Graph neural networks (GNNs) can capture interdependencies between data with the structured data modeling ability, and have received much attention from industry professionals in remaining useful life (RUL) prediction tasks. However, the existing methods assume that graph nodes and edges are of the same homogeneous attributes, which leads to information loss and cannot fully capture the complex degeneration pattern and topological relationship of the bearings. To solve this problem, a novel heterogeneous graph representation-driven multiplex aggregation graph neural network is proposed for bearing RUL prediction. Different from the conventional methods based on homogeneous graphs, we model the heterogeneous attributes of bearing data and parameterize the representation of node relationships in heterogeneous graphs. The node adjacency is represented as the heterogeneity belonging to the designed spatial meta-path and temporal meta-path, respectively. In addition, a multiplex aggregation heterogeneous graph neural network (MAHGNN) is proposed to extract heterogeneous features of the graph as well as temporal dependencies of each node and achieve the bearing RUL prediction. In particular, a novel hierarchical aggregation mechanism for graph heterogeneous attributes is designed, which includes node-level aggregation, path-level aggregation and time-level aggregation. This mechanism can capture the diverse relationships and significance of various types of nodes and edges in heterogeneous graphs, so as to aggregate the feature information of nodes within a meta-path and different meta-paths as well as extract the temporal dependencies. The experiments conducted on two datasets provide evidence for the superiority of the proposed method in comparison to other state-of-the-art RUL prediction methods based on homogeneous graphs.

中文翻译：

---

异构图表示驱动的多重聚合图神经网络用于轴承剩余使用寿命预测


图神经网络（GNN）可以通过结构化数据建模能力捕获数据之间的相互依赖关系，并且在剩余使用寿命（RUL）预测任务中受到行业专业人士的广泛关注。然而，现有方法假设图节点和边具有相同的同质属性，这会导致信息丢失，并且不能完全捕获轴承的复杂退化模式和拓扑关系。为了解决这个问题，提出了一种新颖的异构图表示驱动的多路聚合图神经网络来承载 RUL 预测。与基于同质图的传统方法不同，我们对轴承数据的异构属性进行建模，并对异构图中节点关系的表示进行参数化。节点邻接性表示为分别属于设计的空间元路径和时间元路径的异质性。此外，提出了一种多重聚合异构图神经网络（MAHGNN）来提取图的异构特征以及每个节点的时间依赖性，并实现轴承RUL预测。特别是，设计了一种新颖的图异构属性分层聚合机制，包括节点级聚合、路径级聚合和时间级聚合。该机制可以捕获异构图中各种类型的节点和边的不同关系和重要性，从而聚合元路径内和不同元路径内节点的特征信息并提取时间依赖性。 在两个数据集上进行的实验证明了所提出的方法相对于其他基于齐次图的最先进的 RUL 预测方法的优越性。