With the overarching goal of understanding inner dynamics of the blast furnace and electric arc furnace, several perspectives of achieving this goal were established. These are focused around: (1) the detection of deviations from normal operating conditions, (2) modeling of the temporal behavior of the thermal and chemical properties during blast furnace operation (3) characterization of the raw material in electric arc furnace operation and (4) prediction of deviations from established EAF process models. These goals are achieved through the application of data analytics techniques and the development of suitable machine learning and deep learning models based on the findings from the evaluation of existing process and control data. The results provided by these machine learning models should be supported by explainable AI methods to provide a reasoning for the results of the algorithmic decision making. In addition, large emphasis is placed on utilization of causality-based methods which utilize observational process data for discovery of underlying causal generative mechanisms. Inputs from these methods not only improve the performance and transparency of machine learning models, but also drive a knowledge discovery process.

In order to model the behavior of these Ironmaking processes several approaches are utilized. In practice, when dealing with industrial data, there are distinct challenges like: discrepancies arising from sensor and equipment wear and tear, periods of shutdown, correlated data, data imbalance, missing values, and temporal inconsistencies due to manual input. Data pre-processing phase comprises three steps: data cleaning, filtering out downtime periods, and finally adjusting data through domain-specific rules (see Figure 1).
In the modeling phase, due to transparency requirements, methods such as RandomForest, Gaussian Regression, and Support Vector Machine were used to establish a baseline performance. This baseline performance was compared with more complex models based on LSTM mechanism, Autoencoders, and even more complex approaches such as Temporal Fusion Transformers to establish performance-transparency tradeoff. From the perspective of causal modeling, we utilized methods such as: FCI, LiNGAM and ANM. Additionally, methods such as PCMCI were used to model the temporal behavior of the previously mentioned ironmaking processes. Finally, state-of-art approaches for xAI (SHAP, Anchors, LIME, Ceteris Paribus, etc.) were utilized to discover explanations on the local and global level. Furthermore, from the perspective of reliability of the ML models, uncertainty metrics can be utilized to address issues with the model parameters (epistemic uncertainty). Confidence intervals and Trust Scores are some of the methods which were utilized in identification of ML model uncertainty.

The overall objective of the project is to develop a transparent approach for modeling and understanding certain aspects of the process behavior in ironmaking processes, i.e., blast furnace and electric arc furnace operation. In the first stage of the project focus is put on modeling and predicting thermal behavior of the blast furnace. In order to achieve this goal, we developed a transparency approach, which is based on the well-known CRISP-DM methodology (see Figure 2) and takes into account intricate specifics of the blast furnace operation. This approach goes beyond development of a ML model for predicting thermal state indicators and provides an overall set of methods which should be utilized to ensure stable operation and utilization of ML model results.
During the model development, methods for causal discovery and global explainability have been deployed to ensure valid basis model training. Furthermore, we proposed additional model performance metrics developed from the case specific blast furnace operation requirements. Additional model performance metrics provided an insight into ML model stability and additional basis for model evaluation and finally, ML model selection. We identified use case appropriate methods for monitoring data provided to the ML model and reducing model uncertainty through concept drift detection. Finally, we proposed a collection of methods that, once deployed, provide an insight into the reasoning behind model prediction after inference time through model simplification and counterfactual instance explanation. By providing model simplification and counterfactual explanations, additional insights are provided, not only into the reasoning behind the ML model, but also into the complex blast furnace process. This additional layer of transparency allows the extraction of various process rules and overall inference logic.