一种QCN火焰图像特征提取的转炉炼钢碳含量实时预测方法

A real-time prediction method of carbon content in converter steelmaking based on QCN flame image feature extraction

钢水碳含量是影响转炉炼钢出钢质量和冶炼效率的主要因素,而实现碳含量的连续实时预测是炼钢终点控制的关键和难点.针对不同碳含量对应火焰图像呈现出的随机自然纹理相似性较高的问题,根据炉口火焰纹理独有的多方向多尺度不规则特点,提出了四元数复杂网络(QCN)彩色纹理描述符.首先,利用四元数等距映射融合火焰图像颜色通道信息,且用幅值谱和二次量化后的相位谱描述映射后图谱以增强颜色信息描述;其次,采用复杂网络的方式,以相位为条件并利用幅值信息构造一种网络连接权重公式,于幅值谱构建炉口火焰图像的多尺度不规则彩色纹理复杂网络模型;最后,计算复杂网络的拓扑特征度和聚类系数,以其相关特征量化复杂网络拓扑连接模式,构建炉口火焰QCN特征,并通过KNN回归模型预测终点碳含量.结果表明,碳含量在±0.01%误差范围内的预测准确率为85.65%,在±0.02%误差范围内预测准确率达到91.83%,且所提算法满足实时性要求.

The carbon content of molten steel is the main factor that affects the tapping quality and the smelting efficiency, and realizing continuous real-time prediction of carbon content is the key and difficulty to the endpoint control of steelmaking. Aiming at the problem of the highly similar random natural textures in flame images corresponding to different carbon contents, and according to the unique multi-directional and multi-scale irregular characteristics of the furnace mouth flame texture, the quaternion complex network(QCN) algorithm for image color texture feature extraction is proposed. Firstly, using the quaternion equidistant mapping to fuse the color channel information of the flame image, and using the amplitude spectrum and the secondary-quantized phase spectrum to describe the mapped spectrum to enhance the color information description. Secondly, using a complex networks method, a weighted formula of network connection that with phase and amplitude information is constructed, and construct a complex network model of the multi-scale irregular color texture of the furnace mouth flame image amplitude spectrum and finally, the topological feature degree and clustering coefficient of the complex network is calculated, the complex network topology connection mode is quantified with related features, the furnace mouth flame QCN feature is constructed, and the end carbon content is predicted by the KNN regression model. The results show that the prediction accuracy rate of carbon content within the ± 0.01% error range is85.65% and the prediction accuracy rate reaches 91.83% within the ± 0.02% error range, and the proposed algorithm meets the real-time requirements.