针对炮位侦察校射雷达在弹丸上升段长时间预测落点偏差大的问题,文中提出了一种基于长短时记忆(LSTM)网络的弹丸长时间落点预测模型修正算法。针对训练样本少的问题,采用知识驱动和数据驱动相结合的双驱动方法,结合传统无迹卡尔曼滤波预测作为先验知识,利用LSTM网络对射向和侧向偏差量进行学习拟合构建误差模型,并利用误差模型对外推弹道进行射向和侧向修正。仿真验证表明:采用少量训练样本就可显著提升弹丸长时间落点预测的精度,落点预测射向精度提升55%,落点预测侧向精度提升85%~91%。

Aiming at the large deviation problem in impact point prediction during the rising period of fire locating radar, a novel long-term projectile impact point prediction correction method based on long short-term memory (LSTM) network is proposed. Due to the small sample problem, a dual-drive method combining knowledge-driven and data-driven method is exploited, with unscented Kalman filter predictions as prior knowledge and the LSTM network to learn the longitudinal and lateral deviations and then correct the impact point of extrapolated trajectory. Simulations show that the accuracy of long-term impact point prediction is improved significantly with a small number of training samples. The longitudinal prediction accuracy is increased by 55%, and the lateral accuracy is increased by 85% to 91%.

TN957.52