Research And Verification Of Non-Cooperative Emitter Location Sensing Technology Based On Deep Learning

With the gradual implementation of national strategies such as "Belt and Road" and "Going Global",Our country has put forward more urgent requirements for the control capability of the electromagnetic spectrum at all times and in all regions.Compared to ground-based monitoring systems with limited perception range and traditional large satellite monitoring systems with high costs and orbital saturation,constructing low-earth-orbit(LEO)electromagnetic monitoring constellations based on micro-nanosatellites,carrying out wideband spectrum awareness has become the optimal alternative.The key to controlling spectrum situations is maintaining the order of spatial frequency and discovering and eliminating abnormal frequency signals.Non-cooperative location(passive positioning)technology is the core of the above process.Due to the broad coverage of LEO constellations and the continuous growth of global electronic devices,the electromagnetic data faced by the system is developing rapidly towards a massive scale.The traditional system architecture of "onboard collection and processing" is limited by satellite payload performance and system replacement costs,making it difficult to adapt to an efficient non-cooperative location under massive electromagnetic data.Therefore,based on the system architecture of "onboard collection and ground processing," this paper focuses on the inefficiency of traditional methods for processing massive amounts of ground-based data.We conducted research on non-cooperative location technology based on deep learning and achieved the following results:1)In response to the low applicability and generalization of constellation positioning scenarios caused by the lack of frequency difference information and emitter-receiver location coordinates in the current research on source localization,we propose a deep learning-based single-target end-to-end non-cooperative location method.We effectively extract complex electromagnetic signals’ time and frequency difference information by simultaneously modelling the real and complex timefrequency data in all heterogeneous information of passive positioning problems using convolutional neural networks and fully connected neural networks,respectively.We also demonstrate the consistency between the proposed method and traditional direct positioning determination from a formula-inductive perspective and provide some theoretical basis for this class of methods.Experimental results show that under various signal-to-noise ratios and similar positioning error conditions,the proposed method achieves an average positioning time of 1% of the traditional direct positioning determination in LEO constellation positioning scenarios considering both time and frequency difference information and has a lower false alarm rate.2)In response to the multiple overlapping signals received by satellite receivers and dynamically changing numbers of satellites in the co-coverage area in actual positioning scenarios that have not been considered in current research,we propose a multi-target positioning method based on deep learning under the variable number of emitter-receiver conditions,which builds upon the foundation of single target positioning research.In this positioning scenario,zero-padding and sort-invariant training strategies are proposed to resolve the matching and permutation problems of the network’s inputs and outputs.Additionally,a dataset has been made publicly available to promote the development of deep learning techniques in this positioning scenario.Experimental results demonstrate that the proposed method reduces positioning error and time by 50% and 2%,respectively,compared to the traditional direct positioning determination in this positioning scenario.3)In response to the suboptimal positioning performance caused by difficulty in efficiently modelling frequency domain information with existing image processing networks as the basis for positioning models,we propose a model that combines selfattention mechanisms and a decreasing convolution structure.By utilizing the lowpass filtering properties of the self-attention mechanism and the high-pass filtering properties of the decreasing convolution structure,the model achieves efficient modelling of frequency domain information in the data.Additionally,the relationship between Transformer performance and the diversity of its multi-head attention distances is explored,providing new ideas for improving its performance further.Experimental results demonstrate that in an image classification task,the proposed method achieves the same performance as the baseline model but with 29% and 57%fewer parameters and computations,respectively.In a multi-target positioning task,compared to the optimal method this study proposed above,the proposed method reduces positioning time by 49% with a similar positioning error,further improving the positioning performance.