| 引用本文: | 杨任农,房育寰,张振兴,等.变分自编码器结合聚类算法在空战态势评估问题上的应用.[J].国防科技大学学报,2019,41(4):144-155.[点击复制] |
| YANG Rennong,FANG Yuhuan,ZHANG Zhenxing,et al.Application of variational autoencoder combined with clustering algorithm in air combat situation assessment[J].Journal of National University of Defense Technology,2019,41(4):144-155[点击复制] |
|
| |
|
|
| 本文已被：浏览 7612次 下载 5655次 |
| 变分自编码器结合聚类算法在空战态势评估问题上的应用 |
| 杨任农1, 房育寰2, 张振兴1, 左家亮1, 黄震宇1, 张滢1 |
|
(1. 空军工程大学 空管领航学院， 陕西 西安 710051;2. 中国人民解放军95939部队， 河北 沧州 061736)
|
| 摘要: |
| 针对传统态势评估方法确定权值的主观性强、处理大数据能力弱、特征提取能力不足等问题，提出基于改进变分自编码器和聚类算法的无监督空战态势评估方法。根据态势变化连续性特点，提出基于时间段的空战态势分类方法，将敌我双方态势划分为四类。在变分自编码器的基础上，提出了VAE-WRBM-MDN特征提取模型，即使用混合密度网络优化变分自编码器的特征提取能力和生成数据的相似度，使用权值不确定限制玻尔兹曼机优化网络的初始权值。将提取的特征分别输入到两种典型的聚类算法中进行聚类，并结合态势函数和实际战场情况修正聚类结果，形成正确的态势分类标准。在实验部分，分别进行了最优参数调整、关键特征提取、聚类以及修正实验。实验结果表明，模型态势分类正确率和运行时间均满足应用需求，实例评估结果与客观态势一致性强，所提方法具有实际应用价值。 |
| 关键词: 态势评估 变分自编码器 混合密度网络 权重不确定限制玻尔兹曼机 聚类算法 |
| DOI：10.11887/j.cn.201904021 |
| 投稿日期：2018-04-19 |
| 基金项目:国家自然科学基金青年基金资助项目（71501184）；航空科学基金资助项目（201551960322）；陕西省自然科学基金资助项目（2016JQ6050） |
|
| Application of variational autoencoder combined with clustering algorithm in air combat situation assessment |
| YANG Rennong1, FANG Yuhuan2, ZHANG Zhenxing1, ZUO Jialiang1, HUANG Zhenyu1, ZHANG Ying1 |
|
(1. School of Air Traffic Control and Navigation, Air Force Engineering University, Xi′an 710051, China;2. The PLA Unit 95939， Cangzhou 061736, China)
|
| Abstract: |
| Aimed at the problems of traditional assessment methods, such as the subjectivity of determining the weights, the weak ability of processing big data and the lack of feature extraction ability, an improved air combat situation assessment method based on VAE (variational autoencoder) and clustering algorithm was proposed. Firstly, according to the characteristic of continuity of situation changes, a situation classification method based on time period data was proposed, and the situation of both sides was divided into four categories. Then, on the basis of VAE, a VAE-WRBM-MDN feature extraction model was proposed, which used the MDN (mixed density network) to optimize VAE feature extraction capability as well as the similarity of generated data, and to optimize initial weights of the network with WRBM (weighted uncertainty restricted Boltzmann machines). Finally, the extracted features were input into two typical clustering algorithms for clustering, and then the situational function and actual battlefield conditions were used to modify the clustering results, so as to forming a correct situation classification criteria. In the process of experiments, the optimal parameters adjustment, key feature extraction, clustering and correction were performed. Experimental results show that the model classification accuracy rate and the model runtime both meet the application requirements. In addition, the assessment results of the example are consistent with the actual situation. Therefore, the proposed method is of practical value. |
| Keywords: situation assessment variational auto encoder mixture density network weighted uncertainty restricted Boltzmann machines clustering algorithm |
|
|
|
|
|
