Gaussian Noise-Augmented Machine Learning for Reliable Friction Estimation in Bolted Joints

Summary & ConclusionsAccurate prediction of the functional behavior of bolted joints is essential for ensuring reliability in mechanical systems. Traditional analytical, numerical and empirical methods often fall short in capturing nonlinear and multivariate interactions among key parameters such as preload force, torque, and friction coefficients, especially under variable operating conditions. Adding to this existing models frequently rely on limited data, which fail to represent the diverse conditions. To address these limitations, this study proposes a machine learning-based reliability framework that leverages synthetic data generation and advanced modeling techniques to enable scalable, low-cost prediction of critical performance parameters. A rule-based data generation pipeline was developed using deterministic equations from established standards and augmented with 2D Gaussian noise modeled from experimental deviations to simulate real-world variability. The synthetic datasets enable efficient training of predictive models without requiring large volumes of experimental data. Three modeling approaches were evaluated: gradient boosting (XGBoost), multilayer perceptrons (MLPs), and one-dimensional convolutional neural networks (1D-CNNs). These models were trained exclusively on synthetic data and validated on unseen experimental datasets representing M6 and M10 bolted joints. Results show that the 1D-CNN outperforms other methods in capturing hierarchical and nonlinear interactions, leading to robust generalization across joint configurations. This approach reduces reliance on resource-intensive testing, enhances reliability prediction, and supports data-driven decision-making for design and maintenance processes.