热轧C-Mn钢工业大数据预处理对模型的改进作用  

Hot rolling C-Mn steel industry big data preprocessing for improvements on the model

作　　者：吴思炜 周晓光[1] 曹光明[1] 史乃安[2] 刘振宇[1] 王国栋[1] 

机构地区：[1]东北大学轧制及连轧自动化国家重点实验室,辽宁沈阳110819 [2]鞍钢股份有限公司第三炼钢连轧厂,辽宁鞍山114021

出　　处：《钢铁》2016年第5期88-94,100,共8页Iron and Steel

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)

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。It is easy to construct a model that meets a certain requirement of precision through neural network base on big data of C-Mn steels. In this case,the original industry data are usually without preprocessed or preprocessed by remove the abnormal value simply. However,there will comes a situation that is contrary to the objective laws if the regularity of the model is further studied. This is due to a large amount of data in the original data to interfere with each other and the discrete distribution of the industry data. Therefore,in order to construct a reasonable model,redundant and large error data must be removed,while the distribution of train data and prediction data must be uniform. In this way,the amount of calculation of the model is reduced while a significant regularity of data is excavated. For the sake of verify the hypothesis of ways to use big data,Bayes regularization neural network was selected to construct a model for mechanical properties of multi- steel number. At the same time,the process parameters which influence on yield strength were analyzed. By statistics,the prediction accuracies of yield strength and tensile strength data are 96.64% and99.16%,respectively,of which the absolute error between the predicted value and the measured value lies in the ±30MPa. Among the predicted data of the elongation rate,85.71% of the data absolute error between predicted value and measured value is within ±4%.

关 键 词：大数据 建模 Bayes神经网络 C-MN钢 屈服强度 

分 类 号：TG335.11[金属学及工艺—金属压力加工] TP274[自动化与计算机技术—检测技术与自动化装置]