Machine Learning-Based Predictive Health Model of Turbofan Engine

Abstract

Turbofan engine is one of the major elements providing power and thrust for aircraft. Maintaining the engine is vital for both safety and economy of aircraft operation. Besides, for an engine original equipment manufacturer, aftermarket services take account approximately 60% of company's revenue. Furthermore, predictive maintenance can reduce up to 30% of unscheduled removals which will lead to extended on-wing time for airliners. In this context, the impact of predicting in-service engine performance or deterioration has been of interest for many years. The idea is also aligned with digital twins supporting decision making for engine maintenance, repair, and overhaul. And it is not limited to safety improvement, engine life extension, efficiency in maintenance, and financial benefit. However, there are practical challenges in pursuing a predictive engine health model for in-service engines. One of these challenges is that it is difficult to consider all factors affecting the engine performance while building the model. Because each engine experiences a variety of environments as it travels worldwide. The emergence of artificial intelligence and machine learning (ML) opens opportunity to circumvent those challenges by building a surrogate model over data. For in-service turbofan engines, there is engine health monitoring (EHM) system that collects sensor signals during flights. EHM system provides limited but numerous information such as pressure, temperature, vibration, fuel flow, aircraft data, etc. It becomes useful data source as it conveys engine operational data for over its life cycle spanning over 20 years. The aim of this research is to build a framework of predictive engine health model applying ML to EHM data, by evaluating various approaches in each process of building a ML-based model. Turbine gas temperature was target variable as it is one of useful health indicator correlated with engine life. Additionally, it aims to provide insights and ideas that are deemed valuable to share within this community as a further contribution. The observation on EHM data provided a useful insight on its characteristics related to engine maintenance interval. Based on the observation, the new training approaches were proposed using data segmentation based on time between major overhaul. The proposed approaches were to use much less data (up to 65%) than a conventional train-test split method for model training. A sensitivity study to select predictor variables were conducted after Pearson covariance coefficient analysis. The models were built with long short-term memory (LSTM) network, however, linear and nonlinear regression algorithms were compared. Missing values were identified from the data observation and several missing value imputation (MVI) methods were discussed to evaluate the impact of those methods. One of the MVI methods used a physics-based engine performance model embedding the performance maps of engine components such as compressors and turbines. Furthermore, cross-engine Transfer Learning approach was proposed to explore a scalability and versatility of large generalized model. The results demonstrated that the predictive model using LSTM showed robust performance considering its prediction accuracy and number of outliers. There were promising results among linear and nonlinear regression algorithms. Furthermore, the proposed training approaches by the data segmentation based on time between major overhauls achieved prediction accuracy with a RMSE value of 4--6 C, even with 65% less amount of data than train-test split method. It was shown that a large generalized model with 45--300 times more data enhanced robustness of the predictive model by substantially reducing outliers and variability. Among various MVI methods, using the engine performance model showed the most improvement in prediction, 36% better than remove rows method. The detrending with extended Kalman filter significantly improved prediction accuracy across all algorithms and training approaches, with improvements reaching up to 87% in mean RMSE value, or 1.7C. The results of the Transfer Learning across different engines implied that the one ML model can be employed to different family engines when there is a commonality between them. Finally, to enhance the model's reliability for decision-making in safety-critical applications, uncertainty quantification was performed. An evaluation of Delta and Bayesian methods showed that the Delta method provided robust prediction intervals with high Prediction Interval Coverage Probability, demonstrating its effectiveness in quantifying prediction uncertainty.