基于早期时间序列分类的可解释实时机动识别算法

An interpretable real-time maneuver identification algorithm based on early time series classification

战斗机机动识别是判断战斗机战术意图的基础,然而现有的机动识别方法实时性不强且不具有可解释性,无法满足空战中对实时性的要求且不利于人机互信。设计基于早期时间序列分类的实时机动识别算法,将完整机动切分为机动单元,使用集成学习算法对机动单元进行识别并实时监控,以满足实时性要求并获得高识别精度。算法使用可解释模型,通过特征贡献度进行模型解释,使模型更透明从而降低空战决策者的决策风险。选择盘旋、斤斗等9种不同机动动作进行仿真实验,结果表明:在完整机动动作执行到20%时,所提算法即可识别其机动类别,识别准确率可达93%。

The maneuver identification of fighter aircraft is the basis for judging their tactical intentions,but the existing maneuver identification methods have weak real-time performance and lack interpretability,which cannot meet the real-time requirements in air combat and are not conducive to human-machine trust.This paper designs a real-time maneuver identification algorithm based on early time-series classification,which divides the complete maneuver into maneuver units and uses ensemble learning algorithm to recognize and monitor the maneuver units in real-time,in order to achieve real-time requirements and obtain high recognition accuracy.The algorithm uses interpretable models and explains the model through feature contribution,making the model more transparent and reducing the decision risk for air combat decision-makers.Nine different maneuvers,such as hovering and jackknifing,are selected for simulation experiments,which proves that the algorithm can complete the identification with only the first 20%of the sample data of the time series observed,and the identification accuracy can reach 93%.