Data-Driven Real-Valued Timed-Failure-Propagation-Graph Refinement for Complex System Fault Diagnosis

Timed Failure Propagation Graphs (TFPGs) have been widely used for the failure modeling and diagnosis of safety-critical systems. Currently most TFPGs are manually constructed by system experts, a process that can be time-consuming, error-prone, and even impossible for systems with highly nonlinear and machine-learning-based components. This letter proposes a new type of TFPGs, called Real-Valued Timed Failure Propagation Graphs (rTFPGs), designed for continuous-state systems. More importantly, it presents a systematic way of constructing rTFPGs by combining the powers of human experts and data-driven methods: first, an expert constructs a partial rTFPG based on his/her expertise; then a data-driven algorithm refines the rTFPG by adding nodes and edges based on a given set of labeled signals. The proposed approach has been successfully implemented and evaluated on three case studies.