基于多层LSTM模型的雷达目标航迹快速识别算法

Rapid Identification Algorithm of Radar Target Track Based on Multi-layer LSTM Model

雷达目标航迹的快速识别对指挥员战场决策具有重要的参考作用,传统雷达目标航迹识别算法对于目标特征分析效果差,导致航迹识别效果不理想,为此,设计了基于多层长短期记忆网络(Long Short-Term Memory, LSTM)模型的雷达目标航迹快速识别算法。对雷达目标航迹信息进行采集与去噪处理;构建多层LSTM模型,提高对时间序列数据处理的性能,将采集的数据输入多层LSTM模型中;通过多层LSTM网络自主学习获取雷达目标航迹特征,并设计融合模块对多个特征进行融合处理,得到多特征子集,改善单一特征分析的不足;基于适应性矩估计(Adaptive Moment Estimation, Adam)算法优化模型超参数,训练损失函数,通过构建多层LSTM模型分类器完成雷达目标航迹快速识别。仿真实验结果显示,该算法能够精准提取雷达目标的多特征信息,多特征融合效果良好,航迹识别结果精准,目标位置平均识别误差为0.31 m,雷达目标航迹识别时间平均值为0.56 s,说明该方法能够精准、快速完成航迹识别。

The rapid identification of radar target track plays an important reference role in the battlefield decision of commanders. The traditional radar target track recognition algorithm has poor effect on target feature analysis, resulting in unsatisfactory track recognition effect. Therefore, the radar target track recognition algorithm based on multi-layer Long Short-Term Memory(LSTM) model is designed. Firstly, the radar target track information is collected and denoised. Secondly, a multi-layer LSTM model is constructed to improve the performance of time-series data processing, and input the collected data into the multi-layer LSTM model. Then, the multi-layer LSTM network is independently learned to obtain the radar target trace features, and the fusion module is designed to integrate the multiple features to obtain the multi-feature subset, which improves the shortcomings of the single feature analysis. Finally, based on Adaptive Moment Eestimation(Adam) algorithm, the model hyperparameters are optimized, the loss function is trained, and radar target track recognition is completed by constructing multi-layer LSTM model classifier. The simulation experiment results show that the algorithm can accurately extract the multi-feature information of the radar targets;the multi-feature fusion effect is good;the track identification results are accurate;the average recognition error of the target position is 0.31 m, and the average radar target track recognition time is 0.56 s, which shows that the method can complete the track identification accurately and quickly.