Tribological performance and wear prediction of spur gears lubricated with palm oil and lithium grease: A hybrid physics-informed machine learning approach

Gear wear prediction and sustainable lubrication are critical for Industry 4.0 predictive maintenance. In this study, spur gear wear under lubrication with palm oil and a lithium complex grease was evaluated using a Taguchi L32 design and ANOVA. Results showed palm oil outperformed grease, yielding lower mass losses (6–17 mg vs. 9–18 mg), with minimum loss (6–8 mg) at 192 rpm. Vibration analysis revealed lower Root Mean Square (RMS) values for palm oil at 480 rpm (0.0017–0.0081 mm) than grease (0.0033–0.0119 mm), while grease reached Peak-to-Peak amplitudes up to 0.034 mm. Taguchi S/N ratio analysis identified optimal conditions (palm oil, 192 rpm, 40 min) with palm oil achieving a higher mean S/N ratio (39.7 vs. 37.5). ANOVA indicated significant contributions from duration (20.96%, p = 0.037) and lubricant type (16.76%, p = 0.009). In contrast, 50.54% unexplained error highlighted complex interactions. To address data constraints, 32 physical samples were augmented using the Archard equation, yielding a total of 192 samples. A CatBoost model trained via 5-fold GroupKFold cross-validation achieved high accuracy in the LOW wear regime (mean R 2 logK = 0.7198, linear R 2 K = 0.6994, MSE logK = 0.0092, linear RMSE K = 6.9 × 10 −9 ). Conversely, the HIGH wear regime exhibited non-linear instability (mean R 2 logK = −2.7017, linear R 2 K = −11.7590). Feature importance confirmed regime-dependent behavior: revolutions (∼28%) and duration (∼22%) dominated in the LOW regime, whereas sliding distance (∼90%) dominated in the HIGH regime. Overall, this hybrid physics-informed ML approach robustly predicts gear wear under stable operating conditions while identifying critical wear-transition boundaries.