SurFSW – Automated visual inspection of friction stir welds for large-scale space structures

The goal of the research project SurFSW is an independent and reproducible inspection of friction stir weld seam surfaces through automated acquisition, processing and evaluation using artificial intelligence. This allows a reliable surface inspection and ensures the production of components with high seam qualities.

Motivation

Friction stir welding (FSW) is a solid-state welding process, which is applied in the aerospace sector. A high joint quality is important, which is why the welds are inspected after the welding process. The first step is visual testing during which the seam surface is examined with regard to surface defects and other characteristic features. The high cost for specially trained personnel as well as the subjective evaluation are still unsolved problems with the visual inspection performed by humans. This results in an uncertain assessment. In addition, the documentation of the surface features is currently limited with human visual inspection due to the necessary manual data acquisition.

Research objective

The overall technical objective of the research project SurFSW is an automated visual inspection system, which is integrated into a welding machine. This allows a detection, localization, quantification and documentation of various seam surface features and makes it possible to identify irregularities in the welding process and to eliminate their causes.

Approach

The first step in the project is to design and implement a system for automatic digitalization of the weld seam surface. For this, a measurement setup, consisting of an industrial camera and a suitable illumination system, will be integrated into the welding machine at the iwb. A suitable image processing pipeline is then identified and set up with which relevant seam surface features can be extracted. The features are afterwards detected using artificial neural networks (ANN). The necessary image data set will be generated from welding samples. To increase the number of training data, an algorithm will be developed to generate artificial and labelled images of welds. The labelled training data will then be used to further improve the ANN. The algorithm determines weld surface features based on camera images. This includes feature recognition, localization on the workpiece, and quantification of the features using image segmentation. To account for a real production environment, a variety of disturbances (e.g., reflections) are considered during testing. The models and the camera system are tested in a real production environment.

Results

An automated visual inspection allows for a person-independent assertion of the weld quality and documentation of the surface features. By classifying the seams based on probability, the usability of the result can be quantitatively assessed.
In addition, the project will gain new insights regarding the optical properties of friction stir welds, how they can be highlighted by different image processing methods, and how an image-based system can be designed on the hardware and software side for robust detection of the features.

Acknowledgements

The presented project is funded by the German Federal Ministry of Economic Affairs and Climate Action (BMWK) under the funding number 50RL2240 and is supervised by the German Aerospace Center (DLR). We thank the BMWK as well as the DLR for their support and for the trusting cooperation.