Safety soft sensor development for pilot-scale ilmenite electric arc furnace using long short-term memory-based architecture

Abstract

Ilmenite electric arc furnaces (EAFs) are used for smelting titanium-iron oxide ore at high temperatures generated by electrical arcs to produce titanium slag and pig iron. As these units are pushed to their limits, ensuring safe and reliable operation becomes challenging. The main risk is molten material leaking out of the unit, known as a run-out. Run-out results from degradation of the EAF sidewalls due to direct contact with high-temperature molten material. To ensure safe operation, the sidewall temperature profile and the thickness of solid material protecting the sidewall are closely monitored by process operators. Operators commonly monitor this using temperature sensors and manual measurements of the molten metal bath size at each tapping cycle. Knowledge of sidewall temperature prediction and estimation of the thickness of solid material protecting the sidewall between tapping cycles can support operators' decision-making process and help avoid run-outs. To achieve this, the present work proposes two data-driven models: (1) a sidewall temperature prediction model and (2) a bath size estimation model. Both models are trained solely on historical process data using long short-term memory (LSTM) and gated recurrent unit (GRU) neural networks. The selected models achieve acceptable accuracy with an averaged mean absolute percentage error (AMAPE) of 0.29% for the sidewall temperature model and a mean absolute percentage error (MAPE) of 9.5% for the bath size model. Furthermore, to ensure model transparency and facilitate industrial adoption, a Shapley additive explanations (SHAP) analysis is integrated to interpret the model outputs and validate physical consistency.