From fault tree to fault graph: Bayesian network embedding-based fault isolation for complex equipment

Abstract

Fault isolation (or fault location) aims to identify the component responsible for the defective behavior's symptoms, which is a significant part of predictive maintenance. However, it is difficult to sort out which component is accountable for the fault characteristics in complex equipment, due to limited space for sensor setup and its synergetic process. Different from fault tree, this paper proposes a large-scale fault graph to locate the fault components through probability inference, which leverages domain expertise. Specifically, a fault graph (type of knowledge graph) has been established firstly using hierarchical structure and domain-specific knowledge. Following, a Multi-field hyperbolic embedding method has been applied to vectorize the node and edge, maximally preserving the logical rules within the fault graph. Next, Bayesian network has been utilized to model the causality of fault graph with the well-trained graph embedding, then reasoning the possible fault component. Lastly, a maintenance case of oil drilling equipment is carried out, where the proposed method has been contrasted with other cutting-edge models to demonstrate its effectiveness in embedding properties and inference performance. This study is expected to provide insight into the potential application of knowledge graphs for fault isolation