Being able to predict future temperatures on the wall lining is key when controlling and scheduling maintenance for large industrial smelting furnaces. In this paper, we propose and test a machine learning approach for predicting lining temperatures in a ferronickel smelting furnace. This approach was deployed and evaluated in a real-world scenario, i.e., in one of Cerro Matoso S.A.'s (CMSA) industrial plant furnaces. Different techniques were tested, and finally, a multitarget regression (MTR) model showed the best performance. Previous state of the art focused on predicting only one target sensor; in contrast, our model is capable of predicting up to 12 targets. Two MTR models were tested: the incremental structured output prediction tree (iSOUP-Tree) and the stacked single-target Hoeffding tree regressor (SST-HT). The SST-HT method had the best behavior in terms of the average mean absolute error (AMAE) and average root mean square error (ARMSE). The results indicate that the developed MTR models can accurately predict the measured temperature on multiple point sensors. Results of this work are expected to help the process of structural control and health monitoring of the furnace linings located at CMSA's plant.

中文翻译：

---

屏蔽电弧炉耐火内衬的监测：在线多目标回归树方法

在控制和安排大型工业熔炼炉的维护时，能够预测壁衬的未来温度是关键。在本文中，我们提出并测试了一种机器学习方法，用于预测镍铁熔炼炉的炉衬温度。这种方法在真实场景中进行了部署和评估，即在 Cerro Matoso SA (CMSA) 的一个工业工厂熔炉中进行了部署和评估。测试了不同的技术，最后，多目标回归 (MTR) 模型显示出最佳性能。以前的技术水平专注于仅预测一个目标传感器；相比之下，我们的模型能够预测多达 12 个目标。测试了两个 MTR 模型：增量结构化输出预测树 (iSOUP-Tree) 和堆叠单目标 Hoeffding 树回归器 (SST-HT)。SST-HT 方法在平均平均绝对误差 (AMAE) 和平均均方根误差 (ARMSE) 方面表现最佳。结果表明，所开发的 MTR 模型可以准确地预测多点传感器上的测量温度。预计这项工作的结果将有助于对位于 CMSA 工厂的炉衬进行结构控制和健康监测。