The always changing data problem of using AI in manufacturing - Using synthetic data from the digital twin to feed AI models

Abstract

Production is becoming increasingly flexible, which also requires the flexibility of the support system for the production. And the key here is the speed of decisions, in which the support of modern artificial intelligence systems can be crucial. Flexible production is based on a well-planned control of production, which increasingly uses some artificial intelligence component. Artificial intelligence can already be useful in the early stages of planning the production line, and of course it can also control the daily operation of the production line after the installation of the production place or line. The biggest problem is supplying the neural network that controls the artificial intelligence with training data. The production lines typically change every 2-4 months, new products appear, the layout changes, and the main process data also changes due to the development of the processes. This results in the training data becoming outdated or obsolete very quickly and thus cannot be used to train models anymore. High-quality learning data can be produced by digital twin models of production lines. Such synthetic data has several advantages over data collected from production. In this article, we investigate how useful this synthetic data is during the life cycle of the production line.