The Feature Extraction Method Of Underwater Target Radiated Noise Based On Manifold Learning

In the underwater sound domain,the radiation noise signal reflects the related information of underwater target propeller rotate speed,power system configuration and target navigation mode.It is a very significant research topic to recognize target with the difference of the radiated noise.Large amounts of initial data describe the target in detail and also bring a lot of redundancy.So it becomes more and more important to explore and research a new effective method to obtain the target feature.Manifold learning method acts as a rapid and effective method of data reduction.It can explore the inherent law of the data and find the substantive characteristics for completing the feature extraction task.By combining the theory simulation with the experimental treatment,the thesis will analyze and research the following contents:1)Explaining the basic theory of manifold learning.Introducing the basic concepts of manifold,topological structure and manifold learning.Describing the main design and implementation steps of the four manifold learning methods such as LLE,LE,Isomap and LTSA.Comparing the characteristics of each manifold learning method and analyzing the influence factors of manifold structure.2)Applying manifold learning method to make the research of noise reduction.Constructing the time-domain signal into a high-dimensional phase space to obtain highdimensional data.Then applying the manifold noise reduction methods of the PCA and LTSA to reduce the noise interference and comparing the reduction effect.3)Applying the manifold learning methods to extract the low-dimensional manifold feature of the underwater target radiated noise.Making simulated analysis the characteristics of variable speed and smooth sailing for the underwater target.Introducing the general situation of the test data,using the STFT and WT time-frequency analysis technique to perform pre-phase analysis and treatment to the test data.Then extracting the manifold characteristics of the underwater target with three manifold learning methods such as LLE,LE and Isomap.4)Adopting SVM method to recognize the underwater target and obtain the recognition rate curve.Feeding the low-dimensional manifold characteristics extracted by different manifold learning methods into the SVM for making the classification and recognition.Comparing and analyzing the influence of the manifold learning feature extraction technology to the underwater target recognition rate.