Predictive models and machine learning algorithms as a step towards adaptive weld process control A pre-study

Important results from the project

This pre-study has demonstrated the possibility to predict a number of relevant weld quality measures based on models derived from data sampled from controlled experimental designs as well as how to utilize these models to optimize welding system settings to account for incoming part variations. Examples of quality measures are penetration depth, weld toe radii, throat thickness and potential asymmetry effects. The project has also identified potential application areas for machine learning (ML) as well as suitable ML methods that can be further explored in upcoming projects.

Expected long term effects

The results from this project showed great potential in developing regression models based on data extracted from controlled experimental designs. Several of the predicted quality measures showed very good correlation with experiments even though the experimental designs were reduced to increase efficiency. Some quality measures, such as weld toe radii, were however difficult to predict and additional work is required within algorithm development in the data sampling stage to develop more accurate models for these types of quality measures.

Approach and implementation

The nature of the projects research question demanded close collaboration between several disciplines, both within academia and industry. In this project the industry partners were compiled of end users who utilize welding in their production while the academic partners were from completely different backgrounds ranging from computer science to production technology and quality engineering. The project managed to meet all goals and an important lesson for future projects is to involve additional industry partners from robot manufacturers or robot integrators.