Machine learning (ML) is a promising modeling framework that has previously been used in the context of optimizing steel processes. However, many of the more advanced ML models, capable of providing more accurate predictions to complex problems, are often impossible to interpret. This makes the domain experts in the steel industry, to a large extent, hesitant to adopt these models. The valuable increase in model accuracy is diminished by the lack of model interpretability. Herein, Shapley additive explanations (SHAP) is applied to an advanced ML model, predicting the electrical energy (EE) consumption of an electric arc furnace (EAF). The insights from SHAP reveal the contributions from each input variable on the EE for every single heat in the prediction domain. These contributions are then evaluated based on process metallurgical experience.
The non-linearity of the Electric Arc Furnace (EAF) process and the correlative behavior between the process variables impose challenges that have to be considered if one aims to create a statistical model that is relevant and useful in practice. In this regard, both the statistical modeling framework and the statistical tools used in the modeling pipeline must be selected with the aim of handling these challenges. To achieve this, a non-linear statistical modeling framework known as Artificial Neural Networks (ANN) has been used to predict the Electrical Energy (EE) consumption of an EAF producing stainless steel. The statistical tools Feature Importance (FI), Distance Correlation (dCor) and Kolmogorov–Smirnov (KS) tests are applied to investigate the most influencing input variables as well as reasons behind model performance differences when predicting the EE consumption on future heats. The performance, measured as kWh per heat, of the best model was comparable to the performance of the best model reported in the literature while requiring substantially fewer input variables.
The melting time of scrap is a factor that affects the Electrical Energy (EE) consumption of the Electric Arc Furnace (EAF) process. The EE consumption itself stands for most of the total energy consumption during the process. Three distinct representations of scrap, based partly on the apparent density and shape of scrap, were created to investigate the effect of scrap on the accuracy of a statistical model predicting the EE consumption of an EAF. Shapley Additive Explanations (SHAP) was used as a tool to investigate the effects by each scrap category on each prediction of a selected model. The scrap representation based on the shape of scrap consistently resulted in the best performing models while all models using any of the scrap representations performed better than the ones without any scrap representation. These results were consistent for all four distinct and separately used cleaning strategies on the data set governing the models. In addition, some of the main scrap categories contributed to the model prediction of EE in accordance with the expectations and experience of the plant engineers. The results provide significant evidence that a well-chosen scrap categorization is important to improve a statistical model predicting the EE and that experience on the specific EAF under study is essential to evaluate the practical usefulness of the model.
Predictive models are used in many different fields of science and engineering and are always prone to make faulty predictions. These faulty predictions can be more or less malignant depending on the model application. We describe fibers of failure (FiFa), a method to classify failure modes of predictive processes. Our method uses Mapper, an algorithm from topological data analysis (TDA), to build a graphical model of input data stratified by prediction errors. We demonstrate two ways to use the failure mode groupings: either to produce a correction layer that adjusts predictions by similarity to the failure modes; or to inspect members of the failure modes to illustrate and investigate what characterizes each failure mode. We demonstrate FiFa on two scenarios: a convolutional neural network (CNN) predicting MNIST images with added noise, and an artificial neural network (ANN) predicting the electrical energy consumption of an electric arc furnace (EAF). The correction layer on the CNN model improved its prediction accuracy significantly while the inspection of failure modes for the EAF model provided guiding insights into the domain-specific reasons behind several high-error regions.
The present work focuses on predicting the steel melt temperature following the vacuum treatment step in a vacuum tank degasser (VTD). The primary objective is to establish a comprehensive methodology for developing and validating machine learning (ML) models within this context. Another objective is to evaluate the model by analyzing the alignment of the SHAP values with metallurgical domain expectations, thereby validating the model’s predictions from a metallurgical perspective. The proposed methodology employs a Random Forest model, incorporating a grid search with domain-informed variables grouped into batches, and a robust model-selection criterion that ensures optimal predictive performance, while keeping the model as simple and stable as possible. Furthermore, the Shapley Additive Explanations (SHAP) algorithm is employed to interpret the model’s predictions. The selected model achieved a mean adjusted of 0.631 and a hit ratio of 75.3% for a prediction error within ±5 °C. Despite the moderate predictive performance, SHAP highlighted several aspects consistent with metallurgical domain expertise, emphasizing the importance of domain knowledge in interpreting ML models. Improving data quality and refining the model framework could enhance predictive performance.
The present work aims to answer three essential research questions (RQs) that have previously not been explicitly dealt with in the field of applied machine learning (ML) in steel process engineering. RQ1: How many training data points are needed to create a model with near-upper-bound predictive performance on test data? RQ2: What is the near-upper-bound predictive performance on test data? RQ3: For how long can a model be used before its predictive performance starts to decrease? A methodology to answer these RQs is proposed. The methodology uses a developed sampling algorithm that samples numerous unique training and test datasets. Each sample was used to create one ML model. The predictive performance of the resulting ML models was analyzed using common statistical tools. The proposed methodology was applied to four disparate datasets from the steel industry in order to externally validate the experimental results. It was shown that the proposed methodology can be used to answer each of the three RQs. Furthermore, a few findings that contradict established ML knowledge were also found during the application of the proposed methodology.