Research On Identification Of Unknown Radio Emitters Based On Deep Learning

The identification of radio emitters is vital to the public safety and military electronic warfare.With the development of modern information technology and the growing popularity of various electronic devices,the electromagnetic environment is becoming more and more complex.The complex electromagnetic environment brings difficulties in feature extraction to emitter identification.It also brings an issue that how to distinguish unknown emitters from known emitters.In order to solve these problems,thesis studies the main modules of radio emitter identification system and the unknown recognition method.The main modules of the recognition system include data preprocessing,feature extraction and classifier based on neural network model.This thesis also studies and compares the different methods of these modules by experiments.As for the unknown recognition,this thesis draws on the field of computer vision,and uses three deep metric learning methods to realize a radio emitter identification system with unknown recognition.The main work of this thesis has the following two aspects:1.Close-set radio emitter identification.It is intended to explore how to conduct data preprocessing and feature extraction,and which classifier based on neural network model can jointly achieve better recognition results.In data preprocessing,thesis proposes an estimation method of in-band signal-tonoise ratio(SNR),which can effectively estimate the target signal quality in the complex electromagnetic environment and then filter the data;in feature extraction,thesis proposes zoom-STFT and STFT feature fusion method,zoom-STFT zooms in and refines the interest part of frequency spectrum without changing the feature size,thereby filtering unnecessary information and enriching slight features.In the recognition of 10 LTE mobile phones,the recognition accuracy of the zoom-STFT is 10% higher than that of the ordinary STFT of the same size.This thesis also sets up multiple sets of controlled experiments to study and analyze how different experimental setup such as data amounts,signal filtering methods,and CNN model affect the results.2.Open-set radio emitter identification.It is intended to solve the problem of identifying unknown radio emitter.In order to recognize unknown pattern,this thesis innovatively applies various deep metric learning methods into the identification of unknown radio emitters.These metric learning methods include Triplet Loss,Center Loss,and the improved Triplet Loss this thesis proposed.To compare the performance differences among these metric learning methods and Softmax Loss,this thesis uses different hyperparameters to train models on four different loss functions,and uses different plans to test these models.In the end,the aforementioned metric learning methods perform well in unknown emitter recognition,them can achieve more than 90% accuracy both in the known set and the unknown set.For the identification task for the same model of communication station,at 1% unknown recognition error rate,the improved Triplet Loss this thesis proposed can achieve 94%accuracy in the known set,which is 3% and 46% higher than that of Triplet and Softmax respectively.