基于知识图谱多集池化的健康状态智能评估方法

Health intelligent evaluation based on knowledge graph multi-set pooling

为了从装备传感器监测数据和其他关联数据中提取更全面的时间域和空间域特征信息,提出一种基于知识图谱多集池化的健康状态评估方法。构建了带时间标签的健康知识图谱,以建模装备一段时间内监测数据、部件组成数据和先验知识间的时空依赖关系。在此基础上,设计了图多集池化网络模型,该模型通过节点特征学习、第一级图池化、自注意力特征学习和第二级图池化能够生成图谱的整体向量表示,将健康状态评估转换为基于表示学习的图谱分类任务。在公开的发动机数据集上对所提方法进行了实验评价,结果表明,该方法能够获得较高的评估准确度,在小样本情况下也表现出良好的优势。

To extract more comprehensive features in time domain and space domain from equipment sensor monitoring data and other related heterogeneous data,a health intelligent evaluation method based on knowledge graph multi-set pooling was proposed.A Health Temporal Knowledge Graph(HTKG)was constructed to fuse spatiotemporal features of monitoring data,component data and priori knowledge.The overall spatiotemporal features of HTKG were embedded into graph-level representation vectors with a graph pooling network,which included node feature learning,first level graph pooling,self-attention feature learning and second level graph pooling.The health evaluation was transformed into a graph classification problem based on representation learning.The proposed method had been evaluated on public engine datasets.Experiments results showed that the method could achieve high evaluation accuracy and also shows good stability in few-shot situations.