Study Of Signal Feature Extraction Methods And Applications

Signal feature extraction is a process that obtains information from signals and a foundational and key technique for many fields such as pattern recognition, intelligent system and machinery fault diagnosis. Feature extraction method has been a research direction followed with interest duo to the complexity of signals and the combinability of multidisciplinary knowledge for feature extraction. Following the latest progress of signal processing techniques, this dissertation explores ways of studying feature extraction methods and develops feature extraction methods by closely combining multidisciplinary knowledge with specific engineering applications. The research work is introduced as follows:1.Effect of spectral leakage in discrete Fourier transform on frequency, amplitude and phase estimates of multifrequency signals is studied. Based on the interpolated fast Fourier transform (IFFT), and considering the long-range leakage effect, an iterative IFFT algorithm is proposed. The novel method can eliminate the long-range leakage effect and improve the accuracy of the parameter estimates effectively. A comparative study of the proposed method with the well-known Grandke's IFFT and Liguori's IFFTc is presented. It is found that, in the noise circumstance and with the diverse frequency, amplitude and phase parameters, the proposed method outperforms the other ones, and provides the best estimates.2.Exponentially decayed sinusoidal function is suggested as a dictionary atom after having a good understanding of machinery impulse response waveforms. A method of adaptive signal representation with the atom is proposed based on the matching pursuit and the genetic algorithms. By using the proposed method, both the periodic and the impulse response waveforms can be separately extracted from the signal. Experimental results of machinery dynamic signal decompositions show that the proposed method can efficiently yield representation which is sparser and physically more interpretable than using the well-known Gabor atom.3.A nonparametric method for extracting feature waveform from signal is studied. Using template signal that contains prior information, a set of nonparametric basis functions is obtained firstly by means of a filter bank, and then a feature waveform atom that is described without any parameters is constructed. The feature waveforms extracted from signal using the method is physically interpretable duo to the employ of template signal. The filter bank makes the dictionary atom shape adaptive. Simulated and experimental results...