PCA降维技术在弯辊力预设定中的研究与应用

为了提高冷连轧带钢弯辊力预设定模型的计算效率, 在原有基于GA-BP神经网络的弯辊力预设定模型基础上, 引入主成分分析(PCA)数据降维技术, 通过PCA将原有10个轧制参数变量转换为3个主成分变量, 降维后的主成分变量包含了原始实测轧制参数93.55%的信息, 实现了轧制参数特征的有效提取; 将其作为神经网络的输入, 建立PCA-GA-BP新形态弯辊力预设定模型, 简化了模型结构。以某1 450 mm冷连轧生产线数据作为样本比较了2种模型的计算性能, 结果表明, 2种模型均具有较好的泛化能力, 在保证带钢头部板形精度的基础上, PCA-GA-BP模型与原模型相比迭代次数减少86次, 计算时间缩短73 ms, 预报效率显著提高。

Research and Application of PCA Dimensionality Reduction Technology in Presetting Bending Force

In order to improve the calculation efficiency of the bending force preset model for cold-rolled strip, the principal component analysis (PCA) technique for data dimensionality reduction was introduced based on the bending force preset model with the original GA-BP neural network. The original 10 rolling parameter variables were converted into 3 principal component variables through PCA. The principal component variable after dimensionality reduction contained 93.55% of the information of originally measured rolling parameters, achieving the effective extraction of rolling parameter features. With it as the input of the neural network, the PCA-GA-BP new bending force preset model was established, leading to the model structure simplified. With the data from a 1 450 mm tandem cold rolling production line as a sample, the two models were compared in terms of the calculation performance. The results show that both two models are good in generalization ability. On the basis of ensuring the accuracy of the strip head flatness accuracy, the PCA-GA-BP model has iteration numbers 86 less than the original model and the calculation time shortened by 73 ms, but it has the forecasting efficiency significantly improved.