Recognition of precession angles of non-cooperative targets based on deep learning with privileged information

With an increasing number of countries engaging in space activities worldwide, space non-cooperative target tracking and identification technology has become a prerequisite for safely conducting space operations. In order to the identify distant non-cooperative targets performing complex motions, this paper proposes a method to recognize difficult parameters by using easily available signal labels as privileged information, which is named Pi-FcResNet. The privileged information is connected to the output end of the network through a fully connected network and coupled with the linear layer of the main network. Through testing, our network achieved a recognition accuracy of 94.45 % for precession angles under high signal-to-noise ratio conditions. After incorporating the Convolutional Block Attention Module (CBAM), our method demonstrates fast fitting speed and robust performance. Testing on experimental data shows that, compared to traditional methods, our approach offers better stability and reproducibility in recognizing micro-motion parameters. This approach of using known information as additional information for deep learning networks holds great potential in the field of feature extraction for space non-cooperative targets undergoing complex motions.