【作者】 李志鹏；

【导师】 陈积明； 程鹏；

【作者基本信息】 浙江大学 ， 控制科学与工程（专业学位）， 2017， 硕士

【摘要】 钢铁冶炼在我国国民经济中占有十分重要的地位,自2014年起,我国的钢铁产量占世界的比重已经高达40%。高炉炼铁是现代炼铁工艺的主要方法,高炉形体庞大、内部反应复杂、生产能耗极高。保证高炉在生产过程中的平稳运行是保持高产率、高能量利用率的基本条件,具有十分重要的意义。本文针对高炉炉况平稳性判别展开研究。目前,对高炉异常炉况的诊断已经有较多工作。先后出现了专家系统、人工神经网络和支持向量机等研究方法。但是,这些工作的关注点主要针对异常炉况发展的后期,是在异常炉况已经较为严重时对异常的原因进行诊断。而早期的炉况平稳性监测在实际生产中依然依赖工程师的个人经验。并且,这些模型的训练方法多为有监督训练,而实际生产中炉况的异常标签很难获得。针对这些问题,本文提出了一套基于混合高斯-朴素贝叶斯的无监督分类模型,着眼于早期的炉况平稳性监测,并且解决了数据无标签或者标签不可信的问题。同时,基于本文提出的炉况判别模型,设计并实现了一套高炉炉况智能判别系统,利用该系统可以针对高炉炉况数据进行可视分析、模型训练和实时监测。首先,本文对所用数据集进行了介绍与分析,对数据集中的异常值和缺失值进行了剔除、插补等处理。通过相关性分析发现了炉况参数数据的第一主成分与炉况的平稳性密切相关,据此选定了第一主成分的均值、标准差和最大值三个表示炉况平稳性的特征。然后,本文构建了一套基于混合高斯-朴素贝叶斯算法的、用于判别炉况平稳性的无监督分类模型,该算法可以在无监督的情况下进行训练,解决了传统模型面临的无标签或标签不可信的问题。本文将模型的性能与决策树、逻辑回归和支持向量机进行对比。本文模型在综合错误率、第一类错误率和第二类错误率等指标上均优于对比模型。最后,本文基于提出的模型,设计并实现了一套高炉炉况智能判别系统,可以用于高炉历史数据可视分析、模型训练和现场数据的实时监测。更多还原

【Abstract】 Iron and steel industry plays an important role in China’s national industry.Since 2014,China’s steel production has accounted for 40%of the world’s production.Blast furnace(BF)is an important object of this industry,with a huge body,complex chemical reactions,consuming much energy.Smooth operation state is a pre-condition of a high production and energy utilization efficiency.That why keeping the BF running in a smooth state is of great importance.Currently,there has been many researches about the irregular state diagnosis of the blast furnace,such as expert system,artificial neural network and support vector machine,but those researches mainly focus on the later period the irregular operation state,which means a diagnosis after the occurrence of the problems.The diagnosis of earlier period still depends on human experiences in the plant.Moreover,the existing researches are basically based on supervising learning methods,which means a set of labels for historical data is necessary.However,there is often not any label set or the label set is not quite reliable in reality.Based on the facts above,this paper propose a unsupervised method based on Gaussian Mixture Model(GMM)and Naive Bayes Classifier(NBC),aimed to the earlier period diagnosis of BF’s irregular operation state,solving the inaccessible label set problem.Firstly,this paper conducts a brief introduction and analysis of the related datasets and deal with the missing and irregular records in the datasets,concluding that the first principal component(FPC)of BF’s parameters is correlated to the stability of the operation state.Expectation,standard deviation and maximum value of FPC are chosen as the features representing the stability.Secondly,this paper constructs a method based on GMM-NBC to diagnose the irregular state,solving the inaccessible label set problem.Decision Tree,Logistic Regression and Support Vector Machine are chosen as the baseline of the proposed method.The proposed method is better all the baselines in terms of the first type error rate,the second type error rate and the comprehensive type error rate.Lastly,this paper builds an irregular state identification system used for historical data analysis,model training and real time surveillance of the BF.更多还原