剩余使用寿命(Remaining useful life,RUL)预测是大型设备故障预测与健康管理(Prognostics and health management,PHM)的重要环节,对于降低设备维修成本和避免灾难性故障具有重要意义.针对RUL预测,首次提出一种基于多变量分析的时序图推理模型(Multivariate similarity temporal knowledge graph,MSTKG),通过捕捉设备各部件的运行状态耦合关系及其变化趋势,挖掘其中蕴含的设备性能退化信息,为寿命预测提供有效依据.首先,设计时序图结构,形式化表达各部件不同工作周期的关联关系.其次,提出联合图卷积神经网络(Convolutional neural network,CNN)和门控循环单元(Gated recurrent unit,GRU)的深度推理网络,建模并学习设备各部件工作状态的时空演化过程,并结合回归分析,得到剩余使用寿命预测结果.最后,与现有预测方法相比,所提方法能够显式建模并利用设备部件耦合关系的变化信息,仿真实验结果验证了该方法的优越性.
Remaining Useful Life Estimation of Facilities Based on Reasoning Over Temporal Graphs
Remaining useful life(RUL)estimation is an important component of the prognostics and health man-agement(PHM)of large-scale equipment,which is of great significance for equipment maintenance and avoiding catastrophic failures.In this paper,a multivariate similarity temporal knowledge graph model(MSTKG)is pro-posed for remaining useful life evaluation.The model can capture the dynamic information such as time-evolving changes in the coupling relationship and stability of the various components of the equipment,and mine the equip-ment performance degradation information accordingly.Firstly,a temporal and sequential graph structure is de-signed to capture the relationships among multiple sensors in adjacent time period,and to formally represent the evolving correlations among different components in continuous working cycle.Secondly,we propose a deep infer-ence network which integrates relation-aware convolutional neural network(CNN)and gated recurrent unit(GRU).The inference network explicitly learns the spatial and temporal evolution of the state of each component of the equipment.The results of remaining useful life estimation are obtained by integrating regression analysis.Finally,compared with existing estimation methods,the proposed method is able to explicitly model and utilize the chan-ging information of equipment component coupling relationships.Extensive evaluation results on benchmark show that the proposed model outperforms previous solutions on RUL estimation.
万方数据
维普
