Industry 4.0 provides an opportunity for realizing converter intelligent steelmaking. Based on a cyber-physical system (CPS) framework for steelmaking plants, we developed a non-contact intelligent prediction model for determining molten steel carbon content. We achieved this by leveraging big data in-depth learning and optical information from converter mouth flame characteristics to realize accurate control of end-point carbon content in the process of intelligent steelmaking. The spectral information of a converter port flame was gathered using a USB2000+ spectrometer, and its main characteristics were extracted via factor analysis. The sampling frequency of the converter flue gas analysis mass spectrometer was set to obtain the carbon content of continuous molten steel during the late stages of steelmaking consistent with the spectral information collection frequency. A large sample dataset was constructed to continuously predict the carbon content of the molten steel based on the flame spectrum information of the furnace orifice. Furthermore, the parameters of a support vector regression algorithm were optimized, and the quantitative relationships among the sample data were intelligently mined to obtain a dynamic prediction model of the carbon content of 150 furnace samples. To improve the universality of the model, the initial steelmaking conditions were applied as the optimization criteria of the carbon content prediction model generated by matching different furnace samples. A support vector machine algorithm was applied to construct a model optimization classifier to obtain a universal prediction model. The results indicate that in the late stages of smelting, the spectral information of the flame in the furnace exhibited a high correlation with the carbon content of the molten steel. After testing, the prediction accuracy of this model was more than 90%. With the help of big data deep learning and CPS framework, we can build robust and accurate steelmaking intelligent control systems.

工业4.0为实现转炉智能炼钢提供了契机。基于炼钢厂的信息物理系统（CPS）框架，我们开发了一种用于确定钢水碳含量的非接触式智能预测模型。我们通过利用大数据深度学习和转炉口火焰特性的光学信息来实现智能炼钢过程中终点碳含量的精确控制。光谱信息使用USB2000+光谱仪采集转换器端口火焰的特征，并通过因子分析提取其主要特征。设定转炉烟气分析质谱仪的采样频率，以获得与光谱信息采集频率一致的炼钢后期连续钢水的碳含量。构建大样本数据集，根据炉口火焰光谱信息连续预测钢水碳含量。此外，a的参数优化支持向量回归算法，智能挖掘样品数据间的定量关系，得到150个炉样碳含量动态预测模型。为提高模型的通用性，采用初始炼钢条件作为匹配不同炉样生成的含碳量预测模型的优化准则。支持向量机应用算法构建模型优化分类器，得到通用的预测模型。结果表明，在冶炼后期，炉内火焰的光谱信息与钢水碳含量呈现出高度相关性。经测试，该模型的预测准确率在90%以上。借助大数据深度学习和CPS框架，我们可以构建稳健、精准的炼钢智能控制系统。