基于自注意力机制增强的CNN-LSTM的榴弹轨迹多步超前预测

Multi-step-ahead Prediction of Grenade Trajectory Based on CNN-LSTM Enhanced by Deep Learning and Self-attention Mechanism

由于榴弹飞行轨迹呈现复杂性、时变性和突变性等特点,给近程防空拦截系统带来了极大的挑战。针对目前轨迹数据时空特征捕捉困难且只能进行较少步数预测的问题,提出一种引入自注意力机制的基于卷积神经网络和长短期记忆神经网络(1dimension Convolutional neural network-Long short-term memory-Attention, 1D CNN-LSTM-ATT)的一维轨迹多步超前预测模型。将所提模型与CNN-LSTM、LSTM模型分别进行单步和多步预测对比分析;实现对于目标轨迹的从T时刻到未来任意T+K时刻的高精度实时多步超前预测。实验结果表明:无论是单步还是多步预测,1D CNN-LSTM-ATT模型预测的评价指标明显优于其他2个模型;1D CNN-LSTM-ATT模型预测500步(即10 s)的累计预测误差在射程方向为82.83 m,高度方向为11.68 m,横偏方向为0.07 m,为实施弹体拦截及时响应提供了重要保障。

Since the grenade flight trajectory presents the characteristics of complexity,temporal variability and sudden change,it brings great challenges to the close-range air-defense interception system.Aiming at the current problem that the spatio-temporal features of trajectory data are hardly captured and only a small number of steps can be predicted,a trajectory multi-step-ahead prediction model based on 1-dimension convolutional neural network-long short-term memory-attention(1D CNN-LSTM-ATT)with the introduction of self-attention mechanism is proposed.The proposed model is compared and analyzed with CNN-LSTM and LSTM models in single-step and multi-step predictions,respectively,and realizes a high-precision real-time multi-step-ahead prediction of target trajectory from T moment to any T+K future moments.The study shows that the evaluation indexes of the 1D CNN-LSTM-ATT model are significantly better than those of the other two models for both single-step and multi-step prediction;the cumulative prediction error of 500 steps(i.e.,10s)of the 1D CNN-LSTM-ATT model is 82.83 m in the direction of range,11.68 m in the direction of altitude,and 0.07 m in the direction of transverse deviation,which provides an important guarantee of timely response to the implementation of the projectile interception.This provides an important guarantee for the timely response to the implementation of projectile interception.