Sequence Learning And Individual Identification Methods For Satellite Communications

Individual identification of communications radiation sources is at the forefront of electronic countermeasures research.In military applications,Specific Emitter Identification(SEI)plays an important role as one of the main technical tools for Measurement and Signature Intelligence(MASINT)reconnaissance activities.It plays an important role in communications signal signature mining,battlefield intelligence analysis,associated target tracking and identification,and battlefield situational awareness.At present,the radiation source identification technology in the field of satellite communications is not mature enough,some areas are still in the gap,lack of systematic and in-depth research.Traditional individual radiation source identification techniques in the field of communications rely on expert knowledge,which is highly complex and requires a priori knowledge as a known condition,thus limiting the scope of application.In this thesis,we use deep neural network technology to analyse the workflow of satellite communication systems according to the characteristics of satellite communication radiation source signals,and design an applicable neural network model to effectively shorten the evaluation cycle of target signals and improve the accuracy and reliability of individual identification of satellite communication radiation sources.The main work of this paper is as follows:1)Modelling the satellite communication system in transparent forwarding mode,analysing and discussing the signal characteristics and difference information that can be associated with a unique individual radiation source caused mainly by active devices such as high power amplifiers.A simulation environment will then be set up and link module parameters adjusted to effectively simulate a specific type of satellite communication and signal transmission process.2)Research on the design of deep learning networks based on Long Short Term Memory(LSTM)model.In this study,various data pre-processing methods and timing learning methods are investigated for simulated picking signals of the same type and different radiation sources using binary phase-shifted key modulation,and their effects on the network training process and classification results are analysed comprehensively.3)Using the characteristics of multi-domain signals and multimodal neural networks,we construct a deep fusion neural network model that can improve the signal feature mining ability;based on the measured and simulated data sets,we conduct training tests on various fusion network models,evaluate the comprehensive characterization ability of signal time-frequency features and higher-order spectral features,and finally provide an optimal solution to improve the recognition accuracy.4)Introducing the Transformer neural network architecture for signal analysis,exploring the feasibility of the self-attentive mechanism in extracting signal features,and evaluating the influence of communication link parameters on the analysis results of the network architecture in a simulation environment.The experiments demonstrate the superiority of the new network architecture,which can provide new ideas for deep learning methods of target recognition.