Deep Learning Methods For Space-Based Electromagnetic Signal Recognition

Electromagnetic space is the main channel for information generation transmission,perception,and utilization,and it is also an important support for new forms of combat in the era of information warfare.The space-based electromagnetic situational awareness system acquires electromagnetic signals such as radar,communication,navigation and data link in a large-scale space for situation analysis and prediction,which helps to realize the reconnaissance,surveillance,tracking and identification of various radiation sources of land,sea,air and space,so as to realize the control of the whole electromagnetic space situation.With the rapid growth of space electromagnetic equipment,electromagnetic signals are becoming more and more dense.In the face of complex and diverse electromagnetic targets,traditional methods based on target modeling and artificial features are becoming more and more difficult to deal with.Data-driven deep learning has powerful feature expression capabilities and is very suitable for dealing with big data problems.Electromagnetic signal recognition is the core component of electromagnetic situational awareness.Its goal is to obtain the modulation of radiation source and individual identity.Electromagnetic signal recognition based on deep learning has important research value and application prospects.There are many different types of electromagnetic signals in space.If the respective identification systems are designed according to different frequency bands,different types of radiation sources and different applications,the engineering application cost is high.This paper regards automatic modulation recognition and radio frequency fingerprint recognition as a continuous signal pattern recognition problem,using an external perspective,taking the overall external behavior of the transmitter as the research object,and learning the recognition model from the transmitter output signal to realize electromagnetic signal recognition.Based on this design,the electromagnetic signal intelligent processing architecture,IQ-related feature convolution network structure,and compressed signal intelligent processing framework can handle two applications: automatic modulation recognition and radio frequency fingerprint recognition.Although deep neural networks are good at mining the complex features of high-redundancy data,they also have the disadvantages of large models and high computational complexity.Due to the limitations of power consumption and space radiation environment,the storage space and computing power of space-based platforms are particularly limited.In order to bridge the gap between space-based platform performance and deep learning resource requirements,a lightweight intelligent processing architecture is proposed.The data stream is first split into small data segments,and then directly processed by convolutional neural networks,and finally information fusion combined by sequence joint decision algorithm.Processing the same length signal,the parameter amount and calculation amount of the lightweight intelligent processing architecture are only 0.037% and 40% of Alex Net,and 0.170% and 14.3% of Google Net.The overall performance of the lightweight intelligent processing architecture depends on a single classifier.In order to achieve efficient use of information,an IQ-related feature convolution network structure is proposed.In the application of automatic modulation recognition,when signal-to-noise ?0d B,the recognition accuracy can be increased by 9.88% with only 50% of the calculation amount,and the recognition accuracy is 37.63% higher than that of the traditional high-order statistics method.In the application of radio frequency fingerprint identification,the accuracy rate can be increased by 12.28% with only 50% of the calculation amount,and the accuracy rate is 33.16% higher than the traditional power spectral density method.Communication and radar signals continue to expand to high frequency bands,the sampling pressure is increasing with signal bandwidth rapidly increasing.Facing the problem of wideband signal recognition,a compressed signal intelligent processing framework combining compressed sensing and deep learning is proposed.In automatic modulation recognition and radio frequency fingerprint recognition,even if the sampling rate is reduced by 16 and 64 times respectively,the same recognition accuracy rate of the Nyquist sampling signal can be achieved,which greatly broadens the recognition range of the signal.Finally,based on a lightweight and intelligent processing architecture and a low-cost signal acquisition and processing platform,a sub-Nyquist sampling signal radio frequency fingerprint recognition real-time processing system with a close to100% recognition accuracy is realized.