Research On Underwater Test Environment Inversion Method Based On Deep Learning

In recent years,with the continuous improvement of the comprehensive national strength of the country,the utilization efficiency of the ocean is getting higher and higher.To develop and utilize Marine resources more efficiently,it is necessary to have a full understanding of the Marine environment.Therefore,in recent years,research on inversion methods for underwater test environment is increasingly strengthened.The inversion of underwater test environment is an important content in the field of underwater acoustic research.It is also one of the bases for understanding and studying underwater acoustic to study how to obtain the information of sound velocity,temperature,salinity and other parameters of underwater environment.Above these information,although you can directly through a variety of instruments to artificial direct measurement for the specified location and underwater environment of information about a specific depth,but if you want a faster,more convenient and more accurate access to a wider range of underwater environment parameters,this way of using artificial measurement,obviously not practical need.In the past,underwater acoustic experts proposed a method based on matched field inversion to deduce the required underwater test environment parameters,but the matched field inversion must first establish the underwater acoustic propagation model suitable for measuring the underwater test environment area as a forward model.However,the underwater test environment is very complicated,and the method based on matched field inversion has some problems in practical application,such as complicated solution and inaccuracy.In recent years,with the breakthrough of deep learning technology in many research fields,how to use deep learning technology to realize the inversion of environmental parameter information of underwater test is the key of current research.Based on the above background,a feature extraction method and inversion method for underwater test environment inversion proposed in this thesis.The traditional feature extraction method of acoustic signal based on Fourier analysis relies on the prior knowledge,and is not ideal in accuracy and effect.In this thesis,a multiscale attention network(MSAN)based environmental feature extraction method is proposed by combining convolutional network,self-attention mechanism network and one-dimensional residual network to extract multi-scale features from input data in the network,so as to solve the feature extraction problem of underwater test environment inversion.Against traditional matched field inversion method in underwater test inversion calculation in complex environment,such as inaccurate problem,the transpose convolution,one dimensional residual,convolution method such as network combined in this thesis,through the network to the input characteristics of multistage hybrid fusion,a network based on hybrid fusion(MFN)underwater test environment inversion method proposed in this thesis.To sum up,the underwater test environment inversion method based on deep learning proposed in this thesis can be more convenient and accurate for the parameter information of the underwater test environment inversion.Through the research of this thesis,a new method based on deep learning is provided for the study of underwater test environment inversion.The feature extraction and inversion methods proposed in this paper have achieved better results than the traditional methods in experiments.