Research On The End Effects Mitigation And Application Of Hilbert-Huang Transform

Hilbert-Huang Transform (HHT) is a powerful tool for non-linear and/or non-stationary signal processing developed over the past decade which has shown strong vitality and lots of advantages since it was first introduced. Though HHT is remarkably novel and creative, there are many open problems need to be solved. The paper is mainly divided into two parts: theoretical research and applied research.Firstly, a series of theoretical researches and explores are carried out to mitigate the end effect of HHT. One of the difficult problem to which HHT inevitably confront is end effect, which prevents HHT to extract essenctial characteristics from the signal effectively. That is to say, the intrinsic mode function (IMF) and Hilbert spectrum obtained from HHT may not exact. Aiming at mitigating end effect of HHT, a method is introduced, which is based on the well-known grey prediction model. Without changing the properties of the original signal, this method reduces the end effect effectively. Based on the typical deficiencies of this method, two improved methods are presented. Due to the particular cause of end effects in empirical mode decomposition (EMD) process, whose extreme may be non-equal interval, a non-equidistance grey model is introduced to improve the prediction accuracy. On the other hand, since grey prediction model may fail to predict complicated signal from the real world, a non-linear grey model termed GM (1,1,α) is introduced to enhance effectiveness of the method.Secondly, a series of applied researches and explores are focused on the hyperspectral classfication. It is a difficult problem for hyperspectral classfication as the hyperspectral image contains enormous bands. We explore a novel hyperspectral classification algorithm which integrates BEMD and support vector machine (SVM). By virtue of BEMD, the selected hyperspectral bands are decomposed into several bi-dimensional intrinsic mode functions (BIMFs), which reflect the essential properties of hyperspectral image. We further make full use of SVM, which is a widely accepted supervised classification tool, to classify the suitable sum of BIMFs. Experimental results indicate that though the proposed method has no advantage in computing time, it exhibits higher classification accuracy and stability than the classical SVM.