Scholarlink

<p>The friction coefficient and film thickness are crucial factors influencing the tribo-dynamic performance of gears. This study presents three artificial neural network (ANN) models, each designed to rapidly predict the friction coefficient and film thickness in elastohydrodynamic lubrication (EHL) contact of face-gear drives, accounting for both non-Newtonian and thermal effects. The architecture and hyperparameters of the ANN models are optimized through Bayesian optimization. A dataset derived from a non-Newtonian thermal EHL model is used for model training, and the impact of dataset size on model performance is analyzed, providing recommendations for the optimal dataset size. The importance of input features is assessed using Shapley additive explanations (SHAP), revealing that input features related to non-Newtonian and thermal effects play a pivotal role. A refined feature set is obtained by excluding less important input features, leading to the generation of improved ANN models. The improved models yield coefficient of determination values of 0.99744, 0.99444, and 0.99597 for predicting friction coefficient, minimum film thickness, and central film thickness, respectively. These values surpass those of the original models and empirical formulas, demonstrating the superior predictive performance of the improved models.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Application of artificial neural networks for predicting friction coefficient and film thickness considering non-Newtonian and thermal effects

  • 期刊论文

  • 发表日期：

    2025年5月23日

查看全文
  • Lei Wang,
  • Ning Zhao,
  • Xiaotao An,
  • Jinran Li

摘要

The friction coefficient and film thickness are crucial factors influencing the tribo-dynamic performance of gears. This study presents three artificial neural network (ANN) models, each designed to rapidly predict the friction coefficient and film thickness in elastohydrodynamic lubrication (EHL) contact of face-gear drives, accounting for both non-Newtonian and thermal effects. The architecture and hyperparameters of the ANN models are optimized through Bayesian optimization. A dataset derived from a non-Newtonian thermal EHL model is used for model training, and the impact of dataset size on model performance is analyzed, providing recommendations for the optimal dataset size. The importance of input features is assessed using Shapley additive explanations (SHAP), revealing that input features related to non-Newtonian and thermal effects play a pivotal role. A refined feature set is obtained by excluding less important input features, leading to the generation of improved ANN models. The improved models yield coefficient of determination values of 0.99744, 0.99444, and 0.99597 for predicting friction coefficient, minimum film thickness, and central film thickness, respectively. These values surpass those of the original models and empirical formulas, demonstrating the superior predictive performance of the improved models.