A Research On Transfer Learning Method Based On Data Feature Analysis

With the rapidly development of the Internet,today's society has entered the era of network and information,and in order to obtain the necessary data accurately and efficiently,the first step is to classify the data.Transfer Learning can use a small amount of labeled training samples or source data,to build a reliable model to forecast the target domain data(source data and target data can do not have the same data distribution).It extends the limitation on the distribution between source samples and target samples in traditional classification learning.On the other hand,due to the influence of underwater complex environment and the uncertainty of sound signal transmission,underwater sound signal processing and classification has always been recognized as a difficult problem in the world.However,the traditional methods mostly use pure mathematical methods to process and classify sound signals,which often fail to obtain satisfactory classification results.In recent years,due to the continuous development of deep learning,it is more and more feasible to use deep transferable network to classify sound signals.Various kinds of methods for transfer learning are analyzed and summarized,using the advantages of existing methods,as well as for its shortcomings,this paper proposes a new transfer learning method based on the analysis of underwater sound characteristics.On data feature processing,in the light of the non-stationary underwater sound signal,this paper proposes a time-frequency signal analysis method,that is suitable for non-stationary signal,called NSST(Non-Stationary Signal Transform,NSST),and it can make the voice into 2-dimen spectrum,then the corresponding MFCC coefficient is obtained by cepstrum distribution,and use it to be training data for transfer learning model.In terms of transfer learning,this paper uses a Siamese neural network,based on convolution layer of network respectively to extract the characteristics of the source domain and target domain data sets,and add the L2 norm in the loss function to prevent the overfitting phenomenon for the model because of increasing the full connection layers.The traditional transfer learning methods only consider either the distribution of differences between data sets,or the invariance between domains.In this paper,we consider the functions of the two methods are similar,so we combine them innovatively,and propose a new transfer learning method DFA(Data Feature Analysis,DFA),using both the Maximum Mean Discrepancy algorithm to match the distribution differences between the two data sets,and the gradient inversion to prevent the gradient descent of the domain classifier in the phase of err propagation,so that can preserve the domain invariance between domains.A parameter dynamic adjustment method is also proposed to dynamically adjust the importance degree of the two methods in the iterative training process.After the transfer learning model has been trained,this paper innovatively applies it to the recognition and classification of underwater sound signals.Finally,the NSST method and the DFA method proposed in this paper are verified by experimental test.Through the analysis of the experimental results show that the proposed transfer learning method based on the underwater sound signal analysis achieves the desired ideal effect.