Presentation on September 29, 2016, at the 4th Annual BayesiaLab Conference:
Supervised Learning to Understand Root Causes of Steelmaking Slivers Using BayesiaLab
Floyd E. Demmon
Lead Statistical Engineer, ArcelorMittal Indiana Harbor
Floyd Demmon is a statistical resource in the Quality Department at Arcelormittal USA-Indiana Harbor, in East Chicago, Indiana. He works on projects that cross departmental and facility boundaries within the company. These complex projects are aimed at optimizing quality performance for a broad range of customers and product types.
Steelmaking slivers form as a result of entrapment of non-metallic inclusions in steel as it crystallizes from liquid to solid. The inclusions are exposed, in greatly elongated form, as linear flaws called “slivers” on the finished surface of the steel after thickness reduction. Their presence in steel produced for exposed applications such as car hoods, appliance cabinets, or casket tops results in rejection of the steel, since they are visible even after painting.
Major reductions in both frequency and severity of slivers have been made in the steel industry over the years, but at some juncture each producer “hits a wall” and finds each incremental improvement smaller in magnitude—and typically at increasingly significant cost. Factors that act over an entire batch (heat) of steel have at this point been largely controlled. Remaining causes reside on the scale of inches of cast slab length. At this level of detail, serial correlation and multi-collinearity render traditional statistical tools largely useless at best, misleading or just wrong at worst.
BayesiaLab is proving to be extremely useful for understanding mechanisms at work in the mold of the continuous caster from one second to the next. Supervised learning is being used in this regard. Augmented Markov blankets are used to model the relationships of the large number of variables that potentially impact the entrapment of the precursor defects that later become slivers. This presentation will describe the approach taken, the issues encountered, and outcomes to date. The ultimate aim is to build a model that can be installed on-line to alert operators when the likelihood of an inclusion becoming entrapped has exceeded a threshold that requires operator intervention.