基于低碳钢连铸坯低倍组织图像颜色特征的碳含量预测综合模型  

Carbon content prediction comprehensive model based on color feature of macrostructure image in continuous casting billet of low-carbon steel

作　　者：侯自兵[1] 郭坤辉 岑叙 朱程赫 HOU Zibing;GUO Kunhui;CEN Xu;ZHU Chenghe(College of Materials Science and Engineering,Chongqing University,Chongqing 400044,China)

机构地区：[1]重庆大学材料科学与工程学院,重庆400044

出　　处：《冶金自动化》2024年第6期2-10,47,共10页Metallurgical Industry Automation

基　　金：国家自然科学基金面上项目(52274318)。

摘　　要：针对现有连铸坯碳含量检测方法的不足,前人基于低倍组织图像灰度与碳含量的对应关系尝试建立了高碳钢连铸坯碳含量预测指数函数模型,但对于碳偏析程度更加难以表征的低碳钢还尚未研究。低碳钢局部区域偏析程度明显,最大偏析指数大于3.0,故有必要展开碳含量高效表征研究。选取典型低碳钢连铸坯试样,首先建立基于低倍组织图像灰度的碳含量预测指数函数模型,函数拟合结果R^(2)系数为0.62,平均相对误差(average relative deviation,ARD)为29.7%。然后,建立了基于低倍组织图像颜色参数的碳含量预测神经网络模型,训练结果ARD为19.5%。最后,结合神经网络模型和指数函数模型预测结果特点,构建了碳含量预测综合模型,测试集数据的ARD得到了进一步降低,为14.27%。同时,预测结果平均值与电子探针检测结果平均值的相对偏差为3.43%,与常用的碳含量检测方法相比,误差已基本达到相同的数量级,部分预测结果误差已低于常用检测方法。由于低倍组织图像及其颜色信息获取过程操作简单、成本低、像素尺度可处于微米级且获取范围可针对整个连铸坯断面或大区域范围,故本模型可为类似钢种连铸坯中碳元素偏析状况精细自动检测和评级以及数字化智能分析提供参考。In view of the shortcomings of the existing carbon content detection methods of continuous casting billets,the predecessors tried to establish an exponential function model for predicting the carbon content of high-carbon steel in continuous casting billets based on the corresponding relationship between the grayscale of the macrostructure image and the carbon content,but the degree of carbon segregation is more difficult to characterize for low-carbon steel.The partial segregation degree of low carbon steel is obvious,and the maximum segregation index is greater than 3.0,so it is necessary to carry out efficient characterization of carbon content.A typical low-carbon steel continuous casting billet sample was selected and an exponential function model for carbon content prediction based on macrostructure image grayscale was established.The R-square coefficient of the function fitting result was 0.62,and the average relative deviation(ARD)was 29.7%.Then,a carbon content prediction neural network model based on the color parameters of macrostructure images was established,and the ARD of the training results was 19.5%.Finally,a comprehensive model for carbon content prediction was established based on the characteristics of the prediction results of the neural network model and the exponential function model.The ARD between the prediction results of the test set of comprehensive model and the results of electron probe detection was 14.27%.The relative deviation between the average value of the prediction results and the average value of the electron probe detection results is 3.43%.Compared with the commonly used carbon content detection methods,the error has basically reached the same order of magnitude,and some prediction results have lower errors than the commonly used detection methods.Since the macrostructure image and its color information acquisition process are simple to operate,the cost is low,the pixel scale can be at the micron level,and the acquisition range can be targeted at the entire continuous cast

关 键 词：连铸坯 低倍组织 碳偏析 颜色特征 神经网络模型 自动评级 

分 类 号：TG142.1[一般工业技术—材料科学与工程] TF777[金属学及工艺—金属材料] TP391.41[金属学及工艺—金属学]