Towards Generalisable Predictive Health Models For Diverse Turbofan Engines Via Knowledge Transfer

Abstract Predictive models for aircraft engines are being developed to forecast engine health conditions based on available operational data. In-service engine data will be the most valuable information to build such a model but may not always be accessible or sufficient. A pre-trained model can be a possible alternative, if it can be extended to new and other engine families. The objective of this paper is to demonstrate knowledge transfer by evaluating a predictive engine health model across different applications of engines. A predictive engine health model was developed using Machine Learning techniques applied to Engine Health Monitoring (EHM) data gathered from the fleets of three different turbofan engines. Each model was built exclusively by training on the EHM data from each fleet of turbofan engines. The pre-trained models made predictions for their source (baseline case), and for other fleets of turbofan engines (transfer knowledge cases). The results show that the transfer knowledge cases have mean Root-Mean-Square-Error values, between 5°C and 8°C, which is about 2 times higher than the baseline case. The baseline case, represented by a smaller interquartile range (IQR), between 0.4°C and 1.6°C, have less variation in the prediction results than that of the transfer knowledge cases which have IQR values ranging from 1.6°C to 6°C. Such results indicate a generative predictive engine health framework may be developed with the capacity to be scaled across multiple gas turbine engine classes and applications.