【作者】 张少宇；

【导师】 张秀玲；

【作者基本信息】 燕山大学 ， 控制理论与控制工程， 2012， 硕士

【摘要】 板带钢是主要的钢材产品之一，广泛应用于汽车制造、食品包装等行业，对国民经济发展有着举足轻重的影响。随着科技的进步，用户对板带钢质量的要求越来越高。板形是带钢的重要质量指标，也是轧制领域研究的热点。近年来，人工智能方法以其在建模和控制方面的优势，在工业过程研究中得到了广泛的应用。本文选择基于改进最小二乘支持向量回归机(LS-SVR)的冷带轧机板形智能控制为研究课题，分析了现有板形智能控制方法的优缺点，重点研究了LS-SVR算法的改进，并利用它对板形控制系统进行了深入的研究。首先，针对标准LS-SVR算法只适用于单输出系统回归估计，而工业过程多为多输出系统的问题，提出了一种多输出最小二乘支持向量回归机(MLSSVR)新算法，该算法秉承结构风险最小化原则，具有良好的泛化性能；针对MLSSVR超参数难以确定的问题，提出了一种基于粒子群算法的超参数优化方法，该方法不仅运算速度快，而且搜索能力强。板形模式识别是板形控制的关键。考虑到板形精确控制的要求，采用一次、二次、三次和四次勒让德正交多项式表示板形的基本模式；并采用MLSSVR方法建立识别模型。研究结果表明，该方法可以有效确定板形缺陷的类型和含量，识别精度高，泛化能力强。板形预测模型是板形控制系统设计的重要基础。为提高板形预测模型的准确性，以生产实测数据为基础，建立了基于MLSSVR的板形预测模型。仿真实验表明，MLSSVR板形预测模型具有预测精度高，鲁棒性强的优点。最后，综合分析了影响矩阵控制方法和预测控制方法的特点，取长补短，提出了一种板形影响矩阵-预测控制方法。将该方法应用于某900HC可逆冷轧机进行仿真实验，结果表明，该方法较影响矩阵控制方法有更好的控制效果，是一种有效的板形控制方法。更多还原

【Abstract】 Plate and strip steel, which is the main composition of the steel products and widelyused in automotive, food packaging industry, has significant influence to nationaleconomy. With the progress of science and technology, higher requirement to steel qualitywas made by customers. Flatness is one of the most important quality indexes of strip steel,and flatness controlling technique is the hot topic in the rolling area. In recent years,artificial intelligence has been used widely in industrial process study for its merits inmodeling, optimization and control. This paper choose the cold strip mill flatnessintelligent control based on improved Least Squares Support Vector Regression(LS-SVR)algorithms as research object. On analyzing the merits and defects of the existingintelligent control approach, improved LS-SVR algorithms was studied andcomprehensive research on flatness controlling system was made.First of all, a novel Multi-output Least Squares Support Vector Regression (MLSSVR)approach was proposed to overcome the defects that standard LS-SVR algorithm, whichapplies only to more inputs single output system, can not use on more inputs more outputsindustrial process directly. The MLSSVR algorithm still meets the principle of StructuralRisk Minimization, therefore, keeps good generalization performance. Furthermore, totackle the difficulty in determining the hyper-parameters of MLSSVR, an optimizationmethod based on particle swarm optimization algorithm was adopted. This approach cannot only compute effectively, but also has strong searching ability.Flatness pattern recognition is the key constituent of the flatness control system. Inorder to adapt to the higher demand of flatness controlling, flatness basic patternsexpressed by the linear, quadratic, cubic and quartic Legendre orthogonal polynomial wereproposed. And a novel flatness pattern recognition method based on MLSSVR was putforward. The results of experiment demonstrate that the proposed approach can distinguishthe types and define the magnitudes of the flatness defects effectively with high accuracy,high speed and strong generalization ability.Flatness predictive model is the most important foundation of control system. Inorder to have higher precise flatness predictive model, a MLSSVR flatness predictive model is designed on the basis of measured data in production. Simulation experimentdemonstrates that the MLSSVR flatness predictive model has higher predictive accuracyand strong robustness.Finally, effective matrix--predictive control approach was put forward oncomprehensively analyzing the characteristics of effective matrix control method andpredictive control method and combining the merits of the tow methods. Then, simulationexperiment on testing the performance of the control model was conducted on900HCreversible cold roll. It demonstrates that effective matrix--predictive control approach hasbetter control effect than effective matrix control method, therefore, is an effective flatnesscontrol method.更多还原