Research Of Classification Method For Ship-Object Based On Higher-Order Statistics

Higher-order Statistics (HOS) is the efficient analysis tool in analyzing non-Gaussian signal and nonlinear system.It has plenty of applications in signal detection,feature extraction and parametric estimation. Extracting frequency features of the ship-radiated noise and constructing classifier applied to object classification and recognition is the main contributiqns in this paper.Firstly,some definition and properties of HOS are introduced.HOS is insensitive to Gaussian process and HOS can completely suppress Gaussian noise in theory .This is the theoretical basis of signal detection and estimation,too. In order to display strong point of HOS in suppressing noise,three kinds of methods compared with HOS in harmonic signal with additive Gaussian color noise.Secondly,some methods of analyzing nonlinear phenomenon are presented based on three-order cumulant and four-order cumulant.An improved parametric bispectrum estimation,based on three-order cumulant, is brought forward for estimating frequency of quardratic phase coupling, and the simulation displays that this method can obtain a higher resolution in lower SNR and short data length.The mechanism of ship-radiated noise is analyzed. And then,the methods of extracting frequency features of the ship-radiated noise envelop are presented by using Bispectrum.A line spectrum search algorithm is put forword for locating the peaks of line spetra,which is used in HOS based fearure extraction.And four linespectrum features are extracted respectively based on 1(1/2)—spectrum and2(1/2)—spectrum.Lastly,an improved RBF neural network algorithm is chosen to classify three type ships. The method, combining auto-clustering algorithm and symmetry distance, improve each vector of hide layer neuron. And the weight from hide layer neuron to output layer neuron are determined by pseud-inverse method.In experiment, the feature extraction methods studied above are used to extract the frequency features of the practical ship-radiated noise samples.The ships are recognized with the features.The highest recognizable rate reaches 92.50% for B type.This further proved that the method of combining HOS and RBFNN is very efficient in analyzing the ship-radiated noise.