徐钢, 黎敏, 徐金梧. 机器学习在深冲钢质量自动判级中的应用[J]. 工程科学学报, 2022, 44(6): 1062-1071. DOI: 10.13374/j.issn2095-9389.2021.05.08.002
引用本文: 徐钢, 黎敏, 徐金梧. 机器学习在深冲钢质量自动判级中的应用[J]. 工程科学学报, 2022, 44(6): 1062-1071. DOI: 10.13374/j.issn2095-9389.2021.05.08.002
XU Gang, LI Min, XU Jin-wu. Application of machine learning in automatic discrimination of product quality of deep drawn steel[J]. Chinese Journal of Engineering, 2022, 44(6): 1062-1071. DOI: 10.13374/j.issn2095-9389.2021.05.08.002
Citation: XU Gang, LI Min, XU Jin-wu. Application of machine learning in automatic discrimination of product quality of deep drawn steel[J]. Chinese Journal of Engineering, 2022, 44(6): 1062-1071. DOI: 10.13374/j.issn2095-9389.2021.05.08.002

机器学习在深冲钢质量自动判级中的应用

Application of machine learning in automatic discrimination of product quality of deep drawn steel

  • 摘要: 在流程工业中,生产过程需根据客户对产品质量要求进行判级,以满足客户提出的产品质量需求。目前,企业主要采用“事后”抽检方式,但因无法对所有产品实现在线自动判级,常发生索赔和退货,导致我国钢铁企业每年近100亿元损失。为了实现产品质量在线自动判级,提出基于高维数据非线性同等缩放与核简支集类边界确定相结合的质量在线智能判级方法。首先,将高维的工艺参数通过非线性同等缩放算法变换成低维的数据集,并对缩放后数据集进行聚类,分析工艺参数的类分布特征。然后,根据分类后样本的质量指标值分布,采用核简支集类边界算法来确定不同产品质量级别的类边界。最后,依据已确定的类边界,通过质量指标预测实现产品在线判级。通过深冲钢(IF钢)应用实例,证实该方法在训练阶段的在线自动判级准确率达到97.2%,测试阶段的准确率为96%。

     

    Abstract: In process industries, the discrimination of final product quality must be implemented in the manufacturing process. At present, the primary method is “after spot test ward,” but there is no other way to realize online automatic discrimination for all products, which frequently leads to customer return purchases and complaints about the product quality, and annual economic loss of 10 billion Yuan in Chinese steel enterprises. This paper proposed online product quality automatic discrimination method based on machine learning to realize online automatic discrimination for all products. First, multidimensional process parameters were mapped into a low-dimensional data set using nonlinear multidimensional parity scaling (MDPS), and the data set is clustered. The distribution feature in the data set was analyzed. The quality index values were then transformed into a low-dimensional map with the class labels determined by process parameter clustering, and the diverse class margins were determined using a support vector machine (SVM) with L2-soft margins. The kernel method set was used to reduce the number of support vectors to simplify the class boundary, and the reduced set determined the actual class margins. Finally, the quality indexes were predicted using machine learning algorithms, such as back-propagation network (BPN), long short-team memory (LSTM), kernel partial least squares (KPLS), and k-nearest neighbors (KNN), including the online automatic discrimination of product quality was realized using the determined class margins and the predicted values of quality indexes. The accuracy of the online automatic discrimination of steel types is up to 97% in the training stage and up to 96% in the testing stage based on industrial production data of interstitial-free (IF) steel.

     

/

返回文章
返回