工程科学学报，第40卷，第11期：1373-1379,2018年11月 Chinese Journal of Engineering,Vol.40,No.11:1373-1379,November 2018 DOI:10.13374/j.issn2095-9389.2018.11.011;http://journals.ustb.edu.cn 基于CART决策树的冲压成形仿真数据挖掘 王 伟12)四，黄宇星)，余鸿敏) 1)福州大学机械工程及自动化学院，福州3501002)福建省高端装备制造协同创新中心，福州350100 ☒通信作者，E-mail:mkwang@fz.edu.cn 摘要油箱壳外形复杂，拉深成形过程中容易出现侧壁起皱和圆角处破裂的缺陷，成形工艺参数的确定非常重要.结合分 类与回归决策树(classification and regression tree,CART)的人工智能技术和模型交叉验证方法，通过调用Python平台开源库 Scikit-.Learn对油箱壳拉深成形数值模拟结果进行知识挖掘，筛选出对油箱壳拉深成形影响大的工艺参数：以基尼指数(Gini idex)最小化作为最优特征值及最优切分点选择的依据，构建了工艺参数与性能指标关系的CART决策树，提取出了可靠的 工艺设计规则.油箱壳拉深实例表明，CART决策树理论的知识发现技术是实现板料成形过程数值模拟结果潜在知识挖掘的 可行途径 关键词知识发现：数据挖掘：数值仿真：分类与回归树：主成分分析 分类号TG302 Data mining of deep drawing simulation results based on CART decision tree theory WANG Wei,HUANG Yu-xing",YU Hong-min 1)School of Mechanical Engineering Automation,Fuzhou University,Fuzhou 350100.China 2)Fujian Provincial Collaborative Innovation Center of High-End Equipment Manufacturing,Fuzhou 350001,China Corresponding author,E-mail:mkwang@fzu.edu.cn ABSTRACT Numerical simulation technology is widely used in material forming process optimization and mold design.Although large volumes of simulation result data can be obtained,it is difficult to directly derive the relationship between the forming quality and the forming process parameters.To extract the potential knowledge latent in the simulation results,a systematic,robust,and efficient knowledge discovery technology is necessary,such as artificial intelligence technology,which has become one of the important research directions of material forming and processing.In this study the deep drawing process of a motorcycle fuel tank cover was taken as an example.A motorcycle fuel tank has complicated surfaces and local small fillets,and during its formation,the side wall and fillet are likely to wrinkle and rupture,respectively,because of local deep and uneven deformation.It is important to determine the forming pa- rameters to produce high quality tank cover that satisfies the surface quality requirements.Compared with the iterative dichotomiser 3 (ID3)decision tree algorithm,the classification and regression decision tree (CART)algorithm is advantageous in terms of faster com- putation speed,higher stability,and supporting multiple segmentation of continuous data.Furthermore,compared with other algorithms such as support vector machines (SVM)and logistic regression (LR),using the CART decision tree algorithm,the decision tree dia- gram can be established,and knowledge rules can be visually extracted.Combining the artificial intelligence technology of CART deci- sion tree and the model cross validation method of FI score,Scikit-Learn,an open-source library of Python platform was used to carry out knowledge discovery from the numerical simulation results of the tank cover deep drawing process.The key forming process parame- ters of the tank cover,which are blank holder force,the height of the draw bead,and radius of the die fillet,were identified.The opti- mal eigenvalues and the optimal segmentation points of CART decision tree were selected according to the minimization criteria of Gini index,and the process rules were extracted from the CART decision tree of the forming quality index and the established process 收稿日期：2017-10-17 基金项目：国家自然科学基金资助项目(51375094)：福建省科技计划资助项目(2018H0015)工程科学学报,第 40 卷,第 11 期:1373鄄鄄1379,2018 年 11 月 Chinese Journal of Engineering, Vol. 40, No. 11: 1373鄄鄄1379, November 2018 DOI: 10. 13374 / j. issn2095鄄鄄9389. 2018. 11. 011; http: / / journals. ustb. edu. cn 基于 CART 决策树的冲压成形仿真数据挖掘 王 伟1,2)苣 , 黄宇星1) , 余鸿敏1) 1) 福州大学机械工程及自动化学院, 福州 350100 2) 福建省高端装备制造协同创新中心, 福州 350100 苣 通信作者, E鄄mail: mkwang@ fzu. edu. cn 摘 要 油箱壳外形复杂,拉深成形过程中容易出现侧壁起皱和圆角处破裂的缺陷,成形工艺参数的确定非常重要. 结合分 类与回归决策树( classification and regression tree,CART)的人工智能技术和模型交叉验证方法,通过调用 Python 平台开源库 Scikit鄄Learn 对油箱壳拉深成形数值模拟结果进行知识挖掘,筛选出对油箱壳拉深成形影响大的工艺参数;以基尼指数(Gini index)最小化作为最优特征值及最优切分点选择的依据,构建了工艺参数与性能指标关系的 CART 决策树,提取出了可靠的 工艺设计规则. 油箱壳拉深实例表明,CART 决策树理论的知识发现技术是实现板料成形过程数值模拟结果潜在知识挖掘的 可行途径. 关键词 知识发现; 数据挖掘; 数值仿真; 分类与回归树; 主成分分析 分类号 TG302 收稿日期: 2017鄄鄄10鄄鄄17 基金项目: 国家自然科学基金资助项目(51375094); 福建省科技计划资助项目(2018H0015) Data mining of deep drawing simulation results based on CART decision tree theory WANG Wei 1,2)苣 , HUANG Yu鄄xing 1) , YU Hong鄄min 1) 1) School of Mechanical Engineering & Automation, Fuzhou University, Fuzhou 350100, China 2) Fujian Provincial Collaborative Innovation Center of High鄄End Equipment Manufacturing, Fuzhou 350001, China 苣 Corresponding author, E鄄mail: mkwang@ fzu. edu. cn ABSTRACT Numerical simulation technology is widely used in material forming process optimization and mold design. Although large volumes of simulation result data can be obtained, it is difficult to directly derive the relationship between the forming quality and the forming process parameters. To extract the potential knowledge latent in the simulation results, a systematic, robust, and efficient knowledge discovery technology is necessary, such as artificial intelligence technology, which has become one of the important research directions of material forming and processing. In this study the deep drawing process of a motorcycle fuel tank cover was taken as an example. A motorcycle fuel tank has complicated surfaces and local small fillets, and during its formation, the side wall and fillet are likely to wrinkle and rupture, respectively, because of local deep and uneven deformation. It is important to determine the forming pa鄄 rameters to produce high quality tank cover that satisfies the surface quality requirements. Compared with the iterative dichotomiser 3 ( ID3) decision tree algorithm, the classification and regression decision tree (CART) algorithm is advantageous in terms of faster com鄄 putation speed, higher stability, and supporting multiple segmentation of continuous data. Furthermore, compared with other algorithms such as support vector machines (SVM) and logistic regression (LR), using the CART decision tree algorithm, the decision tree dia鄄 gram can be established, and knowledge rules can be visually extracted. Combining the artificial intelligence technology of CART deci鄄 sion tree and the model cross validation method of F1 score, Scikit鄄Learn, an open鄄source library of Python platform was used to carry out knowledge discovery from the numerical simulation results of the tank cover deep drawing process. The key forming process parame鄄 ters of the tank cover, which are blank holder force, the height of the draw bead, and radius of the die fillet, were identified. The opti鄄 mal eigenvalues and the optimal segmentation points of CART decision tree were selected according to the minimization criteria of Gini index, and the process rules were extracted from the CART decision tree of the forming quality index and the established process