ENABLING TRUST IN ENGINE HEALTH MONITORING THROUGH MODEL ENSEMBLES

Rotorcraft turboshaft engines within the Navy and DoD fleet are frequently exposed to harsh environments, particularly during takeoff and landing, which contribute to degraded engine health.   Existing modeling approaches of engine health are highly sensitive to the operational conditions on which they are trained while individual models have varying degrees of accuracy for a given set of operating conditions.  To circumvent these shortcomings CFD Research has identified the integrated multimodel scheme proposed here. CFD Research, in collaboration with University of South Carolina, will address each challenge encountered during this effort through the development of Data-Driven Monitoring and Maintenance (D2M2) engine software, with a focus on Navy-relevant engine platforms.  A prototype ensemble of predictive models for remaining useful life will be defined during Phase I by leveraging benchmark datasets of simulated engine health to be generated by the NASA C-MAPSS code. The accuracies and inaccuracies of the individual models will be improved through a graph-based hybrid approach capable of autonomously deciding which models are most trustworthy for a given set of operating conditions. Uncertainty quantification and model justification techniques will also be investigated. Additional health monitoring metrics such as anomaly detection and fault prediction will be explored during the Phase I Option and beyond.  D2M2 capabilities for any Navy provided dataset will also be investigated thereby demonstrating the broad applicability of D2M2. A fully functional prototype of D2M2, including graphical user interface with continuous model refinement through a robust data acquisition and active learning pipeline will be developed by the end of Phase II.