Abstract: To exploit the temporal and spatial features of the sensor monitoring signals,which can make a comprehensive reflection of health state and improve failure prediction accuracy,an aero-engine RUL prediction method based on an Adaptive Spatial-Temporal Graph Convolution Network (ASTGCN) was proposed.An adaptive adjacency matrix was built by combining a static adjacency matrix based on mutual information and a dynamic adjacency matrix with learnable parameters,which could automatically adjust the spatial interconnection of sensor nodes to achieve high-quality construction of graph-structured data in the aero-engine health monitoring scenario.A spatial-temporal graph convolutional network was established to capture the dynamic spatial-temporal correlation of monitoring data through synchronous learning of the temporal and spatial dependencies of the monitored signals based on the one-dimensional and graph convolutional networks respectively.Finally,the fully connected layer was adopted for degradation feature fusion and accurate prediction of RUL.The public aero-engine degradation dataset was employed to validate the effectiveness and superiority of ASTGCN.

Key words: aero-engine, remaining useful life prediction, data-driven, spatial-temporal graph convolutional network, adaptive adjacent matrix

摘要： 为了深入挖掘传感器监测信号的时间域和空间域特征,全面反映健康状态进而提高故障预测精度,提出一种基于自适应时空图卷积网络(ASTGCN)的航空发动机剩余使用寿命(RUL)预测方法。首先以基于互信息的静态邻接矩阵为基础,结合参数可学习的动态邻接矩阵表示方法建立自适应邻接矩阵,自动调整传感器节点的空间关联,高质量构建航空发动机健康监测场景下的图结构数据;其次建立时空图卷积网络模块,分别利用一维和图卷积网络同步学习监测信号的时间和空间依赖关系,捕捉监测数据的动态时空相关性;最后将全连接层用于退化特征融合和RUL预测。采用公开的航空发动机退化数据集验证了ASTGCN的有效性和先进性。

关键词: 航空发动机, 剩余使用寿命预测, 数据驱动, 时空图卷积网络, 自适应邻接矩阵