基于人工神经网络的凸轮轴铸造过程数值模拟优化研究

在实测铸铁凸轮轴铸造温度场的基础上，研究了人工神经网络技术在铸造数值模拟优化中的应用。首先采用三维有限元方法模拟了凸轮轴充型凝固过程的温度分布。在温度场实测方案中，设计了7个热电偶测温点。通过实测数据与模拟数据的比较，确定有限元模拟的最大相对误差为4．54％，CPU时间为3200s。人工神经网络采用了基于自适应学习率一动量项的误差反向传播梯度下降算法，并以温度场实测数据及有限元模拟数据为样本，进行了充型凝固数值模拟的优化。神经网络优化处理后模拟的最大相对误差为1．98％，CPU时间为670s，从而在模拟精度和效率上均优于传统有限元法。在铸造过程模拟中引入神经网络优化具有良好的可行性。

Optimization of Numerical Simulation for Solidification Process of Camshaft Based on Artificial Neural Network

Based on tested data of solidifying temperature field of cast-iron camshaft, optimization of numerical simulation in casting process was investigated based on artificial neural network. First, three-dimensional finite element modeling was used to simulate the temperature distribution of camshaft during filling and solidification. Seven sites on the camshaft were selected for temperature testing by thermo-couples. By contrasting the data of the simulation with those of the testing, it was estimated that the maximum relative error of finite element modeling was 4.54%,and the CPU time was 3200 s. In neuro-optimization of solidification process modeling, a gradient-descendent algorithm of error back-propagation with adaptive learning rate and momentum was applied to train the data of specimens obtained by finite element modeling and the testing results of temperature field. After the optimization, the maximum relative errors of simulation is 1.98%,and its CPU time is 670 s. It can be concluded that the modeling precision and efficiency by neural network are higher than those by finite element modeling, and good feasibility of the application of neuro-optimization to solidification process modeling is demonstrated.