Interpretable physical-informed machine learning for dynamic characteristics prediction of air spring with limited measurement information.

Pipe-type air spring with auxiliary airbag (ASAA) are important vibration isolation component due to its tunable volume and adjustable dynamic characteristics. However, its dynamic modeling is challenging because airbag expansion work, rubber bladder viscoelasticity, and turbulence-dominated throttle dynamics are strongly coupled under large-amplitude excitations. Traditional data-driven models often require large datasets and lack interpretability. To address this issue, this study proposes an interpretable physics-informed machine learning (PIML) framework for ASAA dynamic prediction. A reduced-order physical model is first developed to provide baseline predictions of stiffness and lag angle, and the discrepancies between physical predictions and experimental measurements are then used to construct residual datasets. An interpretable Kolmogorov-Arnold Network (KAN) is employed to learn these residuals through adaptive B-spline activation functions and derive explicit symbolic correction terms. Experimental validation shows that the proposed hybrid model improves prediction accuracy while retaining physical interpretability, outperforming pure data-driven, advanced physics-data and physical models.