基于机器学习模具钢大型锻件力学性能预测  

Prediction of Mechanical Properties of Large Forgings in Mold Steel Based on Machine Learning

作　　者：王孟超 武川 孟亚飞 姚磊 WANG Mengchao;WU Chuan;MENG Yafei;YAO Lei(National-local Joint Engineering Laboratory of Intelligent Manufacturing Oriented Automobile Die&Mould;Tianjin University of Technology and Education,Tianjin 300222,ChinaTianjin Key Laboratory of High Performance Precision Forming Manufacturing Technology and Equipment,Tianjin University of Technology and Education,Tianjin 300222,China)

机构地区：[1]天津职业技术师范大学汽车模具智能制造技术国家地方联合工程实验室,天津300222 [2]天津职业技术师范大学天津市高性能精准成形制造技术与装备重点实验室,天津300222

出　　处：《精密成形工程》2024年第12期198-208,共11页Journal of Netshape Forming Engineering

基　　金：国家自然科学基金(52075386);天津市自然科学基金多投入重点项目(22JCZDJC00650);中国博士后科学基金第67项研究金(2020M672309);陕西省高性能精密成形技术与装备重点实验室项目(PETE2019KF02)。

摘　　要：目的为了预测模具钢大型锻件热处理后的力学性能,构建多种机器学习算法,探索出一种高效可行的算法来预测模具钢大型锻件的硬度、抗拉强度、屈服强度。方法以热作模具钢为研究对象进行热处理试验和力学性能检测得到模具钢大型锻件热处理后的力学性能数据集,先对其进行预处理,包括噪音消除和标准化等,以得到高质量的力学性能数据集。使用随机森林特征重要性分析法对输入参数进行筛选,去除弱相关变量降低预测算法的复杂度。随后构建随机森林算法、决策树算法、BP神经网络算法。综合对比各算法预测精度,选出最优的算法模型,对算法进行验证并应用。结果最终建立了硬度、抗拉强度、屈服强度预测算法,其中随机森林算法的确定系数R2分别为0.9539、0.9241、0.9320,决策树算法的确定系数R2分别为0.9485、0.9069、0.9280,BP神经网络算法的确定系数R2分别为0.8047,0.7921,0.7937。结论通过对比分析,随机森林算法表现出最高的预测精度,且随机森林算法对力学性能的预测误差基本保持在5%以内。说明通过该算法,可以实现模具钢大型锻件的力学性能预测,大大节省了研究和实验成本,能够极大地加快高性能材料的筛选。The work aims to predict the mechanical properties of large forgings of mold steel after heat treatment,construct multiple machine learning algorithms,and explore an efficient and feasible algorithm to predict the hardness,tensile strength,and yield strength of large forgings of mold steel.With hot work mold steel as the research object,heat treatment experiments and mechanical performance testing were conducted to obtain the mechanical performance dataset of large forgings of mold steel after heat treatment.First,it was preprocessed,including noise elimination and standardization,to obtain a high-quality mechanical performance dataset.Then,the random forest feature importance analysis method was used to screen the input parameters and remove weakly correlated variables to reduce the complexity of the prediction algorithm.Then a random forest algorithm,a decision tree algorithm,and a BP neural network algorithm were constructed.The prediction accuracy of various algorithms was compared comprehensively,the optimal algorithm model was selected,and the algorithm was verified and applied.As a result,a prediction algorithm for hardness,tensile strength,and yield strength was ultimately established.The determination coefficients R2 of the random forest algorithm were 0.9539,0.9241,and 0.9320,respectively.The determination coefficients R2 of the decision tree algorithm were 0.9485,0.9069,and 0.9280,respectively.The determination coefficients R2 of the BP neural network algorithm were 0.8047,0.7921,and 0.7937,respectively.In conclusion,through comparative analysis,the random forest algorithm shows the highest prediction accuracy,and the prediction error of the random forest algorithm for mechanical performance is basically maintained within 5%.Through this algorithm,it is possible to predict the mechanical properties of large forgings of mold steel,greatly saving research and experimental costs,and greatly accelerating the screening of high-performance materials.

关 键 词：机器学习 模具钢 预测模型 力学性能 随机森林 

分 类 号：TG142.1[一般工业技术—材料科学与工程]