Data analytics of temperature profiles in induction heating prior to steel extrusion

Sammanfattning

Predicting continuous output signals from heterogeneous observational data –where input sources differ not only in content but in their fundamental temporal structure– poses a significant challenge in applied data science. This study investigates the extent to which preprocessing and structural representation decisions determine the predictability of such signals, and which input variables and temporal representations retain the most predictive information. The experimental work was conducted in collaboration with Alleima, a Swedish metallurgical company, using real production data from the induction heating stage that precedes hot steel extrusion – a setting that naturally combines static process parameters with time-varying sensor measurements. The proposed pipeline encompasses signal segmentation of the output profile, three input aggregation strategies of increasing domain dependency, and dimensionality reduction of the target signal prior to modeling. A total of 3.004 billet heating observations were used across three aggregation strategies –one of which was further combined with dimensionality reduction of the target signal– yielding four experimental configurations thatwere systematically compared. The best-performing configuration achieved an RMSE of 10.55◦C and an R2 of 0.74. ASHAP-based feature importance analysis consistently identified a small set of dominant predictors across experimental configurations, with the most informative temporal representations differing from those initially anticipated by domain experts – a finding that highlights the value of data-driven feature evaluation in heterogeneous input settings. The results demonstrate that preprocessing design is the primary driver of predictive performance in this class of problem, and that domain-informed aggregation strategies yield measurable improvements over purely statistical alternatives. The methodology and findings generalize beyond the specific industrial context studied here, with implications for any supervised regression setting involving heterogeneous data sources of differing temporal structure.