Research On Passive Recognition Methods Of Marine Targets

The passive recognition of marine targets has been the key technology in underwater signal processing.The problem is highly complicated due to the reduction of ship radiation noise level and difficulty on collection of marine targets.Extracting effictive features under low signal noise ratio(SNR)and identifying targets rapidly and accurately with small sample data are new demands for passive sonar recognition.In order to satisfy the demands,the extraction method of ship radiated noises from ocean ambient noises under low SNR is studied,the effictive feature extraction algorithms of ship radiated noises are discussed,the classifier suitable for small sample is established,the decision level fusion algorithm based on multi-features is developed,thus under low SNR,the recognition of two typical marine targets is realized efficiently with small sample data.The thesis mainly consists of four aspects:(1)The weak signal extraction method based on nonlinear dynamical model of ship radiated noises and ocean ambient noises.Based on phase space reconstruction theory,the second-order Volterra series filter is applied to establish a nonlinear dynamical model.As the Volterra series filter converges slowly and the convergence parameters are difficult to determine,the estimating method of the Volterra kernel fuction is modified by Kalman filter,which helps to raise forecast accuracy and convergence speed.The method of extracting weak signals is built based on the nonlinear dynamical model.Under low SNR,the time-domain waveforms of regular signals and chaotic signals are extracted from chaotic background noises,also the time-domain waveforms of different kinds of ship-radiated noises are separated from ocean ambient noises.The weak signal extraction method based on nonlinear dynamical model can gain higher input SNR for the marine target recognition system.(2)The effective features extraction of marine targets radiated noises.Based on the physical mechanism of various noise sources,characteristic parameters are extracted from time-domain waveforms,PSD(Power Spectral Density)spectrums,LOFAR(Low-Frequency Acquisition and Ranging)spectrums,DEMON(Detection of Envelope Modulation on Noise)spectrums and topological structures of the phase space attractors.It is studied whether the two kinds of marine targets radiated noises have separability on time-domain waveform structure characteristics and nonlinear characteristics.It is verified that the two targets radiated noises are statistically independent and differ from each other in the phase space time delays,correlation dimensions and the largest Lyapunov exponents.Using principal component analysis(PCA)and linear projection analysis(LPA),the feature dimesion is compressed,which avoids data redundancy.Thus,the features of marine targets radiated noises are extracted effectively.(3)The classifier used for marine targets.To solve the small sample,nonlinear classification problem,the marine target classifier is optimally designed based on support vector machine(SVM).The modified methods of selecting SVM parameters are proposed.The grid search method is improved by a second search with a smaller step,raising searching speed.The algorithm DEPSO is proposed by combining the differential evolution(DE)and particle swarm optimization(PSO)algorithm,obtaining global optimum solution of penalty parameter and kernel function parameter of SVM.The generalization capability are improved greatly of the SVM classifiers.Classification experiments are carried out based on time-domain waveform features of six marine targets radiated noises.The experimental results examine the influence of the modified grid search method,the hybrid optimization algorithm DEPSO and the feature screening method on the SVM classifiers,raising the classification accuracy and reducing the calculation cost.(4)The decision-level fusion method based on multiple characteristics of marine targets.As D-S evidence theory can't deal with the cases with highly conflict evidences,a modified weighted mean method of the basic probability assignment(BPA)is applied to preprocess the conflict evidences.As single sensor obtains incomplete information,the BPA function is established based on posterior probability and classification rates of SVMs.A decision-level fusion method of marine targets is proposed,screening time-domain waveform features,power spectrum features,LOFAR spectrum features,DEMON spectrum features and nonlinear features.The proposed method considers comprehensively classifying results of SVMs for the multi-characteristics.Sea experiment results validate the decision-level fusion algorithm utilizes advantages of multiple features,raises recognition rate of two typical marine targets greatly,enhances the fault tolerance capability and the detection performance of the passive recognition system.Synchronously,multi-characteristics are selected to decide which kind is more effective.To sum up,an effective way to classify marine targets is presented in the condition of low SNRs and small sample.The estimation method of Volterra kernel function had been improved,increasing the prediction accuracy by three to five orders of magnitude and speeding up the convergence.The model is built to separate ship radiated noises from ocean ambient noises under very low SNRs.The parameter optimization method of SVM had been modified.DE algorithm and PSO algorithm are combined,ensuring that the particle swarm converges to global optimal results.Thus the modified SVMs with hybrid optimization algorithm are suitable for marine target recognition with small training sample.A decision-level fusion method had been proposed based on multiple features of marine targets,which overcomes the influence of individual feature,such as incomplete information and low reliability.Sea experiment results indicate that the robustness and recognition rate of water-surface/underwater targets are improved greatly.The research results provide new ways to enhance the performance of passive sonar recognition system in the complex undersea environment.