热轧C-Mn钢工业大数据预处理对模型的改进作用

发布时间：2018-03-12 15:24

本文选题：大数据　切入点：建模　出处：《钢铁》2016年05期 　论文类型：期刊论文

【摘要】：在应用C-Mn钢工业大数据进行神经网络建模时,如果将大量原始数据不加处理或者经过简单的剔除异常值处理后进行建模,很容易建立满足一定精度要求的模型。但是,如果进一步研究模型的规律性,却常常有违背客观规律的情况。这是由于原始数据中大量的数据相互干扰和生产数据的离散分布造成的。因此在建模过程中,需要将冗余和误差较大的数据剔除,保证训练数据和预测数据的均匀分布,这样能够在减小建模的计算量的同时保证数据具有显著的规律性,从而建立出合理的模型。文章利用Bayes神经网络建立了多种牌号C-Mn钢力学性能预测模型,并对影响屈服强度的工艺参数进行了分析。经统计,屈服强度和抗拉强度的预测数据中分别有96.64%和99.16%的数据预测值和实测值绝对误差在±30 MPa之内,伸长率的预测数据中有85.71%的数据预测值和实测值绝对误差在±4%内。
[Abstract]:When using C-Mn steel industry big data to model the neural network, if a large number of raw data are not processed or the outliers are simply removed, it is easy to establish a model that meets the requirements of certain precision. If we further study the regularity of the model, there is often a violation of the objective law. This is caused by the large amount of data in the original data and the discrete distribution of the production data. Therefore, in the process of modeling, In order to ensure the uniform distribution of the training data and the prediction data, the redundant data and the large error data should be eliminated, so as to reduce the computational cost of the modeling and ensure that the data have remarkable regularity. In this paper, several prediction models of mechanical properties of C-Mn steel grades are established by using Bayes neural network, and the process parameters affecting the yield strength are analyzed. In the yield strength and tensile strength prediction data, the absolute errors of the predicted and measured values of 96.64% and 99.16% are within 卤30 MPa, respectively, and 85.71% of the predicted data of elongation are within 卤4%.
【作者单位】：东北大学轧制及连轧自动化国家重点实验室;鞍钢股份有限公司第三炼钢连轧厂;
【基金】：钢铁联合基金重点资助项目(U1460204) 辽宁省自然科学基金资助项目(2015020180) 中央高校基本科研业务费专项资金资助项目(N140704002)
【分类号】：TG335.11;TP274

【参考文献】