Research On Nonlinear Feature Extraction Of Underwater Acoustic Target

In the complex and changing marine environment,it is becoming increasingly difficult to extract features from underwater targets.The traditional feature extraction method is relatively simple and cannot be applied well to the underwater target,especially the feature extraction of shipradiated noise.This thesis focuses on the nonlinear feature extraction of ship-radiated noise.First,the thesis proposes a novel ship-radiated noise feature extraction algorithm based on variational mode decomposition,weighted permutation entropy and local tangent space alignment.The proposed algorithm calculates the mode number of variational mode decomposition based on the variance of the center frequencies of the intrinsic mode functions.Weighted permutation entropy fully considers that neighboring vectors having the same ordinal patterns may heve different amplituds,which solves the problem that permutation entropy cannot effectively distinguish between abrupt regions and stagnation regions.In addition,high-dimensional features obtained by feature extraction of ship-radiated noise often have data redundancy.The algorithm introduces a non-linear manifold learning algorithm with local tangent space alignment to achieve dimensionality reduction.The results of simulation experiments show that the proposed algorithm fully combines the advantages of variational modal decomposition,weighted permutation entropy,and local tangent space permutation,which can better classify the ship-radiated noise.Based on this,as the single-scale entropy analysis often cannot fully reflect the complexity of the time series,the thesis proposes a weighted composite multi-scale permutation entropy to measure the complexity of the ship-radiated noise.The weighted composite multi-scale permutation entropy not only solves the problem of information loss in the coarse graining process of multiscale permutation entropy or weighted multi-scale permutation entropy,but also can effectively detect signal mutations and better realize the multi-scale analysis of ship-radiated noise..Finally,considering that entropy,as a kind of chaotic feature,can accurately reflect the dynamic changes of time series,but still cannot fully measure the complexity of series in some cases.The thesis studies features of chaos and fractal dimensions such as the maximum Lyapunov exponent,correlation dimension,Hurst exponent and box dimension.The experimental results prove that the four characteristic parameters can distinguish the three types of ship-radiated noise well,and the anti-noise performance is better than the weighted permutation entropy.In addition,the joint eigenvectors composed of Hurst exponent and maximum Lyapunov exponent were fed into a multi-class support vector machine based on particle swarm optimization,achieving a 100%recognition rate.