CEO Berk Birand on how AI is making factories greener:
Watch the interview
Outsourcing to Fero

Model a Process in Hours With Machine Learning

Building a production-ready ML system can take a month or even longer, prolonging the gap between effort put in and value extracted. With Fero, however, companies can do it in hours.

• December 2022
 

Industrial data science has rich potential to turn the stream of data generated on the factory floor into value. However, the average data scientist spends more than a third of their workday on data preparation and cleaning, according to a recent survey by data science platform Anaconda.

Collaboration adds another wrinkle to the process. Typically modeling a process requires several rounds of back-and-forth between data scientists, who can code but may be unfamiliar with the ins and outs of the manufacturing process, and engineers, who have in-depth process knowledge but may not boast much coding acumen.

Building a production-ready ML system can take a month or even longer, prolonging the gap between effort put in and value extracted. With Fero, however, companies can do it in hours.

The project: reducing model building time

One of our customers, a major US-based chemicals manufacturer, wanted to cut the time it took to produce a usable model without damaging predictive accuracy. The team was focused on catalyst lifetimes; their objective was to see if they were operating inefficiently and losing money. Put simply, producing a timely forecast was imperative to business growth.

Fero made the process more efficient. Not only did the software automate the initial data preparation and cleaning process, it also allowed engineers to add their domain knowledge into the model without involving data scientists via a simple no-code interface. Without any coding knowledge, the engineers were able to upload historical data into Fero and add specific process information such as equipment configurations and residence times. After a few clicks through Fero’s interface, they’d generated the forecast they needed.

Before Fero, building a model and interface engineers could use took the team three to five weeks. Using Fero to streamline collaboration, the customer was able to recreate the process in less than two hours, with the same accuracy. This helped them make their internal scientists more efficient and meant they didn’t have to hire any additional consultants. Data scientists, meanwhile, were able to use the software as a rapid prototyping tool. They could easily make a project prototype in hours, allowing them to vet whether a use case had a good value proposition and whether it needed additional custom development.

Fero data science timeline

The value: scaling optimization

Outsourcing work like data cleaning to Fero frees up data scientists to pursue more advanced use cases. To prove this, the customer made a set of conservative assumptions, determining similar projects would last about a month and a data scientist would go through 12 projects. They also assumed that rather than reducing that month to 1.5 hours, use of Fero would reduce a month’s work to about a week, and only half of the 12 projects would be suitable for Fero. Even given this set of conservative assumptions, the customer found they would still end up with five additional months of time that data scientists could dedicate to more advanced use cases, such as complex use cases that can’t be addressed using automated machine learning software.

This solution also allows companies to scale optimization faster. A global manufacturer has tens of thousands of use cases; it’s not feasible for a data scientist to model all of them. With machine learning software, the same models can be deployed across similar use cases, rather than being rebuilt each time from scratch.