Convolutional neural network for classifying space target of the same shape by using RCS time series

Abstract

Warhead and decoy classification is one of the most important and difficult technical problems in ballistic missile defence. The conventional methods extract features from the measured data and employ some classification algorithms. However, it is hard to extract all the information embedded in the raw data, and there might be contradictory features lowering the classification ability. A one-dimensional convolutional neural network structure named RCSnet was proposed to classify the warhead and decoy targets of the same shape in midcourse, which directly utilises the radar cross-section (RCS) time series. It was compared with 5 conventional classification algorithms which used 26 selected features on simulation dataset, and it outperformed them in both classification performance and predicting speed. Different training algorithms and networks of the RCSnet structure with different filter numbers were explored for better utilising the RCSnet.