基于注意力机制的雷达信号分选算法

Radar Signal Sorting Algorithms Based on Attention Mechanisms

针对传统雷达信号电子侦察先分选、再融合、后识别的流程繁琐且低效,本文提出直接对疑似敌方雷达的某个脉冲构建单脉冲特征矩阵,基于注意力机制与卷积神经网络(CNN)对其进行高相关脉冲的挑选和脉间调制类型识别.对挑选的高相关脉冲和识别的脉间调制类型结果进行分析,提取出脉冲间特征以及脉冲特征的相对关系,进一步完成后续分选操作.注意力机制分特征空间和脉冲重复间隔(PRI)时序两个层面.特征空间层面以待识别脉冲为焦点,利用位置分布特征和各分布位置脉冲数统计特征提取高相关性脉冲,简化单脉冲特征矩阵,完成预分选.PRI时序层面先通过CNN判断脉冲序列是否为混合型,若为混合型,则存在干扰脉冲,以待识别脉冲为焦点进行二维变换,分析其分布规律,根据杂散程度去除干扰脉冲.由于神经网络训练慢且不同脉间调制类型的脉间特征相对关系是已知的,采用有监督学习方式,提前用CNN对不同脉间调制类型进行学习,从而达到脉间类型识别的目的.通过对常规、抖动、参差、脉组变频、捷变频、脉宽捷变、线性滑变这7种脉间调制类型雷达信号进行分选仿真,验证了基于注意力机制的雷达信号分选算法对同时到达多目标雷达信号进行先识别后分选的可行性.仿真结果表明,该方法较传统信号主+预分选方法更加高效,且识别正确率更高.

For traditional electronic radar signal detection,the process of sorting,fusion,and recognition is cumbersome and inefficient.This paper proposes to construct a monopulse feature matrix directly for a certain pulse of suspected enemy radar.This matrix will select the high-correlation pulses and recognize the type of inter-pulse modulation based on the attention mechanism and the convolutional neural network(CNN).The inter-pulse features and the relative relationship between pulse features are extracted through the analyses of selected high-correlation pulses and recognition results of inter-pulse modulation types.Then,the subsequent sorting operation is completed.The attention mechanism has two aspects:feature space and pulse repetition interval(PRI).On the level of feature space,with a focus on the pulses to be identified,the high-correlation pulses are extracted using position distribution characteristics and the quantitative statistical features of each distribution position in order to simplify the monopulse feature matrix and complete the pre-sorting.At the PRI level,it is determined whether the pulse sequence is mixed through the CNN;in such cases,there are interference pulses.With a focus on the pulses to be identified,a two-dimensional transformation is conducted to analyze the distribution of pulses.Then,the interference pulses can be removed according to the degree of dispersion.Since the training of the neural network is slow and the relative relationship between the inter-pulse feature of different inter-pulse modulation types is known,the supervised learning method is used to learn different inter-pulse modulation types with the CNN in advance so as to achieve the goal of inter-pulse modulation identification.Through the sorting simulation of seven types of inter-pulse modulation radar signals the stable PRI,the jittered PRI,the staggered PRI,pulse group frequency conversion radar,frequency agile radar,pulse width agility radar,and the slip PRI-it is verified that the radar signal sorting algorithm based on the attention mechanism is feasible for recognizing and then sorting radar signals that simultaneously arrive at multiple targets.The simulation results demon-strate that this method is more efficient than the traditional methods of main-and pre-sorting,and the accuracy of recognition is higher.