A Rapid Verification Method for Ball Bearing Interference Fit Based on Automated Dataset Construction and Optimized Neural Network with Uncertainty Quantification

Conventional ball bearing interference fit verification relies heavily on labor-intensive finite element (FE) simulations, while existing data-driven approaches suffer from non-standard dataset preparation, single-objective optimization, and a lack of uncertainty-aware risk assessment. This paper develops an integrated rapid verification framework to address these limitations. First, an automated dataset construction pipeline combining dimension chain decoupling and Latin Hypercube Sampling (LHS) is established, enabling fully automatic CAE modeling, solving and result extraction. Second, a multi-output fully connected neural network is optimized via a composite-objective Bayesian algorithm, which balances prediction accuracy, output stability and model complexity. Third, Monte Carlo Dropout is introduced to quantify epistemic uncertainty and generate probabilistic prediction intervals. Simulation results show that the dataset construction efficiency for 400 samples is more than 90 times higher than the manual approach, with a sample validity rate of 100%. The proposed model achieves an average test SMAPE of 12.48% under the single split used for comparison, and 16.12 ± 2.07% under 5-fold cross-validation. The value of this model lies in rapid evaluation of geometric variants: a single forward pass takes about 2 ms, and the full uncertainty prediction takes about 0.4–0.5 s. It outperforms classical surrogate models, including Kriging and response surface methodology on average, while a classical response surface is more accurate for the ZC_U1 output. The 95% prediction interval yields an empirical coverage rate of approximately 92.5–95.0% for epistemic uncertainty; corresponding 95% confidence intervals (Clopper-Pearson) for these coverage estimates are also reported, providing a quantitative reference for risk-informed design verification. This framework upgrades traditional deterministic point verification to interval verification with uncertainty characterization. It targets the axisymmetric interference fit between a deep groove ball bearing inner ring and a hollow journal between a shoulder and a snap ring groove, under linear elasticity and a constant friction coefficient. The training data cover parameter ranges representative of automotive drive-motor bearing fits; the specific ranges are not disclosed due to commercial confidentiality.