Machine Learning-Based Predictive Quality

摘要

In the metals industry, up to 4% of the production needs to be scrapped even with the highest standards related to quality control. Most defects are related to the early stages of the production process (casting and milling) but often only discovered at the finishing stages of the product. Most defects and scrap could be avoided if discovered early in the process by: · e-assigning the material to a less demanding order · Selecting different production route · Adapting some parameters of remaining production steps · Adding a rework step to correct the defect But the quantity of data to be analyzed makes it impossible for human operators and quality experts to predict these defects. Machine learning based predictive quality can predict up to 75% of the defects, most of which can then be avoided. This will reduce the scrap production of the plant drastically. The big challenge is how to combine mathematics and production, in other words how to use a machine learning model to fit to the operational world. This includes correct choice of problem statement, prediction target and definition of the business use case. One way to integrate a machine learning-based quality prediction in the existing process is by applying a certain set of actions for all coils considered "suspicious". For that, we need to define a threshold for the probability of defect above which we will act on the coils to avoid the defect. This threshold can be calculated based on the cost of corrective action and the cost of defect. It should also take into account some business and technical limitations based on the chosen actions.