Reduction of coil downgrading based on surface defects such as shells and slivers

Customer characteristics and project scope:

Our customer, an Americas-based flat rolling producer runs a mini mill with a capacity of about 1.5 MTPA. The customer is facing an overall coil downgrading and scrape rate of 3%, whereof 25% of downgraded coils show shell-like defects and slivers. Taking into account reduced market prices of hot rolled band (about 210 USD/ton) and potential warranty costs (about 350 USD/ton), we identified with our customer a significant cost saving potential that was estimated with 1.4 million USD per annum considering a split half of shell- and sliver-based defects. The customer requested a machine learning-based approach to identify root causes of downgrading and derive improved adapted PDI (Process Data Input)-settings.


Consulting approach and activities:

Our consulting activities consisted out of a three-fold approach that started with the calculation of the overall cost saving potential and a short feasibility study targeting to identify machine learning-based root causing of coil downgrading.

Based on a positive outcome of the short feasibility study and a positive financial use case evaluation of 1.4 million USD per annum, a machine learning model was trained that used one year of production and quality data, structured from a variety of input sources and features:

  • Product Data – Order details, product hierarchy, customer specifications
  • Production Data – PDI-settings, process control parameter set points
  • Process data – Production stage wise and product-segment level process data (L1/ L2 data)
  • Quality data - labels, targets, tolerances, geometry, etc.
  • Context data – Environment, operation crew, weather, etc.

The machine learning model derived out of more than estimated 400 features, the most relevant impact factors on shell and sliver defects. As a result, about 35 features where identified with a significant, highly correlative impact.

Based on an adapted model taking into account the most relevant features, potential PDI settings where edited and tested to reduce defect rate resulting in new recommended PDI ranges to control product quality.


Benefits and results:

An upfront cost saving calculation and feasibility study ensured a positive project indication for both perspectives, business-wise and technically. With 1.5 million USD per annum, a strong cost saving lever was identified. As we staffed our consulting team interdisciplinary with casting and hot rolling process experts, data engineers and data scientists, we were able to form together with the customer`s experts a competent and fast-moving team.

As the machine learning model resulted in adapted PDI settings for product quality control, our customer implemented the solution within nine months after initial project start and runs the model on regular basis to keep PDI setting improvements up to date based on changing equipment conditions, production planning and order portfolio.