Aircraft Engines Performances Estimation from Multi-Point and Multi-Time Operational Data via Neural Networks

Turbine engine monitoring is an essential enabler for predictive maintenance strategies. Among popular monitoring techniques, estimating performance indicators is a long-standing subject from which a variety of methods have been proposed— including recent applications of machine learning. Traditionally, indicators estimation is viewed under the single-point perspective: at one fixed operating condition, mapping observed measurement data to indicator values. In this perspective, while state-of-the-art methods have promising results in some favorable testing scenarios, none of them have completely resolved under-determined cases in which the number of indicators to-be-estimated is larger than the number of observed measurements.In this work, we study a use case in monitoring aircraft engines that falls into an under-determined setting. In order to mitigate this, we adapt a multi-point perspective that leverages measurements at multiple pre-determined steady-state operating conditions. It is noteworthy that we relax the condition-independence hypotheses made by previous multi-point models to better align with real scenarios. Unlike previous works under the multi-point perspective, we focus on machine learning methods as resolutions; in particular, we conduct a series of experiments with artificial neural networks in the considered use case. Our first contribution is to demonstrate that the estimation precision provided by tested models is significantly improved in the multi-point perspective in comparison with the single-point case. Secondly, we explore a scenario involving temporal data (i.e., data from consecutive flight missions) and show that neural networks with appropriate architectures can well exploit such "multi-time" data structure.