Deep Learning-based Algorithms For Satellite Reception Signal Classification And Target Identification

The classification of satellite reception signals and their target identification is an important research direction in cognitive electronic warfare.With the continuous development of satellite communication and reconnaissance networks in various countries,the classification and identification of target signals based on satellite reception data is becoming more and more important.Traditional deep learning signal classification and recognition algorithms are often vulnerable to channel interference,require a lot of manual annotation,and cannot adapt to the needs of unknown(new)target recognition.In order to overcome these shortcomings,this thesis investigates the problems of classification and target recognition algorithms for satellite received signals in practical engineering projects based on metric learning methods,resulting in the following main work and innovations:(1)A satellite reception target individual recognition algorithm combining deep learning with PCA dimensionality reduction and SVM clustering methods is proposed,which completes the overall process from the classification of the received signal regime to the individual recognition of the classified signals using individual recognition networks for the target communication signals received by the actual engineering satellites.Firstly,the signal is extracted using a feature residual network,then the signal is filtered by PCA dimensionality reduction and SVM clustering,and finally,according to the results of the clustering,the signal of interest for a specific regime is used as the data set for target identification using the individual identification network.It was tested that the clustering accuracy of the three signal regimes received by the satellite in this project exceeded95%,and the accuracy of the individual recognition of the eight targets of the signals of the regime of interest exceeded 95%.(2)A satellite received target open set identification algorithm based on Siamese networks is proposed.The algorithm addresses the scenario that the traditional individual recognition network cannot adapt to new targets,and gives a method to determine the judgment threshold for the corresponding case according to whether there is unknown class a priori information: when there is unknown class a priori information,the algorithm uses the spatial feature centre to find the sample metric distribution and uses this distribution to determine the unknown class judgment threshold; when there is no unknown class a priori information,the algorithm gives the unknown class judgment threshold by minimising the contrast loss.In both cases,the algorithm uses the spatial feature centres as anchor points for the Siamese network when identifying within known classes,in order to improve the known class judgement accuracy and stability at judgement.In the presence of unknown class a priori information,the optimised combined Siamese network achieves 97.76% for known class judgement and 95.98% for unknown class judgement,an improvement of7.94% and 3.84% respectively over the conventional Siamese network.Without the a priori information of unknown class,the accuracy of known class discrimination of the optimized combined Siamese network is 98.75%; the accuracy of unknown class discrimination is 84.46%,the accuracy of known class recognition is improved by 11.7%,and the overall accuracy is improved by 1.30% compared with the traditional Siamese network.(3)An experimental validation system for satellite reception signal classification and target individual recognition was built.The system was constructed through a graphical user interface for front-end display,and the back-end interface was used to invoke the main algorithm,containing three systems: the target signal processing subsystem,the target signal clustering subsystem and the target individual recognition subsystem,completing the overall process from signal pre-processing,signal regime clustering and target individual recognition,and realizing the visual application of the algorithm of this thesis in practical engineering.