摘要
目的通过机器学习方法构建工艺参数与壁厚减薄率之间的映射关系,建立弯管壁厚减薄率高精度预测模型,为缺陷的有效预测提供数据驱动解决方案。方法首先,以直径、弯曲程度等6项工艺参数为变量设计仿真方案,并进行实验验证,建立数据库来源。其次,分别建立随机森林(RandomForest,RF)与梯度提升回归树(GradientBoostingRegressionTree,GBRT)2种机器学习模型。最后,对比分析训练样本量(N=100~808)对2种机器模型泛化性能的影响规律。结果在中等样本量(N=400)下,2种机器模型的预测性能达到最优。其中,GBRT模型测试集决定系数(R^(2))高达0.9769,方差贡献率(VAF)为97.69%,均方根误差(RMSE)和平均绝对误差(MAE)分别低至0.3470和0.1759,所有性能指标均显著优于同等条件下的RF模型的性能指标(R^(2)=0.9229,VAF=92.36,RMSE=0.6335,MAE=0.3638)。当样本量超过400后,2种机器模型均出现预测性能饱和甚至衰减的趋势。采用训练区间之外的工艺参数进行GBRT模型预测精度验证,发现其平均预测误差小于7%。结论构建的GBRT预测模型能够高精度、高效率地映射双金属复合管弯曲工艺参数与壁厚减薄率之间的复杂非线性关系,验证了机器学习方法在该领域替代或辅助传统有限元分析的可行性。
The work aims to construct a mapping relationship between process parameters and the wall thickness reduction rate with machine learning methods and establish a high-precision prediction model for the wall thickness reduction rate in tube bending,thereby providing a data-driven solution for effective defect prediction.Firstly,a simulation scheme was designed with six process parameters,such as diameter and bending degree,etc.as variables,and experimental validation was conducted to establish a database.Secondly,two machine learning models—Random Forest(RF)and Gradient Boosting Regression Tree(GBRT)—were developed.Finally,the effect of training sample size(ranging from N=100 to 808)on the generalization performance of the two models was compared and analyzed.Under a medium sample size(N=400),both machine learning models achieved optimal prediction performance.Among them,the GBRT model exhibited a coefficient of determination(R^(2))as high as 0.9769 on the test set,a variance accounted for(VAF)of 97.69%,and a root mean square error(RMSE)and a mean absolute error(MAE)as low as 0.3470 and 0.1759,respectively.All performance metrics were significantly superior to those of the RF model under the same conditions(R^(2)=0.9229,VAF=92.36,RMSE=0.6335,MAE=0.3638).Beyond a sample size of 400,both models showed a trend of performance saturation and eventual degradation.Using process parameters outside the training range for GBRT model validation revealed an average prediction error of less than 7%.The GBRT prediction model constructed in this study can accurately and efficiently map the complex nonlinear relationship between the bending process parameters of bimetallic composite tubes and the wall thickness reduction rate,demonstrating the feasibility of machine learning methods as alternatives or supplements to traditional finite element analysis in this field.
作者
朱英霞
元琛
王磊
陈炜
许江平
ZHU Yingxia;YUAN Chen;WANG Lei;CHEN Wei;XU Jiangping(School of Mechanical Engineering,Jiangsu University,Jiangsu Zhenjiang 212031,China)
出处
《精密成形工程》
北大核心
2026年第3期184-193,共10页
Journal of Netshape Forming Engineering
基金
国家自然科学基金(52575387)
校企合作(HX20241044)。
关键词
双金属复合管
弯曲
壁厚减薄
机器学习
预测模型
bimetallic composite tube
bending
wall thickness reduction
machine learning
prediction model
