Adopting a machine learning-based workflow can dramatically reduce production risk and cost—but it takes the right practices and mindset.
That was the message underscored at last week’s Digital Transformation Forum in Pittsburgh, where Fero Labs Head of Operations Pamir Ozbay and Carter Almquist from Gerdau’s Petersburg mill took the stage to discuss Gerdau’s experience using Fero’s factory optimization software.
We already took a look at Gerdau’s impressive results, including $3 savings per ton of steel; over 500,000 pounds of manganese, carbon, niobium and vanadium saved; and 15% reduction in quality variation. Here, we take a look at 5 best practices the team used to hit those goals:
1. Focus on training and collaboration. Going from project to deployment is a change management challenge. Even the most accurate ML model won’t benefit the plant if users can’t—or won’t—use it.
Multiple stakeholders across the plant will be involved in deploying ML. Identify these people and engage them as early in the process as possible. As an added benefit, white-box ML can help with the training process by facilitating knowledge transfer, such as helping new operators learn how certain alloys impact the mechanical properties of the final steel.
2. Make software easy to access. Sometimes the simplest solution can have an outsized impact. By observing the user workflow and making adjustments accordingly, Gerdau was able to significantly improve the effectiveness of Fero. After noticing that operators were forgetting to follow Fero's live recommendations, the company installed a dedicated monitor at a central location at the pulpit, which allowed operators to easily view the recommendations. This basic change proved to be transformative, more than doubling the tonnage of steel cast using machine learning in a single month and tripling the amount of realized profits.
3. Prioritize production volume over profits. Although the actual savings potential was 16% of raw material costs, the mill decided to add artificial buffers to the software in order to provide extra-conservative recommendations. While this approach halved the potential savings, it also increased the volume of production using Fero, as plant operations became more comfortable with the software's recommendations. In the 12-month period shown below, the software optimized 25% of all steel produced at the mill. Thousands of heats have been optimized at Gerdau's mill without any failures, and operators have gained invaluable experience in using machine learning to drive production improvements.
4. Increase operational flexibility. Flexibility is crucial when it comes to real-time optimization, and having a well-trained operations team that can adapt to dynamic recommendations is essential. With Fero, the software may recommend reducing alloy additions when it detects that other strengthening elements are sufficient to meet the target mechanical properties, or increasing alloy additions when it predicts a high risk to final quality. As a result, the dynamic targets can fluctuate from heat to heat, requiring the ladle operators to pay close attention to the latest recommendations and optimize the heat's chemistry accordingly. This demands a high degree of attentiveness and agility on the part of the operations team, as well as a willingness to embrace new technologies and approaches to production optimization.
5. Don’t wait for the perfect data. Waiting for the "ideal" data before adopting machine learning is not necessary, as white-box ML algorithms can highlight problems and improve data practices over time. By starting early and continuously refining their approach, companies can achieve the goals they set out to achieve faster. At Gerdau, the use of Fero created a positive feedback loop: the software highlights potential issues, the mills take action to improve the data, and in turn, the models become better.
