full use of the advantages of both, effectively inhibit the end effectproblems.(2) A method combining improved wavelet threshold denoising and EMDdecomposition is applied to bearing fault diagnosis. Firstly the original signal wasdenoised by the improved wavelet threshold method, then the denoised signal wasdecomposed into several IMFs by EMD adaptively, the IMFs reflecting the faultcharacteristics were selected through energy-correlation coefficient method and faultfrequency would be gotten from envelope spectrum analysis. The experiment resultsprove that the method can extract fault characteristic frequency effectively.(3) A method combining EMD decomposition and combined forecasting modelis applied to the sunspots time series, first of all, this thesis preprocessed the originaldata through wavelet denoising method, then the denoised signal was decomposedinto several IMF components and remainder by EMD. In view of the characteristics of the low frequency and high frequency components, RBF neural network model andSVM model were used to predict them respectively, the final predicted value would begotten by adding each component’s result at last. The simulation results show that themodel has better prediction effect and higher prediction accuracy.