Research On Feature Extraction Methods And Its Application To Nonstationary Signal Based On HHT

As an important content of modern signal processing, nonstationary signal analyzing and processing has always been the hot spot of the theory of scientific researches and engineering practices. Although the related processing techniques like short time Fourier transform and wavelet analyzing give a proper description of the time-varying and nonstationary signal indeed. However they all take the theory of Fourier transform as the ultimate basis, using these methods to analyze nonstationary signal may easily produce spurious components and aliasing phenomenon. To void this problem, the most directive method in nonstationary signal analyzing and processing is to use basic measures which have the local property, like instantaneous frequency. Hilbert-Huang transform(HHT) is new nonstationary signal analyzing technique which based on instantaneous frequency. HHT can analyze the inherent frequency character of signal in the whole time domain effectively, and has been widely applied in biomedical sciences, mechanical fault diagnosis, oceanographic engineering and other fields. In this paper, the main existing time-frequency analyzing methods and their targeted and limitations are studied as the basis. Then we focus on the time-frequency analyzing and feature extracting method based on HHT. The research works are introduced as follows:① The basic theory of HHT algorithm is researched, including instantaneous frequency, intrinsic mode function, empirical mode decomposition(EMD) algorithm and Hilbert spectral analyzing. The decomposition principle and steps of the EMD algorithm are introduced in detail, and the completeness and orthogonality analysis about it are also given. The main existing weaknesses of HHT including mode mixing and end effecting problem are deep analyzed, and the relevant improvement measures are given.② On the basis of Hilbert time-frequency image analyzing, a feature extracting method based on histogram to nonstationary signal is proposed. Firstly, the method of time-frequency representation based on HHT is researched, and the mode mixing and end effect problem to the representation is analyzed. A comparative study between the HHT and traditional time-frequency analyzing method is conducted by the simulation experiment. The method of feature extraction based on time-frequency image processing is given at the last.③ On the basis of Hilbert marginal spectrum analyzing, a feature extracting method based on Hilbert marginal spectrum to nonstationary signal is given. Firstly, the marginal spectrum of signal is presented, and then the difference between the marginal spectrum and the Fourier amplitude spectrum is researched by the simulation experiment. Last, the method of feature extraction based on marginal spectrum is introduced and the features of spectral energy and entropy are analyzed in detail.④ The two feature extraction methods are applied in seizure detection in electroencephalogram(EEG) signals. Firstly, the EEG signals are analyzed by using the above two methods and the significant features are extracted. Then the support vector machine is employed for features classification. The experiments show that the two methods can both get effective results in EEG signal classification.