摘要
智能制造、航天探测等领域中的传感器产生了大量的时序数据,这些数据中记录了设备的运行状态信息。如何有效地利用传感器的历史数据对未来时刻的行为进行预测,以便进一步对设备实现维护是非常重要的。在以往的预测方法中,往往只对多元时序数据在时间维度的关联关系进行建模,忽视了多元时序数据在空间维度的依赖关系。为了充分地对多元时序数据在时间和空间两个维度上的特征信息进行提取,从而实现更精准的预测。论文提出了一种基于CNN和GAT的多元时序数据预测方法。分别利用CNN和GAT来提取多元时序数据在时间维度和空间维度的信息,并综合利用提取到的信息来进行预测。在两个真实传感器数据集上的实验表明,论文提出的方法相比于传统的预测方法有着更低的LOSS值。
Sensors in smart manufacturing,aerospace exploration and other fields generate a large amount of time-series da-ta,which records information about the operating status of the equipment.It is very important to use the historical data of sensors ef-fectively to predict the behavior of future moments in order to achieve further maintenance of the equipment.In previous prediction methods,only the correlations of multivariate time series data in the temporal dimension are often modeled,ignoring the dependen-cies of multivariate time series data in the spatial dimension.In order to fully extract the feature information of multivariate time-se-ries data in both temporal and spatial dimensions,so as to achieve more accurate prediction.In this paper,a prediction method based on CNN and GAT for multivariate time-series data is proposed.CNN and GAT are used to extract the information of multivari-ate time-series data in temporal and spatial dimensions respectively,and the extracted information is combined to make predictions.Experiments on two real sensor datasets show that this proposed method has a lower LOSS value compared to traditional prediction methods.
作者
王玉乾
陈雷鸣
董次浩
陈涛
赵宏伟
张宝宇
WANG Yuqian;CHEN Leiming;DONG Cihao;CHEN Tao;ZHAO Hongwei;ZHANG Baoyu(China University of Petroleum(East China),Qingdao 266580)
出处
《计算机与数字工程》
2025年第8期2181-2187,共7页
Computer & Digital Engineering
基金
国家自然科学基金项目(编号:62072469)
中国科学院自动化研究所复杂系统管理与控制国家重点实验室2021年开放课题(编号:20210114)资助。
