Lantek – Forescrap: Prediction Service for Scrap Generated in Sheet Metal Cutting Processes
Lantek
Sector: Industry
Business Case
In metal part manufacturing processes involving cutting, the generation of scrap or unused material is closely tied to how efficiently the parts are arranged on the sheet metal. In order to address this challenge, the Forescrap system has been implemented, which aims to predict the amount of scrap generated taking into account the characteristics of the part and the manufacturing history.
Objectives
Predict the amount of scrap to be generated in the manufacturing process of a metal part prior to production.
Use case
A multi-client prediction model was developed using regression with an ensemble of models. It was then packaged into a Docker image and deployed to a Kubernetes service for deployment in a scalable production environment.
Infrastructure
Cloud
Technology
Machine learning and deep learning
Data
Database tables, structured data, etc. Private datasets
Resources
Personnel: 1 data scientist, 1 MLOps engineer, 1 project manager Infrastructure: AKS scoring, AKS training, blobstorage, AzureML workspace, Vnets
Difficulties and learning
It was necessary to define a MLOps (machine learning operations) process for the standardisation of all Lantek ML processes, from the metrics definition process to the deployment process in the production environment.
KPIs (business impact and metrics of the model)
Error metrics related to scrap generation in production and the capacity to predict at a rate of 2300 bids per minute are enhanced through vertical and horizontal scaling in the AKS scoring system.
Funding
Hazitek
Collaborators, Partners
Internal development with own resources.
