基于物理信息神经网络的柴油机机体疲劳安全系数预测模型研究

蒲博闻, 廖挥亚, 孙兴悦, 王根全, 刁占英, 陈旭

doi:10.3969/j.issn.1001-2222.2025.05.008

车用发动机 >

2025 , Vol. 0 >Issue 5: 52

DOI: https://doi.org/10.3969/j.issn.1001-2222.2025.05.008

基于物理信息神经网络的柴油机机体疲劳安全系数预测模型研究

展开

（1.天津大学化工学院，天津 300350；2.中国北方发动机研究所（天津），天津 300405；3.西北工业大学航空学院，陕西西安 710072）

网络出版日期: 2025-10-31

收起

Fatigue Safety Factor Prediction Model of Diesel Engine Body Based on Physics-Based Neural Network

Expand

(1.School of Chemical Engineering and Technology,Tianjin University,Tianjin 300350,China;2.China North Engine Research Institute (Tianjin),Tianjin 300405,China;3.School of Aeronautics,Northwestern Polytechnical University,Xi’an 710072,China)

Online published: 2025-10-31

Fold

摘要

针对柴油机机体结构疲劳安全系数的高效预测问题，选取机体主轴承隔板典型关注部位，结合现代采样方法与随机有限元仿真计算，提出了一种基于梯度约束的物理信息神经网络（GC-PINN）作为疲劳安全系数预测的代理模型。该模型在机器学习模型中融入螺栓预紧力与安全系数间的物理梯度关系，以引导并约束代理模型的训练过程，从而提升模型的泛化能力和预测精度。将GC-PINN代理模型与基于人工神经网络、随机森林和支持向量回归的代理模型进行预测精度对比，结果表明，该代理模型在小样本条件下表现出较好的预测效果，其均方根误差值（RMSE）中位数最低，仅为0.294。

关键词： 机体; 物理信息神经网络; 高周疲劳; 安全系数; 预测; 代理模型

本文引用格式

蒲博闻, 廖挥亚, 孙兴悦, 王根全, 刁占英, 陈旭 . 基于物理信息神经网络的柴油机机体疲劳安全系数预测模型研究[J]. 车用发动机, 2025 , 0(5) : 52 . DOI: 10.3969/j.issn.1001-2222.2025.05.008

Abstract

To address the efficient prediction of fatigue safety factors (SF) for the block structure of diesel engine, several typical positions of block main bearing were selected and a gradient-constrained physical information neural network (GC-PINN) was proposed as a surrogate model for SF prediction by combining modern sampling methods with stochastic finite element simulation calculations. Through incorporating the physical gradient relationship between bolt preload and SF into machine learning models, the training process of surrogate model was guided and constrained to improve its generalization ability and prediction accuracy. Prediction accuracy comparisons between the GC-PINN surrogate model and models based on artificial neural networks, random forests, and support vector regression showed that the surrogate model had better prediction performance under small sample conditions, exhibiting the lowest median mean square root error value (RMSE) of only 0.294.

Key words： engine body; physics-based neural network; high-cycle fatigue; safety factor; prediction; surrogate model

Options

摘要页面

文章导航

模态框（Modal）标题

摘要

本文引用格式

Abstract