Feature Extraction Of Underwater Transient Signal And The Fusion Of Multiple Classifiers

Feature extraction of underwater transient signal is the key of the airdrop targets classification. Historically, there are many methods of signal analysis which are applied to extract signal characteristics, such as Fourier Transform, Short Time Fourier Transform, Wigner-Ville Transform, Wavelet Transform, and so on. But each of them has difficulties when they are used in the extraction of characteristics of underwater targets. The difficulty of Fourier Transform is that the signal must be linear, strictly periodic and stationary; the difficulty of Short Time Fourier Transform is that size of the window function is unchangeable; the difficulty of Wigner-Ville Transform is that it has severe cross terms; the difficulty of Wavelet Transform is that it can only change the shape of its window. So this paper uses the method of Hilbert-Huang Transform, which is a powerful method for nonlinear and non-stationary time series analysis.In the area of target classification, the Artificial Neural Network classifier is the most widely used. Different characteristics of signal or different classifiers with different structures will give diffentent classification results, which we may use to improve the performance of the system. The weighted vote method of multi-classifier outputs is an effective multi-classifier fusion algorithm, which has simple terms and is suit to deal with actual problems, so we adopt this method in this paper.In this paper, we first review the methods for processing non-stationary data which have been used widely and point out their limitations, and then introduce Hilbert-Huang Transform method. Through the study on the structure characteristics of the signal from airdrop target water-entry, we do the research and extract the characteristics from three parts that are pulse signal, "quiet" interval, and fluctuant signal. We have studyed some artificial neural netwoks and used them respectively to classify the targets, and the results have shown that they have the complementarity. In the last of the paper, we have studyed the multi-calssifier fusion algorithm, and have used it to classify the targets, then we get a good result, which proves that the fusion method can be used in classification problems.