Machine learning-based discrimination of lubrication regimes in journal bearings using acoustic emission and knock sensors

Journal bearings are critical components in a wide range of rotating machinery, from internal combustion engines and industrial turbines to wind turbine drivetrains, where accurate monitoring of the lubrication regime is essential for predictive maintenance and efficiency optimisation. This study investigates the classification of lubrication regimes (boundary, mixed, and hydrodynamic) in a journal bearing test system using acoustic emission (AE) and knock sensors, processed by an artificial neural network (ANN) supported by a genetic algorithm (GA) for feature selection. Sensor signals were segmented and characterised using power spectral, hit rate, and wavelet-based features, with lubrication regime labels derived from the lambda ratio. The AE-sensor achieved a classification accuracy of 95.02%, with predicted regime transitions closely matching those of the lambda ratio, while the knock sensor performed substantially worse at 67.18%, attributed to its lower operating frequency range. The trained ANN generalised successfully to previously unseen operating conditions, including a different oil temperature and sensor configuration, without retraining. GA-based feature selection provided a measurable improvement in classification accuracy over random selection at reduced feature set sizes, supporting its use in computationally constrained applications. These findings demonstrate that AE sensing combined with machine learning offers a robust and adaptable method for real-time lubrication regime monitoring in journal bearings.