Interpretable machine learning for multiparameter analysis of oil supply-scavenge dynamics in ventless bearing chambers

In order to address the persistent challenge of oil supply-scavenging matching in modern aero-engine bearing chambers, the study proposes a multiparameter coupled approach integrating experimental analysis and machine learning techniques. A dedicated test rig was developed for ventless configurations, operating at speeds up to 8000 r/min. Using an Emergency Cut-off Valve (ECV) method, oil residence volumes (Vre) were quantified to investigate the effects of rotational speed (ns), air flow rate (Va), water flow rate (Vw), and scavenge-to-supply ratio (SR). Dimensionless models were developed using fluid dynamics principles, incorporating the rotational Reynolds number (Reu), air supply Reynolds number (Rea), water supply Froude number (Frw), and SR. Spearman correlation analysis revealed complex nonlinear interactions among these parameters. Machine learning models, including Random Forest (RF), Gradient Boosting Decision Trees (GBDT), Extreme Gradient Boosting (XGBoost), and Generalized Additive Models (GAM), were compared, with XGBoost showing the best performance, achieving a coefficient of determination (R2) of 0.96 and a mean absolute percentage error (MAPE) of 4.53%. SHapley Additive exPlanations (SHAP) analysis highlighted Frw as the dominant factor affecting residence volume fraction (Rv), while ns, SR, and Va exhibited significant condition-dependent effects.