| Περίληψη: | In the context of a rapidly changing and fiercely competitive global economy, aeronautical engineering endeavors towards enhancing aircraft safety, reliability and extending its operational lifetime have become of vital importance. To this end, Prognostic and health management (PHM) is deemed as key to revolutionize the maintenance paradigm in this sector. The most critical step in PHM is the remaining useful life (RUL) prognosis; a challenging task that demands special attention. In this work, a novel two-stage data-driven framework for direct RUL prediction of complex aeronautical elements is proposed, combining advanced health indicators and ensemble learning (EL). Multi-stiffener composite panels (MSPs), subjected to variable amplitude compression-compression fatigue, provide the necessary condition monitoring data, in the form of strain measurements. The latter are utilized to construct several damage sensitive features, i.e., health indicators (HIs) that properly capture the degradation process of the MSPs. A feature-level fusion of the aforementioned HIs is performed, utilizing Genetic Algorithms to create an enhanced HI, in terms of monotonicity and prognosability attributes. An upscaling methodology is attempted, by exploiting data from single-stiffener panel (SSP) histories, in order to estimate the advanced HI and eventually the RUL of the MSPs. The proposed framework involves the formulation of an ensemble approach capable of aggregating the RUL predictions in an adaptive way. The latter is obtained by training diverse sub-models built upon each SSP, and then combining the RUL predictions with a dynamic weighting strategy, based on Fuzzy Similarity Analysis (FSA). Two data-driven models, namely Support Vector Regression (SVR) and Long Short-Term Memory Network (LSTMN) are considered as the regression technique to map the input data to its corresponding RUL output. A comparison between the aforementioned machine learning (ML) and deep learning (DL) ensembles demonstrates the capabilities of the proposed framework. Ultimately, prognosis depends heavily on the HI behavior and trend, while confidence intervals get narrower as more condition monitoring data is incorporated, an essential trait of a robust prognostic algorithm. However, it is observed that no regression model significantly outperforms the other, since both SVR and LSTMN achieved better performance for different MSPs.
|
|---|
