基于火焰彩色纹理特征的转炉炼钢碳含量预测

Carbon content prediction of converter steelmaking endpoint based on quaternion direction statistics of flame image

连续实时测量碳含量是转炉终点钢水碳含量控制的关键。针对转炉终点炉口火焰图像相似性高导致难以提取具有较强区分性火焰图像特征的问题,提出一种图像彩色纹理特征提取方法——四元数方向统计量(QDS)算法。首先,利用四元数乘法特性对火焰图像进行旋转操作得到火焰图像的四元数旋转图谱;其次,以旋转图谱指定方向和距离上像素差值的四元数相位为投影轴,并以差值的四元数幅值为权重进行统计得到特征;最后,将提取的特征归一化后作为火焰图像最终的彩色纹理特征。在提取火焰图像特征后,使用K近邻(KNN)回归模型对碳含量进行预测,碳含量预测在0.01%误差范围内的准确率为83.7%,在0.02%误差范围内准确率达到88.2%。

Continuous real-time measurement of the carbon content of molten steel in the molten pool is the key to controlling the endpoint carbon content of the converter.The high similarity of flame images at the end of converter makes it difficult to extract flame image features with strong discrimination.To solve this problem,the Quaternion Direction Statistics(QDS)algorithm for image color texture feature extraction was proposed.The quaternion multiplication property was used to rotate the flame image to obtain the quaternion rotation map of the flame image.The quaternion phase of pixel difference in the specified direction and distance of the rotation map were taken as the projection axis,and the quaternion amplitude of the difference was taken as the weight to carry out statistics to obtain characteristics.The histograms at various angles and distances were connected in series and normalized as color texture features of flame images.After the extraction of the flame image features,K-Nearst Neighbors(KNN)regression model was used to predict the carbon content.The accuracy of carbon content prediction was 83.7%within the error range of 0.01%and 88.2%within the error range of 0.02%.