Radar target recognition method based on single-layer bidirectional recurrent neural network

Abstract

The invention discloses a radar target recognition framework based on a single-layer bidirectional recurrent neural network. The method comprises the steps of firstly, carrying out pretreatment to reduce the sensitivity in an HRRP sample; establishing a dynamic adjustment layer, and dynamically adjusting the sample; enabling a subsequent CNN layer to extract information contained in an HRRP envelope while reserving sequence correlation contained in the HRRP sample; carrying out time sequence correlation modeling through a bidirectional RNN, and extracting high-level features of the time sequence correlation; and finally carrying out target classification through a softmax function.

Description

technical field [0001] The invention belongs to the field of radar target recognition, and in particular relates to a radar target recognition method based on a single-layer bidirectional cyclic neural network. Background technique [0002] The high-resolution one-dimensional range profile (High Resolution Range Profile, HRRP) obtained by the broadband radar in the radar target reflects the distribution of the target scattering center along the radar line of sight, which contains rich target structure and shape information, so the target based on HRRP Identification has attracted extensive attention at home and abroad. Therefore, the HRRP-based radar automatic target recognition method has gradually become a research hotspot in the field of radar automatic target recognition. [0003] Traditional feature extraction methods have achieved good recognition performance in experiments, but there are some problems with their recognition methods, including: 1) The way of feature e...

AI Technical Summary

Problems solved by technology

[0003] Traditional feature extraction methods have achieved good recognition performance in experiments, but there are some problems in their recognition methods, including: 1) The way of feature extraction is mostly unsupervised and lossy, which means that the transformation-based feature extraction The method cannot focus on finding the maximum separable features, and the separable information will inevitably be lost in the process of feature extraction, which is not conducive to the identification of the back-end classifier

2) The selection of feature extraction methods is highly dependent on the researchers' knowledge and experience of HRRP data, and it is difficult to achieve satisfactory results in the absence of prior information

(3) Deep learning method based on recurrent neural network (RNN): This method is based on serial correlation modeling. Although the physical structure features are modeled and described, there are several problems as follows: (1) The distance with small amplitude The unit may contain some highly separable features, but these features are rarely used; (2) The input dimension and duration in RNN are entangled (for example: if HRRP is 1x256 dimension, choose the input dimension is 32, the overlapping area is 16, and the duration is 256 / 16-1=15), it is difficult to independently adjust these two important parameters; (3) the one-way RNN can only use the current moment and current The structural information before time cannot make good use of the overall structural information a priori contained in HRRP

Images