• Peter Aurelius Munk
  • Emil Færgemand Bøgh
4. term, Control and Automation, Master (Master Programme)
This thesis investigates the problem of predict-
ing free burn of the short-circuit phase of a con-
ventional GMAW Short-Circuit welding pro-
cess. The problem is formulated as a binary
classification problem where data samples in the
last period of the short-circuit phase should be
classified differently from the prior samples of
the short-circuit phase. In addition to the mea-
surement data provided from the welding pro-
cess, a set of features is extracted from the mea-
surement data to support the classification. A
state observer is designed for the welding pro-
cess in the short-circuit phase, and the residual
for the control voltage is added to the feature
set. Several classification models are compared
in terms of binary classification performance
measures and the K-Nearest Neighbor m pre-
formed the best. In testing, following optimi-
sation of model hyperparameters, performance
is acceptable on test data of the same origin
as the training data, however, the model de-
liver an unacceptable performance on test data
from different variations of the welding process.
For acceptable performance on variations of the
welding process, the classification model must
be trained on the same variations of welding
measurement data.
LanguageEnglish
Publication date3 Jun 2020
Number of pages118
External collaboratorMigatronic A/S
René Petersen RPE@migatronic.dk
Other
ID: 333531499