Some steel waste occurs only after it has left the mill. When components for excavators, bulldozers and other machines are cut from large steel plates, those shapes can either mesh tightly together or leave substantial areas of unusable offcut.
Finding a better way to use every last square inch of steel with minimal wastage is an engineering problem that Juxihong Julaiti, principal data scientist and founding director of Optima Lab at China Unicom, is trying to tackle. Around 20 to 30% of steel is typically lost as scrap or offcuts during plate cutting, he says, with manufacturers “silently wasting steel”.
China Unicom’s solution is an AI system that predicts how efficiently the parts required for a proposed production schedule will fit on to steel plates, allowing planners to fulfil orders in combinations that use more material from the raw plate.
Doing so is tricky. Factories normally schedule around factors such as plant capacity, order priority and delivery dates. Working out the best arrangement for a large number of irregular shapes is a challenging problem, Julaiti says, because every additional component creates a rapidly expanding number of possible layouts.
Doing this manually was possible, Julaiti says, at a small factory producing about 10 machines a day. There, they could calculate each plate layout from scratch by searching five to 10 years of historical production data for similar combinations of parts, then using their recorded material utilisation to create new schedules.
That approach is impractical at larger plants producing 500 machines or more a day, however. Instead, China Unicom trained a deep learning model to learn the relationship between a batch of components and the plate utilisation it was possible to achieve.
The AI system is relatively simple, says Julaiti. When the scheduling system considers a combination it has seen before, it retrieves the previous result. When it encounters a new one, the model predicts its likely utilisation. Once the schedule has been carried out, the actual result is fed back into the system to retrain it. “The beauty in this is we don’t replace what’s there,” Julaiti says. “We give them a new dimension.”
Preserving previous good work is an important element for potential users of the new system. Julaiti says planners and factory floor staff were initially concerned that the project would disrupt their work or force them to adapt to unfamiliar schedules. To allay those fears, the team tested the AI system in smaller experiments, deliberately selecting batches the model predicted would produce either strong or poor utilisation. Its predictions matched the outcomes seen by staff laying out the parts, giving staff the confidence to adopt it. The system was then incorporated into scheduling and scaled across several plants over the following six months.
Rolling out the AI system has worked well, according to Julaiti. Official notes from a presentation he delivered at the UN AI for Good Global Summit in early July suggest plate utilisation increased from 80% to 82% under identical demand, while more than 90% of schedules were generated automatically and manual planning effort fell by 60%. It also cut costs: more than RMB 10m, or just over £1m, per year.
Julaiti admits that a two percentage point improvement in plate utilisation might appear small, but at a factory consuming large quantities of steel plate it means building the same number of machines from less raw material, cutting purchasing costs and embedded carbon in scrap.
The success in China suggests an unflashy yet potentially significant role for industrial AI. Rather than automating an entire factory or replacing experienced engineers, the system can highlight the problems with a planning decision, then enable existing tools and human planners to fix it.
While Julaiti admits the new system will not eliminate cutting waste, and its accuracy depends on the quality and relevance of historical data, it offers immediate savings without replacing machinery or redesigning products. It is “a working system helping factories build the same machines from less steel and less carbon for the planet,” says Julaiti.
Want to learn more about AI in engineering? Gain an understanding of AI principles and learn how to apply the technology to solve complex problems at IMechE’s AI for Engineers training course, next running in Manchester from 1-2 October. Learn more, check for further dates and register now.
Want the best engineering stories delivered straight to your inbox? The Professional Engineering newsletter gives you vital updates on the most cutting-edge engineering and exciting new job opportunities. To sign up, click here.
Content published by Professional Engineering does not necessarily represent the views of the Institution of Mechanical Engineers.