摘要
飞行动作识别及其对应飞参数据的提取是飞行训练质量分析的关键内容。现阶段的飞参数据量大、维度高、冗余数据多,为此提出了无监督聚合动态时间规整(UADTW)算法,降低动态时间规整(DTW)算法复杂度,帮助人工快速建立样本集,并提取标准序列相关性特征。同时,根据复杂飞行动作特点,构造深度神经网络模型学习飞行动作序列特征、差量特征和标准序列相关性,并在此基础上设计了自选择特征层,提出自选择深度神经网络(SDNN)模型,该模型能够自主选择对飞行动作识别较大贡献的特征,进一步提高特征表示对飞参数据的刻画。本文所提出的UADTW和SDNN飞行动作提取及识别方法能够减少人工成本,并有效提升了飞行动作识别的准确率。
Aircraft flight action recognition and its corresponding flight parameter data extraction are the key contents of flight training quality analysis.At present,the flight parameter data has features of big scale,high dimension and big redundancy.Therefore,this paper proposes an unsupervised aggregation dynamic time warping algorithm(UADTW)to reduce the complexity of DTW algorithm,help manual establish the sample data set quickly and extract the correlation characteristics of standard sequence.At the same time,according to the characteristics of complex flight action,a deep neural network model is constructed to learn the characteristics of flight action sequence,the difference characteristics and the correlation characteristics of standard sequence.Based on the deep neural network model,this paper designs a self selection feature layer and proposes a self-selective deep neural network(SDNN)model,which can independently select the features that contribute greatly to flight action recognition and improve the characterization of flight parameter data by feature representation.The practical application shows that the method of flight action extraction and recognition based on UADTW and SDNN can reduce the labor cost and effectively improve the accuracy of flight action recognition.
作者
李超
张原
汲万峰
司晓锋
李璇
Li Chao;Zhang Yuan;Ji Wanfeng;Si Xiaofeng;Li Xuan(School of Aviation Fundamentals,Naval Aviation University,Yantai 264000,China)
出处
《航空兵器》
CSCD
北大核心
2023年第1期127-134,共8页
Aero Weaponry
基金
国家自然科学基金项目(62076249)
山东省自然科学基金项目(ZR2020MF154)
山东省重点研发计划项目(2020CXGC010701,2020LYS11)。
关键词
飞行动作
自选择机制
动态时间规整
神经网络
飞参数据
aircraft flight action
self-selective mechanism
dynamic time warping
neural network
flight parameter data
