Feature Extractions Of The Signal Based On Non-negative Matrix Factorization

Feature extraction has become indispensable to many research fields such as the signal processing,pattern recognition,geophysical exploration and mechanical fault diagnosis with the development of times and the progress of science and technology.It is obtain to extract characteristic information from the data.For a long time,we treat signals as smoothly for analysis and processing since the backward theoretical level and the lack of research tools.It is convenient,but neglects the inner connection of signal no matter in time domain or frequency domain.It can only present the statistical average results.However,the vast majority of the signal are not cycle and smooth in nature and engineering.Consequently,the study of non-stationary signal has important theoretical significance and application value.In this paper,we research some problems against the Non-stationary signal feature extraction and analyze the advantages and disadvantages of different methods of time-frequency feature extraction.Furthermore,we get the superiority of Non-negative Matrix Factorization(NMF)algorithm against the spotlight of signal feature extraction.Firstly,after introducing some basic concepts of time-frequency(T-F)analysis,this paper explains some methods of T-F,such as Short Time Fourier Transform,Gabor Transform,the Wavelet Transform,Bilinear Time-Frequency Transform,Adapted Signal Decomposition,Empirical Mode Decomposition and so on.Then compare the advantages,disadvantages and applications of the all above.Secondly,on the basis of Non-negative matrix factorization(NMF)algorithm and its advantages,this paper researches the superiority of NMF which applies to time-frequency filtering and feature extraction.An improved Non-negative matrix factorization is proposed.And then demonstrate the applicability of this method.This paper also proposes Spectrum Moments,Phase Moments and Sparse Features which can reflect the signal's time-frequency characteristics.The simulation results show that NMF algorithm is better in calculation and extraction results than traditional methods.Finally,this paper analyzes the performance of NMF algorithm.The simulation results prove that this approach is effective.Then with the help of Hidden Markov Model and the Sparse Features,we successfully identify the target and prove the effectiveness and the applicability of Sparse Features.