基于HJPS算法的神经网络的轧制力修正模型

Rolling Force Modification Modeling of Neural Networks Based on HJPS Algorithm

传统的轧制力模型结构简单、精度较低,即使采用基于有限元的数值积分方式进行精化,出于计算效率的考虑因其有限区域的划分十分有限,因此对于轧制力计算的精度提高有限。直接采用神经网络对轧制力进行建模可以极大地提高模型精度,但是模型对新型材料的泛化能力较差。为此提出简单有限元轧制力模型,并在模型基础上使用HJPS优化算法的神经网络对轧制力进行修正,对该模型的仿真测试表明,该模型具有很强的泛化能力,收敛速度快、不易陷于局部优化,能够极大地提高轧制力模型的计算精度。

The conventional rolling model is too simple and its precision is too low. The rolling model buih by numerical integration based on limited units can improve computing precision, but the improved precision is limited because the number of partitioned computing area is linited. The precision of the rolling model is satisfactory, but its extension ability is not. The model based simple limited unit algorithm and modified by ANN-based HJPS optimization algorithm have the following merits： definite extension ability, high convergent velocity, no ease to get in local optimization and higher precision.