基于GA-BP神经网络晶粒尺寸预测模型的轮端轮毂锻造工艺优化  

Optimization of Wheel End Hub Forging Process Based on GA-BP Neural Network Grain Size Prediction Model

作　　者：孔德瑜 晏洋 张浩[1] 邓磊[1] 王新云[1] 龚攀[1] 张茂[1] KONG Deyu;YAN Yang;ZHANG Hao;DENG Lei;WANG Xinyun;GONG Pan;ZHANG Mao(State Key Laboratory of Material Processing and Die and Mould Technology,Huazhong University of Science and Technology,Wuhan 430074,China;Hubei Triring Forging Co.,Ltd.,Hubei Xiangyang 441700,China)

机构地区：[1]华中科技大学材料成形与模具技术全国重点实验室,武汉430074 [2]湖北三环锻造有限公司,湖北襄阳441700

出　　处：《精密成形工程》2024年第3期44-51,共8页Journal of Netshape Forming Engineering

基　　金：国家重点研发计划(2022YFB3706903);国家自然科学基金(52090043)

摘　　要：目的针对6082铝合金轮端轮毂在热处理过程中出现的粗晶问题,利用基于遗传算法优化的BP神经网络晶粒尺寸预测模型模拟优化锻造工艺方案,避免产生粗晶。方法以遗传算法替代梯度下降法优化神经网络各节点的权值和阈值,建立高精度的GA-BP神经网络晶粒尺寸预测模型,再以轮端轮毂为对象,设计锻造工艺方案并利用Deform进行微观组织仿真,研究压下速率、坯料初始温度对晶粒尺寸的影响,获得最优方案。结果优化模型预测的晶粒尺寸平均值和最大值的平均绝对百分比误差分别为2.55%、0.43%,与常规的BP神经网络相比,准确性有了较大提高。对比不同锻造方案的结果,得到轮毂较优的初始坯料温度为500℃,压下速率为200mm/s,经试验验证,锻件特征位置的晶粒尺寸预测值与实际值之间的误差均在10%以下,表明该预测模型具有良好的工程应用价值。结论遗传算法的引入大大增强了BP神经网络的全局寻优能力,提高了模型的准确性。在Deform中复现的预测模型对锻件的晶粒尺寸分布有较好的预测效果,并基于此成功模拟、优化了轮端轮毂的锻造方案。Aiming at the coarse grains of the 6082 aluminum alloy wheel end hub occurring during heat treatment,the work aims to simulate and optimize the forging process with the BP neural network grain size prediction model based on genetic algo-rithm optimization so as to avoid coarse grains.The weights and thresholds of each node in the neural network were optimized by using the genetic algorithm instead of the gradient descent method.A GA-BP neural network grain size prediction model with high precision was established.Subsequently,taking the wheel end hub as the object,different forging process schemes were designed and microstructure simulation was conducted using Deform to investigate the impact of compression rate and initial billet temperature on grain size,and obtain the optimal scheme.The mean absolute percentage error of the average and maxi-mum grain size predicted by the optimized model were 2.55%and 0.43%,respectively,which was a significant improvement in accuracy compared with the conventional BP neural network.The optimal initial billet temperature of the wheel end hub was determined to be 500,with a compression rate of 200 mm/s,based on comparative analysis of different forging schemes.The℃experimental results demonstrated that the error between the predicted and the actual grain size of the characteristic position was less than 10%,which indicated that the prediction model had good engineering application value.The introduction of genetic algorithm greatly enhances the global optimization ability of the BP neural network and improves the accuracy of the model.The prediction model reproduced in Deform has a good prediction effect on the grain size distribution of forgings,and based on this,the forging scheme of wheel end hub is successfully simulated and optimized.

关 键 词：轮端轮毂 晶粒尺寸预测 遗传算法 神经网络 数值模拟 

分 类 号：TG316.3[金属学及工艺—金属压力加工]