Analysis and modelling of steel hot rolling assisted by machine learning

热变形是钢板制造中的关键阶段，在该阶段结束时的要求是厚度控制。 阿尔戈马热轧机有一个传统的设置七个轧制“站”定位在一条线上，一个接一个。 铸钢“连铸坯”被加热到高温，然后通过这些机架，在每个机架上进行减薄，最后一个机架达到所需的厚度。由于通过所有这些机架的行程时间非常短，因此每个机架的辊缝必须在轧制前预先设定。 无论轧制机架多么“刚性”，辊缝总是会变大，因为即使在高温下，钢也非常坚固。 为了预先设定辊缝，通过数学轧机“模型”对钢的强度进行精确预测，该模型通常基于对每个机架的显微组织和温度变化的预测。轧机模型是半经验的;因此，无论何时轧制钢，测量的轧制负荷都可以用于改进模型的预测。 轧机建模的完全经验方法基于机器学习。 在这里，模型是“纯粹”基于测量数据输入（如辊速）生成的。 不需要过程的物理知识。机器学习意味着开发的模型可以“直观地”链接数据，而不是纯粹的统计方法。 该建议建立在最近开发的阿尔戈马的机器学习模型上。 机器学习模型的第一次迭代远远上级当前的阿尔戈马轧机模型，但是，对第一个轧制机架的预测不如对其余机架的预测准确。 此外，模型的分析揭示了一些非常令人困惑的工艺参数和轧制负荷之间的相关性。 在本建议中，我们将改进对初始轧制机架的预测。 我们还将试图了解上述过程/轧制负荷的相关性，以便进一步了解该过程的冶金学。 我们将通过开发特定于阿尔戈马过程的算法来进一步改进该模型。 最后，我们将生成一个模型，可以预测热轧带钢的轧制工艺参数的机械性能。

Hot deformation is a key stage in the manufacturing of steel sheet, the requirement at the end of this stage being thickness control. The Algoma hot rolling mill has a conventional set up of seven rolling 'stands' positioned in a line, one after the other. The cast steel 'strand' is heated to a high temperature, and passes through these stands, undergoing thickness reduction at each stand, with the required thickness achieved by the last stand. Because the travel time through all these stands is very short, the roll gaps of each stand has to be pre-set before rolling. No matter how 'rigid' the rolling stands are, the roll gaps will always widen because, even at high temperatures, steel is very strong. To pre-set the roll gap, an accurate prediction of the strength of the steel is performed by a mathematical mill 'model', which is conventionally based on predictions of how the microstructure and the temperature changes at each stand. The mill model is semi-empirical; thus, whenever the steel is rolled, the measured rolling load can be used to improve the prediction of the models. A completely empirical approach to mill modelling is based on machine learning. Here, the model is generated 'purely' based on measured data inputs, such as roll speed. No knowledge of the physics of the process is required. Machine learning means that the model is developed that links the data 'intuitively' as opposed to a purely statistical approach. This proposal builds on a machine learning model recently developed for the Algoma for strip. The first iteration of the machine learning model is far superior to the current Algoma mill model However, predictions for the first rolling stands were not as accurate as for the rest to the stands. Also, analysis of the model revealed some very metallurgically puzzling correlations between process parameters and the rolling loads. In this proposal, we will improve the predictions for the initial rolling stands. We will also seek to understand the aforementioned process/rolling load correlations in order to gain further insight into the metallurgy of the process. We will further improve the model by developing algorithms that are specific to the Algoma process. Finally, we will generate a model that can predict mechanical properties of the as-hot rolled strip from rolling process parameters.