Research On Special Emitter Identification Technology Based On Transfer Learning

With the development of The Times,the requirements of modern battlefield for weapons and equipment are increasingly high,among which cognitive ability is one of the essential key abilities.Through the use of artificial intelligence technology,the equipment system can intelligently discover,determine,track,aim and attack the threat target,and evaluate the effect after the attack.The special emitter identification is one of the key links of cognitive reconnaissance,which has attracted widespread attention of scholars at home and abroad.Focusing on the emitter identification,this paper focuses on the problem of low recognition rate due to the deviation of emitter data distribution,and carries out research on emitter identification technology based on transfer learning.To be specific,the main studies on emitter signal recognition using transfer learning in this thesis are as follows:(1)A set of methods for generating individual emitter features and signal fingerprint features are studied,and deep learning is used to realize recognition under the condition of identical distribution of emitter data sets.Signal preprocessing and feature extraction were carried out for the acquired radiation source signals,and the identification network was constructed.Through the training of deep learning,the recognition rate of the data sets of nine radiation sources reached more than 97%under the condition of the same distribution.It is found that the difference of data sets will seriously affect the performance of the recognition network,and make the original training network basically lose the recognition ability,and the recognition rate is less than 15%.(2)A transfer learning algorithm based on model parameters,a transfer learning algorithm based on data feature domain adaptation and a multi-representation adaptation migration algorithm based on data feature adaptation are proposed to solve the identification problems under the conditions of different distributed radiation source data.For data collection of differences caused by the source domain model in the target domain the problem of lower recognition rate,based on the model parameters of migration algorithm text of nine sources of data sets for simulation,by fine-tuning multilayer parameters,stable under large sample can reach more than 96%of the recognition accuracy,a single fine tuning algorithm with 3% accuracy about ascension,under the condition of limited training samples is compared with the traditional CNN algorithm has identified more than 5.5% of the performance improvement;The domain adaptation migration based on data features is to reduce the differences between data features by adding an adaptation layer to the network.The image feature adaptation algorithm is improved by text and applied to the field of emitter source identification.The recognition accuracy rate is 56.6% by simulation under the real data collected from nine emitter sources.Further,in a single adaptive algorithm is proposed based on the characterization of adaptation algorithm,the basic network output were input to the multiple branch network,then each child deep network to extract different features of range error to join the training together with the loss function of the model,finally get the recognition accuracy of 72%,compared with single adaptation algorithm,increased accuracy by 15.4%.In conclusion,compared with the unadapted recognition network,the recognition rate of the adaptive network can be increased by up to 37%.(3)The special emitter identification method based on model and feature domain adaptation integrated transfer and the special emitter identification algorithm based on meta-transfer learning are proposed.The special emitter identification method based on model and feature domain adaptation comprehensive migration is to combine parameter migration with feature adaptation.Firstly,the model is trained by multi-representation adaptation algorithm for the first time,and then a small amount of target domain data is trained for the second time.Finally,the recognition rate of88% can be obtained.The emitter-source identification based on meta-transfer learning is aimed at the problem of small samples,and the MAML algorithm is improved,which combines the ideas of transfer learning and meta-learning,and utilizes the transfer of model parameters and the training of meta-learning machine.In a small sample of nine emitter classification tasks,compared with CNN algorithm,the recognition accuracy can be improved by 22%.