Failure Prognostics using Multi-fidelity Graphic Learning for Enhanced Complex System Resilience

This paper tackles the issues in combining sparse, heterogenous data in order to approximate a desired output. The focus of this paper is on predicting the state of health of systems to determine when a fault will occur. The approach used utilizes a graphic Directed Acyclic Graph (DAG) in a Bayesian Network scheme to relate the information sources. This method can incorporate data with different fidelities as well as units that are related to the desired output to be predicted and detect the underlying relationship between them. This method is very useful in complex systems where the system state is unobservable or where the data is severely sparse. The proposed online learning scheme was tested on a simple math example and was shown to provide accurate predictions of the high-fidelity function given very limited data. The results show that the online algorithm outperforms the all-at-once method for different initial beliefs and maintains a lower mean square error.