机械筒形件离心铸造工艺的神经网络优化  

Neural Network Optimization of Centrifugal Casting Process for Mechanical Cylinders

作　　者：吴兆立[1] 仇多利 陈辉[3] WU Zhaoli;QIU Duoli;CHEN Hui(Jiangsu Vocational Institute of Architectural Technology,Xuzhou 221000,China;School of Management,Huaibei Normal University,Huaibei 235000,China;School of Materials and Physics,China University of Mining and Technology,Xuzhou 221116,China)

机构地区：[1]江苏建筑职业技术学院,江苏徐州221000 [2]淮北师范大学管理学院,安徽淮北235000 [3]中国矿业大学材料与物理学院,江苏徐州221116

出　　处：《热加工工艺》2021年第21期66-69,73,共5页Hot Working Technology

基　　金：江苏建筑节能与建造技术协同创新中心课题(SJXTY1603)。

摘　　要：以浇注温度、旋转速度、离心旋转半径和离心力保持时间为输入层参数,以铸造缺陷等级和抗拉强度为输出层参数,采用4×24×8×2四层拓扑结构构建了机械筒形件离心铸造工艺神经网络优化模型。结果表明:模型经过8886次迭代运算后收敛,训练性能曲线平滑,模型输出的铸造缺陷等级相对预测误差为3.3%~6.7%,平均相对预测误差4.8%;模型输出的抗拉强度相对预测误差为3.0%~5.1%,平均相对预测误差4.3%,具有较强的预测能力、较高的预测精度和较好的实用性。与企业现用工艺相比,采用优化工艺离心铸造的筒形件铸造缺陷从2级变为1级,抗拉强度提高18MPa。Taking pouring temperature,rotation speed,centrifugal rotation radius and centrifugal force holding time as input layer parameters,casting defects(characterized by defect grade)and tensile strength as output layer parameters,the neural network optimization model for centrifugal casting process ofmechanical cylinder parts was established by using 4×24×8×2 four layer topological structure.The results show that the model converges after 8886 iterations,the training performance curve is smooth,the relative prediction error of casting defects is between 3.3%and 6.7%,and the average relative prediction error of casting defects is 4.8%;the relative prediction error of tensile strength is between 3.0%and 5.1%,and the average relative prediction error of tensile strength is 4.3%.The model has strong prediction ability,high prediction accuracy and good usability.Compared with those of the current process of the enterprise,the casting defects of cylindrical parts cast by centrifugal casting with optimized process changes from grade 2 to grade 1,and the tensile strength increases by 18 MPa.

关 键 词：神经网络优化 离心铸造 机械筒形件 输入层参数 输出层参数 

分 类 号：TG292[金属学及工艺—铸造] TG249.4