Publikation
Synthetic Training Data Generation for Deep Learning-Based Billet Detection in Rolling Mills
Maria Luschkova; Christian Schorr; Tim Dahmen
In: Proceedings of the International Conference on NDE 4.0. International Conference on NDE 4.0 (NDE 4.0), October 24-27, Berlin, Germany, DGzfP, 2022.
Zusammenfassung
AI-powered quality assurance solutions are gaining momentum in the steel
industry under the Industry 4.0 paradigm. In rolling mills, knowing the real-time
location of billets, i.e. fast moving bars of hot steel, is important in order to guarantee
a safe process and defect-free end products. To achieve this aim, we present a deep
learning-based detection of these billets in rolling mills using synthetically generated
training data. A core practical challenge for many deep learning projects is the limited
availability of appropriate, annotated training data. We propose a method for
simulating images employing a partial digital twin of the rolling process. Partial
models governing the shape and location of the billets, the layout of the rolling mill
floor, the camera settings, and the lighting situation changing over time are combined
into a scenario model. Choosing different parametrizations of this scenario model
facilitates synthesizing a broad range of images for training. The resulting deep
learning model is utilised to detect billets in real-world images from an actual rolling
mill. We describe the creation of the partial models using aerial photogrammetry,
expert knowledge, and 3D modelling, as well as the choice of the deep learning model.
An evaluation of the model's performance on real-word images shows the
applicability of our synthetic training data approach.