Research On Transfer Learning Based Automatic Modulation Classification

In communication systems,automatic modulation recognition promotes the development of many important signal processing applications,such as cognitive radio,spectrum sharing and so on.In recent years,because of the requirements for the amount of training data and data distribution,the deep learning models,which are often used in modulation recognition,cannot solve the following problems in the actual communication scene:(1)the data distribution is offset due to the complexity of the communication environment.(2)For a specific data distribution,the amount of training data is uncontrollable.To address these challenges,we study the automatic modulation recognition based on transfer learning.Transfer learning can release the constraints of data distribution and the amount of training data for classification model.(1)In this paper,we introduce the adversarial transfer learning architecture(ATLA)into the field of automatic modulation recognition,incorporating adversarial training and knowledge transfer in a unified way.The model mines the potential information in the source domain through adversarial training,and then applies it to the target domain to realize the asymmetric mapping between the two domains.The structure of ATLA is divided into three parts:source model,target model and discriminator.Through sufficient experiments,we observe that our ATLA framework can alleviate the demand of training data for the target model,and outperforms the parameter transfer approaches.However,the above structure is only applicable to the case that the source model and the target model can share the same parameters.When the dimensions of input vector are different,or the source model and target model have different structures,the above adversarial transfer learning will be invalid.(2)In order to ensure the independence between the source model and the target model,this paper proposes to improve the above structure and realizes the cross-model adversarial transfer learning architecture(CATLA).This method mainly makes the following three adjustments:(a)The structures of the target model and the source model are independent and do not interfere with each other.(b)The target model uses the small training dataset of the target domain to train as the initialization.(c)The input of the discriminator is replaced by the output of softmax layer of the classification model.Experimental results show that CATLA can effectively solve the transfer problem between different model structures,and is superior to parameter transfer and ATLA in terms of generalization,tolerance of data distribution differences and so on.