| Communication transmitter individual identification(CTII)technique is an important issue in communication countermeasure and reconnaissance.It can provide reliable evidence for communication intelligence reconnaissance and military action decision.Generally speaking,CTII technique consists of signal processing,machine learning,pattern recognition and many relevant techniques that applying on communication signal analysis tasks,and has the capability to extract individual features for identifying different communication transmitters according to the intercepted transmitted signals.On the basis of the features extracted,CTII is usually then realized by classification or clustering methods.Current CTII technique research is focused on utilizing supervised methods to learn the features of large number of labeled communication signal samples.However,the small sample size problem(which is called “3S”)under practical environment causes serious negativity on the application and development of CTII technique.That is to say,when under practical environment or at wartime,we can capture massive unlabeled communication signals but can only have limited communication signal observations.In such situation,it will apparently lead to a terrible classificaion result if implementing the supervised feature extraction or identification method.In order to solve the “3S” problem,deep learning theory was introduced to make full use of the internal structure information of massive unlabeled samples for extracting the essential feature of communication transmitter individuals.And thus,the CTII techniques were carried out from the point of semi-supervised learning in this dissertation.Firstly,the foreign and domestic research status of CTII was introduced,and the feasibility of CTII based on deep learning was discussed.On the basis above,the feature extraction and classifier model optimization methods based on deep learning were investigated.In this dissertation,4 efficient semi-supervised CTII methods were proposed and validated on the dataset of 4 kinds of communication signal collected practically.Through comparing the methods proposed with popular CTII methods,the feasibility and effectiveness of the methods proposed in this dissertation were validated.The main works of this dissertation is summarized as below:(1)Aiming at solving the “3S” problem under practical environment or at wartime,a communication transmitter individual identification method based on stacked autoencoders(SAE)is proposed.In the SAE method,a designed specifically stacked autoencoder network was utilized for compressing and encoding the massive unlabeled communication signal observations.The essential features that reflecting the communication individual internal structure information were extracted and passed to a softmax classifier model to improve the identification ability of the classifier,which was then trained and optimized by small amount of labeled communication signal observations.The experiment results on the dataset consisting of 4 kinds of communication signals collected from practical environment proved that the method proposed can efficiently solve the CTII task under “3S” prerequisite.(2)Aiming at the noise interference in CTII,a method based on denoising deep learning machine(DDLM)is presented.Firstly the communication signal observations were corrupted by the artificially injected noises.Then the DDLM had to force the autoencoders to reconstruct the original unpolluted observations,thus improving the generalization ability of the network on observations.Finally,the classification and identification phase was operated on the dataset consisting of 4 kinds of communication signals collected from practical environment.Compared with the SAE method above,the DDLM method extracted the features with stronger generalization performance and got significant promotion on identification ability.The correct ratio of classification on kenwood,krisun,USW,SW dataset was promoted by 11%,16%,25% and 22%,respectively.(3)The SAE method merely uses the unlabeled samples for the unsupervised training of stacked autoencoders,neglecting the effect of supervised training by labeled observations.To solve this problem,the graph embedding frame was introduced and,the CTII method based on graph embedding stacked encoders network(GAE)is presented.On the basis of compressive encoding the unlabeled input samples by SAE,the label messages in labeled samples were utilized for guiding the construction of the intra-class compact graph and the inter-class separability graph,which combining the encoding coefficients with discriminant information efficiently and improve the classification performances of the method.The feasibility and effectiveness of the DDLM method was validated by the experiments on the dataset consisting of 4 kinds of communication signals collected from practical environment.(4)In order to further promote the classification performance of DDLM,a CTII method based on denoising rectangular network(DRN)is proposed.The improvements of DRN are shown as followed:(1)Aiming at the problem that DDLM method only use the unlabeled samples for the unsupervised training on DDLM,the DRN is achieved by the parallel semi-supervised mode(nevertheless,the GAE was implemented by a serial semi-supervised mode).And that is to say,when training the denoising autoencoder through labeled samples,the stacked autoencoder was trained by the unlabeled samples at the same time.(2)In the decoder of the denoising autoencoder,the denoising decoding parameters obtained by supervised training were modified by the compressive decoding parameters obtained by unsupervised training.(3)A loss objective function compromising discriminant information specifically was designed to substitute the loss function that calculated by reconstruction error in DDLM method.The results of experiments on 4 kinds of communication signal dataset that collected from practical environment demonstrated that the DRN method achieved better classification performance than the DDLM method. |
PDF Full Text Request