关键词: 转炉炼钢, 碳含量预测, 四元数图像处理, 四元数方向统计量, 彩色纹理特征

Abstract: 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%.

Key words: converter steelmaking, prediction of carbon content, quaternion image processing, quaternion directional statistics, color texture features