Laser Seam Welding optimization using Inductive Transfer Learning with Artificial Neural Networks

Abstract

Transfer Learning aims at transferring knowledge from an already learned task to a different, but related task, in order to accelerate the learning process of the latter. This concept can be applied to manufacturing systems where process models that map process parameters into process quality are used to optimize the calibration phase of new unseen products at the shop-floor. However, these process models often require a great amount of experiments, which normally is costly and impractical for most manufacturing systems. The present work explores a Laser Seam Welding scenario with 3 different product variants where the problem is training one of the process models with a reduced amount of labeled data. Artificial Neural Networks (ANNs) were used to model these processes and Inductive Transfer Learning is then used to tackle the proposed problem. Ultimately, this approach was compared to traditional machine learning where no transfer occurs and a model is trained only using the small amount of labeled data. The results revealed that for all the Laser Seam Welding processes the trained models performed better when using Inductive Transfer.