Research On Radar Signal Sorting Algorithm Based On Machine Learning

Radar signal sorting is an important part of electronic warfare and is an important means of obtaining radar intelligence.With the increasing complexity of the modern electromagnetic environment,the performance of traditional signal sorting algorithms has declined,and it is increasingly unable to adapt to new task scenarios,and it is urgent to develop new methods to deal with complex sorting tasks.In view of this situation,this thesis starts from the characteristics of the interpulse parameter-pulse description word,integrates the traditional sorting algorithm into the machine learning framework,and designs a reasonable and effective radar signal sorting algorithm.This thesis focuses on the pre-sorting and main-sorting parts of signal sorting,with the following main research contents.(1)Aiming at the problems of uncertain number of radar emitters,complex distribution of pulse data and sensitivity to noise in pre-sorting,this thesis proposes a clustering sorting algorithm based on improved spectral clustering combined with data field theory,using the radio frequency,direction of arrival and pulse width in the pulse description word as pre-sorting parameters.The algorithm first uses the data field theory to pre-process the data,removes the interference points according to the potential value,and determines the initial number of clusters using the number of potential centers,then uses the grid density division to obtain reasonable landmark points,and finally completes the cluster sorting based on the spectral clustering via landmark-based sparse representation.Two sets of different types of pulse signal data are set up for simulation experiments,and the sorting accuracy is over 95%.Experimental results show that the proposed algorithm is superior to common clustering algorithms in accuracy and time efficiency.(2)In the mixed radar pulse sequences,the periodicity of the pulse repetition interval is broken and main-sorting cannot be performed well based only on the periodicity law implied in the arrival of time parameter.In this thesis,pulse amplitude and arrival of time are taken as the main-sorting parameters,and the neural network is applied to the mainsorting by using the sequence-to-sequence classification idea.A mixed signal dataset with fixed and sliding pulse repetition intervals is constructed,and the long and short-term memory network,the gated recurrent unit network and the bi-directional long short-term memory network are used as training models,while a comparison experiment with only arrival of time as sorting parameter is set up.The results show that the combination of pulse amplitude and arrival of time can improve sorting accuracy and also validate the applicability and effectiveness of the bi-directional long short-term memory network in the main-sorting task.