【作者】 张强；

【导师】 王文礼； 刘明华；

【作者基本信息】 西安建筑科技大学 ， 材料工程， 2021， 硕士

【摘要】 轧制力数学模型是轧制区域最为重要的数学模型之一,其预测精度直接影响产品的厚度精度和板形质量。传统数学模型的建模过程假设和简化了轧制过程的诸多实际因素,模型计算误差较大,往往不能满足现代化高精度轧制技术的要求。建立高精度轧制力数学模型已经成为近些年研究的热点之一。支持向量回归(SVR)是以统计学理论为基础的一种机器学习方法,不仅可以避免传统数学模型假设脱离实际和简化过于粗糙产生的误差,同时在实验数据有限的条件下,也具有较强的泛化能力。目前,SVR轧制力模型存在一些不足,采用单一核函数难以解决高维、非线性的轧制问题,以及输入变量维数过高的问题。本文利用西姆斯(Sims)热轧公式确定了影响轧制力的主要参数,根据主要参数从国内两个大型钢厂分别采集板材轧制实验数据,并采用T检验准则和归一化等方法预处理实验数据。将A厂预处理数据作为训练集训练模型,B厂预处理数据作为测试集验证模型的准确性,以及采用统计指标评价模型的综合性能。论文的主要研究内容和结果如下:采用混合核函数(HKF)代替单一核函数,提出了粒子群算法(PSO)与天牛须算法(BAS)相结合的混合优化算法(PSO-BAS),并采用PSO-BAS算法优化HKSVR轧制力模型的参数(惩罚因子(c),径向基核函数(g),不敏感损失参数(?),多项式核函数(d),控制参数(m)),建立了PSO-BAS-HKSVR轧制力预测模型。采用测试集验证PSO-BAS-HKSVR的预测精度,并将该模型的计算结果与PSO-HKSVR、PSO-SVR、网格搜索算法(Grid)-SVR、BP神经网络(BPNN)、广义回归神经网络(GRNN)、径向基神经网络(RBF)和传统轧制力数学模型比较,PSO-BAS-HKSVR模型具有更好的泛化能力且平均绝对百分比误差(MAPE)值仅为4.9600%。另一方面,采用主成分分析(PCA)技术降低输入变量的维数,提出了自适应调节惯性权值和加速因子的方法改进PSO算法(IPSO),并采用IPSO算法优化SVR轧制力模型的参数(c,g),建立了PCA-IPSO-SVR轧制力预测模型。采用测试集验证PCA-IPSO-SVR的预测精度,并将该模型的计算结果与PCA-PSO-SVR、PSO-SVR、Grid-SVR、BPNN、GRNN、RBF和传统轧制力数学模型比较,PCA-IPSO-SVR模型具有最高的预测精度且MAPE值为5.1716%。PSO-BAS-HKSVR与PCA-IPSO-SVR模型的预测精度相比提高了4.09%,而对于运算速度降低了7.982s,相对于模型的运算速度,提高模型的预测精度具有更重要的意义。PSO-BAS-HKSVR模型具有较强的泛化性能,为实现高精度轧制力预测提供了新的方法。更多还原

【Abstract】 The rolling force mathematical model is one of the most important mathematical models in the rolling area,and its prediction accuracy directly affects the thickness accuracy and shape quality of the products.The modeling process of the traditional mathematical model assumes and simplifies many practical factors in the rolling process.The calculation error of the model is large,and it often can not meet the requirements of modern high-precision rolling technology.Establishing a high-precision mathematical model of rolling force has become one of the research hot spots in recent years.Support vector regression(SVR)is a machine learning method based on statistical learn theory.It can not only avoid the errors caused by traditional mathematical model assumptions that are out of reality and simplified too rough,but also has strong generalization ability in the case of limited experimental data.At present,the SVR rolling force model has some shortcomings,such as single kernel function in SVR rolling force model may be difficult to solve the high-dimensional and nonlinear strip rolling problem,and the dimension of input variables is too high.In this paper,Sims hot rolling formula is used to determine the main parameters of affecting the rolling force.According to the main parameters,the experimental data of plate rolling are collected from two large domestic steel plants,and the experimental data are preprocessed by T test criterion and normalization method.The preprocessed data of A plant is used as the training set to train the model and the preprocessed data of B plant is used as the test set to verify the accuracy of the model,and the comprehensive performance of the model is evaluated by using statistical indexes.The main research contents and results of this paper are as follows:A hybrid kernel function(HKF)is used instead of a single kernel function,and a hybrid optimization algorithm(PSO-BAS)combining the particle swarm algorithm algorithm(PSO)and the beetle antennae search algorithm(BAS)is proposed.The PSO-BAS algorithm optimizes the parameters of the HKSVR rolling force model(penalty factor(c),radial basis kernel function(g),insensitive loss parameter(?),polynomial kernel function(d)and control parameters(m)),and the PSO-BAS-HKSVR rolling force prediction model is established.The prediction accuracy of the PSO-BAS-HKSVR is verified by the test sets,and the calculation results of the model are compared with PSO-HKSVR,PSO-SVR,grid search algorithm(Grid)-SVR,BP neural network(BPNN),general regression neural network(GRNN),radial basis function network(RBF)and traditional rolling force mathematical models.The PSO-BAS-HKSVR model has better generalization ability,and the average absolute percentage error(MAPE)value of the model is only 4.9600%.On the other hand,the principal component analysis(PCA)technology is used to reduce the dimensionality of the input variables,and the method of adaptively adjusting the inertia weight and acceleration factors is proposed to improve the PSO algorithm(IPSO),and the IPSO algorithm is used to optimize the parameter of SVR rolling force model(c,g),and the PCA-IPSO-SVR rolling force prediction model is established.The test set is used to verify the prediction accuracy of PCA-IPSO-SVR,and the calculation results of the model are compared with PCA-PSO-SVR,PSO-SVR,Grid-SVR,BPNN,GRNN,RBF and traditional rolling force mathematical models.The PCA-IPSO-SVR model has the highest prediction accuracy,and the MAPE value of the model is 5.1716%.Compared with the PCA-IPSO-SVR model,the prediction accuracy of PSO-BAS-HKSVR is increased by 4.09%,but the calculation speed of PSO-BAS-HKSVR is reduced by 7.982 s.Compared with the calculation speed of the model,it is of more important significance to improve the prediction accuracy of the model.The PSO-BAS-HKSVR model has strong generalization performance,and a new method is proposed to achieve high-precision rolling force prediction.更多还原