融合工业大数据与冶金机理的微合金钢性能预报模型研究

The accurate prediction of the mechanical properties of steel can help to control product properties, optimize the product design and improve production organization mode. But the coupling and mechanism between mechanical properties and composition process factors are very complex, which makes the mechanical properties prediction problem present high-dimensional nonlinear, so that it is difficult to model with mechanism. Then due to the hot rolling production data with large detection error and low signal noise ratio, the data modeling is also fail to guarantee the generalization ability of the model. A modeling method combining the industrial big data and metallurgical mechanism is proposed in this project. Based on the sub-model of each factor, a mechanical properties prediction model with high reliability and strong generalization for micro-alloyed steel is built. ①The influence factors of the model are selected by the random forest, causality diagram. The mass fractions of micro-alloyed elements carbonitride precipitation are calculated, and the effect of carbonitride precipitation is analyzed; ②The dynamic box mean processing method is used to clean the actual data. Sub-model for each factor is built, and the reproducibility and mechanism compatibility of model rules are verified; ③The interactions between the various factors are explored, and the interactive sub-models are constructed; ④The relationships between the sub-models are analyzed, the mechanical properties prediction model is built to reveal the mechanism of the composition and process parameters on the mechanical properties; ⑤Product design and optimization auxiliary support tools for micro-alloyed steel are developed. The above research provides a new modeling method for complex systems. And the built prediction model can be used to design and optimize the steel grades, which can reduce the number of physical tests and cut down product development costs.

准确预报钢材力学性能，有助于控制产品性能、优化钢种设计、改进生产组织方式。而力学性能与成分工艺各因素耦合、作用机理复杂，使性能预报问题呈现高维非线性，难以用机理建模；热轧生产数据检测误差大、信噪比低，数据建模也难保证模型的外延性。.本课题提出融合工业大数据与冶金机理的建模方法，构建基于各因素子模型的、可靠性高、外延性强的微合金钢性能预报模型：①用随机森林、因果图筛选影响因素，计算微合金元素碳氮析出物质量分数，解析碳氮化物析出的影响；②用动态分箱均值处理清洗数据，构建各因素子模型，验证模型规律的可重现性、机理相容性；③探明各因素间的交互作用，构建有交互作用的子模型；④分析各子模型间的关系，进而构建性能预报模型，揭示成分、工艺对力学性能的作用机理；⑤研制微合金钢产品设计和钢种优化辅助支撑工具。.上述研究为复杂系统提供了新的建模方法，预报模型可用于钢种设计和优化，减少物理试验次数，降低研发成本。

热轧钢材组织性能预报一直是钢铁冶金行业关注的难点问题，是一项十分复杂的冶金前沿技术。微合金钢力学性能与成分工艺各参数高度耦合、作用机理复杂；而用户对模型的实用性要求很高，具备足够精度、可靠性与泛化能力的模型才具有实用价值。因此，本课题提出融合工业大数据与冶金机理的建模方法，借助数据与机理对成分、工艺各因素的影响进行剖析，将复杂的高维非线性问题拆分为若干子问题。本项目主要研究结果为：.1）结合冶金机理、统计方法及人的先验知识，确定性能预报模型的影响因素。将微合金钢的强度看成碳氮化物的强化作用叠加碳锰钢的作用，建立热力学模型计算奥氏体与铁素体相中碳氮化物析出的成分、质量分数，作为模型的影响因素；采用随机森林算法获得成分、工艺各因素的重要性排序，筛选对性能影响较大的因子，实现对模型的降维。.2）用动态分箱均值处理生产大数据，解决数据信噪比低、分布不均匀问题；采用孤立森林算法对异常数据进行清洗，提高建模数据质量；运用广义可加模型，基于Back-fitting算法迭代计算成分工艺各因素对钢材力学性能的影响，通过三次样条函数拟合估计各因素的单变量子模型，验证子模型规律的可重现性、机理相容性，从而构建微合金钢性能预报模型。利用国内某大型热连轧机组生产的含铌微合金钢进行了性能预报实验，抗拉强度、屈服强度的预报误差分别为2.52%、3.27%，揭示了成分、工艺对力学性能的影响规律。.3）以产品成分、工艺参数为决策变量，以产品力学性能最优、合金成本最小为目标，建立微合金钢产品设计的多目标优化模型，研制基于多精英解引导的多目标鲸鱼优化算法进行求解，获得产品成分、工艺参数的Pareto最优方案，研制微合金钢产品设计和钢种优化辅助支撑工具。钢种优化设计实验表明，不仅力学性能指标得到较大提升，而且成本有所降低。.本项目为钢材力学性能建模提供了新的思路，模型可用于新产品设计和钢种优化，从而减少物理试验次数，缩短产品研发周期，降低研发成本；还可用于产品性能在线动态控制，提高产品合格率及力学性能的控制精度。