【作者】 田慧欣；

【导师】 毛志忠； 王安娜；

【作者基本信息】 东北大学 ， 控制理论与控制工程， 2009， 博士

【摘要】 洁净钢生产工艺开发、炉外精炼是国内外钢铁冶金领域的热门研究课题,LF精炼炉是炉外精炼的重要环节,在我国钢铁企业中已经得到了广泛的使用,LF精炼过程可以提高钢的质量、扩大品种,在电炉或转炉一连铸流程中可以调节电炉或转炉与连铸的生产节奏实现多炉连浇。LF精炼炉钢水温度控制对钢的质量和连铸操作的顺行十分重要,可以缩短冶炼时间、降低成本、节约能源、提高精度,而LF钢水温度预报模型是LF钢水温度控制的先决条件,是实现精炼过程计算机自动控制的重要基础。本文以宝钢一炼钢300吨LF精炼炉为研究对象,针对单一使用机理模型或智能模型的不足,提出了将机理方法与智能方法相结合的建模思想。并将集成学习的思想融入到温度预报模型的建模过程中,取得了较好的效果。首先,结合LF精炼炉冶炼过程及其特点,从能量平衡的角度出发,通过对LF精炼炉冶炼过程能量变化的分析,分别建立基于能量平衡的LF精炼炉钢水温度预报机理模型和基于ELM的钢水温度预报智能模型。为建立LF精炼炉钢水温度预报混合模型奠定了基础。单一使用机理模型,存在着部分参数无法准确计算、模型难以被生产现场采用的不足；而单一使用智能模型,则会有过分依赖数据、缺乏工艺指导、预报精度很难提高的不足。针对上述不足,将传统的机理建模思想与智能建模思想相结合,提出了一种串行方式的LF钢水温度预报混合建模方法。用智能方法确定传统机理模型中难以准确获得的参数,再用机理模型来预报钢水温度。并在智能模块中使用了改进的AdaBoost.RT集成学习算法,有效地提高了单一使用ELM的精度。从并行角度出发,提出了基于Bagging的多模型钢水温度预报混合建模方法。将机理模型与智能模型一起作为Bagging的子学习机,以并行的形式建立钢水温度预报混合模型,有效地将智能方法与机理方法结合在一起,在提高智能方法预报精度的同时使得机理模型与智能模型的优势得到互补。此外,针对Bagging集成算法固有的子学习机精度与数据集多样刑不能兼顾的问题,在建模过程中巧妙地将PCA运用到Bootstrap采样后的子训练数据集上,使得整个算法在保证数据集多样性的同时又能够确保每一个子学习机具有较高的精度。在实际应用中,预报模型必须要面对模型更新的问题。针对现有软测量模型更新方法浪费时间、制约软测量模型中智能算法训练性能等不足,提出了一种基于增量学习思想的软测量建模及更新方法。该方法通过使用新数据集的训练得到新的子模型,再将其依据计算出的权重因子与原模型结合起来,实现整个模型的更新。这种新方法有效地利用了历史训练结果,无须保存历史数据,具有节省存储空间,减少后继训练时间,可以实现延续性学习的特点。在上述的建模与更新过程中所使用的数据均认为是没有过失误差的数据,而实际上,现场采集的数据由于各种原因不可避免地存在着过失误差,因此需要对建模数据进行过失误差侦破,以保障软测量模型的测量精度。针对软测量建模数据过失误差侦破的特殊性,本文提出了一种新的CED聚类算法,并形成基于CED的过失误差侦破方法。考虑单纯使用CED聚类进行过失误差侦破会由于侦破过程缺乏指导而产生过度侦破的现象,又提出将该方法与软测量建模过程相结合,在使用过失误差侦破后的数据进行软测量建模的同时用建模误差指导过失误差侦破过程。使用这种方法可以有效地找出数据集中的过失误差,同时还可以得到准确的软测量模型。更多还原

【Abstract】 Secondary refining has become a hot research topic in steelmaking fields at home and abroad. At present, Ladle furnace (LF) is applied widely in the iron and steel industry. LF refining technology plays a substantial role in secondary metallurgic process. The main purpose of ladle furnace treatment is to ensure that the molten steel has the required temperature when the ladle is taken over at downstream secondary metallurgy units or at a continuous caster. Therefore, the temperature control of LF is essential for improving both productivity and product quality. The temperature prediction is a key for the temperature control of LF successfully. Based on the 300t LF in Baoshan Iron & Steel Co. Ltd., the new hybrid modeling method for molten steel temperature prediction in LF is proposed by combining mechanism and intelligent methods, aiming at the shortages of single mechanism and intelligent models. In order to enhance the performance of prediction model, ensemble algorithm is used during the modeling process.The energy conservation is analyzed during the process of LF refining. The mechanism model based on energy balance and the intelligent model based on ELM are established for predicting the molten steel temperature. These models lay a foundation for following hybrid model of molten steel temperature prediction in LF.Aiming at the disadvantages of using single mechanism or intelligent model, a hybrid model for predicting the molten steel temperature in LF is proposed, in which mechanism method and intelligent method are combined. The intelligent method is used to predict the parameters of mechanism model that are calculated hardly. And then, the mechanism model is used to predict the temperature of molten steel. The modified AdaBoost.RT is presented in intellige0t model to enhance the accuracy of single ELM.Focusing on parallel modeling method, a multi-model hybrid modeling method based on Bagging for molten steel temperature prediction is proposed. In this method, mechanism models and intelligent models are used together as weak learners. They are aggregated to establish the prediction hybrid model. Additionally, aiming at the problem of using Bagging algorithm, PCA is used for sub training data set that has been sampled by Bootstrap. By using PCA, the diversity of data set and the accuracy of every weak learner are all considered carefully.Aiming at the shortages of existing updating method for soft sensor model, such as wasting time, restricting the performance of intelligent algorithm, a new updating method for soft sensor model based on incremental learning is proposed. In the new method, the new data is used to train new model firstly. And then the new model is combined with the old models according to their weights, and the update for whole model is completed. The new updating method using the past training results, and need not the old data any more. So it is benefit to save space and training time by using the new updating method.During above processes of modeling and updating, the training data sets are supposed as the data sets without gross error. Actually the gross errors are inevitable during the process of data acquisition. Therefore a conclusion can be drawn that it is necessary to detect the gross errors of modeling data. A new clustering method is proposed for detecting the gross errors in order to ensure the accuracy of soft sensor model. Considering over-detection will exist during the process of detecting, the new method is combined with the modeling. The modeling errors are used to supervise the gross error detection. This new method can detect the gross errors effectively and obtain a good soft sensor model.更多还原