Research On Technologies Of Communication Emitter Identification Based On Deep Learning

Communication emitter identification is a technology to extract the hardware information that can reflect the individual identity in the signal,and then identify different individuals.The existing methods based on deep learning mainly carry out supervised learning in the case of a large number of labeled samples,but in the actual scene,it will face the problem that the number of labeled samples is insufficient,and the deep learning network is difficult to fully extract effective features,so it is difficult to identify.In addition,since the bit information sent by the communication emitter is the main transmission information,and the individual characteristics of the emitter are slightly different,the neural network is vulnerable to the influence of the bit information when extracting features.In view of the above problems,this paper carries out the following research:(1)Aiming at the problem of small number of labeled samples,previous experiments have verified that the complex domain neural network is more suitable for small sample labeled data than the real domain network.So an algorithm is proposed,which combines the metric-based method in meta learning and complex domain convolution neural network.The feature extraction module in meta metric learning is improved into complex form to extract one-dimensional complex signal features and measure the similarity between them.Three kinds of network structures are improved:complex matching network,complex prototype network and complex relation network.The training strategy of meta learning is adopted to double sample the task and data.The generalization ability of networks learning when the data category changes is tested on a new category.Simulation results show that the algorithm effectively improves the network performance under the condition of small samples,and the recognition performance is improved compared with the complex convolution network and the meta metric method in the real field.(2)Aiming at the problem that the individual characteristics of communication emitter are difficult to extract and often be affected by sending information,a sending information assisted communication emitter individual recognition algorithm is proposed.This method combines recognition and demodulation.Assuming that the demodulation has been completed,the demodulated formed waveform data is input into the network as a priori knowledge.This paper studies the way of auxiliary information joining the network and data fusion,and designs two network structures:double branch network and complex relation network.Simulation results show that the performance of the network with auxiliary information is always better than that without auxiliary information,especially in the case of low signal-to-noise ratio.To some extent,the influence of sending information on the extraction of emitter features by neural network is reduced.In the experiment of adding auxiliary information of 1%and 1‰ bit error rate,it is found that even if there are certain demodulation errors,the proposed algorithm can still bring significant performance improvement,indicating that the algorithm is feasible in practical application.