Stochastic Dynamic Response Analysis of Spherical Roller Thrust Bearings Based on Improved Deep Neural Network

The roller–raceway contact response is a key factor affecting stress concentration, fatigue initiation, and raceway spalling in spherical roller thrust bearings. Uncertainty analysis of this response is therefore important for revealing how practical parameter fluctuations affect bearing contact behavior and for supporting robust bearing design and operating-condition optimization. In this paper, a multibody dynamic model of a spherical roller thrust bearing is established by explicitly considering the main internal contact pairs, including roller–raceway, roller–flange, roller–cage, and cage–guide interactions. The model is used to obtain the transient roller–raceway contact loads under coupled axial loading and rotational motion. The resulting contact loads are introduced into a finite element contact model to evaluate the dynamic contact stress response of the inner raceway. To assess the effects of random uncertainties on this response, an improved deep neural network (DNN) surrogate model is developed. An attention mechanism deep neural network (AM-DNN) is improved by incorporating feature importance information from random forest (RF) into its attention mechanism, and the resulting model is denoted by RF-AM-DNN. Validation on the generated dataset demonstrates that the proposed RF-AM-DNN outperforms conventional surrogate models in prediction accuracy. Finally, the RF-AM-DNN is used to investigate the uncertainty characteristics of dynamic contact stress in spherical roller thrust bearings under multiple uncertainty factors.