Interpretable Deep Learning Techniques and Meta Information for Automated Decision Making in Ironmaking Plants

Modeling complex processes like ironmaking is a demanding task. Approaches based on machine learning (ML) and especially deep learning (DL) offer a considerable option that can complement traditional first principles approaches. They can model complex connections between many input variables and with the possibility to retrain a model, they can adjust to changing circumstances e.g., furnace operating conditions, sensor drift, changing raw materials. Using the latest artificial intelligence (AI) achievements, a multi-step time series framework was developed to forecast key performance indicators e.g., the silicon content and temperature of hot metal of a blast furnace.
However, ML and in particular DL methods are often seen as black boxes, which struggle with a lack of transparency and interpretability. These factors make it hard for new models to be accepted by domain experts, who do not only need to understand what is going on but also why. Hence the implementation of such models into control systems is challenging because an explanation for a model’s decision is required for trust and decision making. In addition, upcoming artificial intelligence regulatory issues could limit the applicability of such DL methods due to the inherent black box characteristic.
Enhancing a ML or DL model with advanced, explainable AI techniques and meta information makes them more transparent. These methods are essential for interpreting the results and are necessary to fulfil regulatory requirements and for optimized decision-making in the iron and steel industry. In this paper we will investigate the application of ML or DL methods combined with explainable AI techniques and meta information in ironmaking, the requirements for successful implementation and the applicability for automated decision-making systems.