Project: Quality assessment of fusion welded joints

Problem definition

Shielded metal arc welding (MSG welding) is frequently used in industrial manufacturing to join steel materials due to high deposition rates. The quality of the component joint has a massive influence on the validity of the dimensioning of the welded structure. Proof of the quality of the welded joint or the welding process is therefore essential and is required by the various regulations. Depending on the application, testing is usually performed both non-destructively and destructively.  Since these test methods have not yet been subject to intensive consideration of the time-dependent measured welding parameters, the potential for automated quality control of welded joints has not yet been exhausted. However, with the aid of artificial intelligence methods, it is possible to view the quality of the joint from a broader perspective on the basis of the time-dependent progression of the welding parameters and thus to assess it holistically.
The aim is to show how the application of machine learning, based on process signals in combination with analyses of metallographic examinations, can be used to set up prediction models for the quality assessment of joints in welded structures.

Objective

The quality of a welded joint cannot be entirely verified by subsequent quality control processes. In standardization, these processes are referred to as "special" processes and this therefore requires consistent quality assurance throughout the entire manufacturing process. For this reason, not only the welding processes themselves but also the groups of people involved (welders, operators) must be qualified.    
Depending on the area of application and design quality, different evaluation groups are consequently used, for example by applying DIN EN ISO 5817. Pre-dimensioning and verification calculations are then carried out on the assumption of a weld quality class and the sometimes greatly reduced geometry parameters.
In the case of dynamically stressed components, notches of different types and characteristics are decisive for the actual service life of the welded joint in addition to the classical geometry parameters of the weld. These are taken into account, for example, in the IIW calculation recommendations by reducing the tolerable stress for the respective shape of the welded joint / notch detail. Internal details of the weld, for example discontinuities in the hardening process, also influence the strength and could be additionally taken into account if they were sufficiently dimensionable and controllable.
Thus, if the position of the weld seam, the penetration conditions or the position of the heat-affected zone could be specifically adjusted, the utilization of the deposited weld metal could be further optimized.

Result

The number of input and output parameters has a major influence on the feasibility assessment of a project; here, the quality of the results usually only becomes apparent after several analyses.  Therefore, downstream validation tests are to be carried out at TIME on the basis of different trained algorithms / networks in order to check the result quality of the forecast parameters.
After successful validation, the prediction models will be transferred to a real-time capable user system on an automated welding system. With the help of a prediction model for welding parameters under specification of a weld geometry by the design department, the parameter selection shall be simplified for the user in the setup process. After the teach-in process, the system can thus preselect the welding parameters itself on the basis of previously prepared ML approaches. If this system proves to be practical, the "initial parameterization" for welded joints can be considerably reduced in effort with regard to the previously existing iteration loops. The resulting factor of resource conservation combined with economic advantages can lead overall to a new efficiency for the start-up of welding processes in the production environment.