Wear Prediction of Grooved Journal-Bearing Shells Using Optimized BP Neural Networks

To predict the mass loss of grooved journal-bearing shells under varying operating conditions while reducing experimental effort, accelerated wear tests were conducted on a component-level journal-bearing test rig. The effects of load, rotational speed, and oil temperature were investigated, and the worn surfaces were characterized microscopically to elucidate the wear mechanism. A backpropagation (BP) neural network was trained on 42 mass-loss samples and evaluated using 18 additional samples. Candidate numbers of hidden-layer neurons were identified using an empirical relation and compared based on prediction performance. A genetic algorithm (GA) and the sparrow search algorithm (SSA) were then used to optimize the initial weights and biases of the BP network. Mass loss increased with load and oil temperature but decreased with rotational speed; all responses were nonlinear. After testing, the groove structure remained visible, while the lead-based overlay became thinner and underwent plastic flow. Pb and Sn were detected on the aluminum-alloy surface, suggesting transfer of overlay material during sliding. The mean absolute percentage error (MAPE) and coefficient of determination (R2) were 26.986% and 0.54552, respectively, for the conventional BP model; 8.436% and 0.9862 for the GA-BP model; and 2.6296% and 0.99714 for the SSA-BP model. For the present dataset, the SSA-BP model achieved lower prediction errors than the BP and GA-BP models and reproduced the variation in mass loss across the investigated operating conditions.