基于集成学习的遥测数据互相关结构知识发现

针对传统遥测数据相关性分析方法仅能发现相关程度知识，无法提供相关结构丰富信息的问题，提出一种神经网络与极限梯度提升（XGBoost）集成的遥测数据互相关结构知识发现方法。在对遥测时间序列进行线性、单调性、序对一致性、散点图形状4个维度相关结构信息标注的基础上，将混合采样、代价矩阵、神经网络、XGBoost算法相结合，直接对遥测数据进行分类得到其相关结构类别或相关关系有无的知识。采用量子卫星任务数据进行实验的结果表明:较之于原始XGBoost模型、融合混合采样与代价矩阵的XGBoost模型，所提方法在受试者工作特征（ROC）曲线、F1-score等性能指标方面具有更高的分类精度，且对类别不平衡数据不敏感，是一种适用于遥测数据互相关结构知识发现的有效方法。 

Knowledge discovery of telemetry data cross-correlation structure based on ensemble learning

Aimed at the problem that traditional telemetry data correlation analysis methods can only discover relevant degree knowledge and cannot provide relevant structural information, an extreme gradient boosting (XGBoost) and neural network ensemble learning method is proposed to discover the cross-correlation structural knowledge of telemetry data. Based on the dimension related structural information annotated by linearity, monotony, order pair consistency and scatter diagram shape, an algorithm combining hybrid sampling, cost sensitive matrix, neural network and XGBoost is developed to directly measure the telemetry data. The data is classified to obtain knowledge of relevant structural categories or related relationships. The results of experiments using quantum satellite mission data indicate that compared with the original XGBoost model, and the fusion-mixed sampling and cost-sensitive XGBoost model, the XGBoost model with neural network ensemble has higher classification accuracy on the performance indicators such as receiver operating characteristic (ROC) curve and F1-score. The proposed method is not sensitive to categorially imbalanced data, making it an effective method for the discovery of cross-correlation structural knowledge of telemetry data.