This report covers a study of the acoustic
response of GMAW's capabilities within machine
learning-based weld quality monitoring.
Initially it is determined to train an artificial
neural network, ANN, to classify two metal
transfer modes - globular transfer and shortcircuit
transfer - and three penetration states
- lack of penetration, full penetration and excessive
penetration - based on related work.
To do so, 1166 features are extracted for each
window of acoustic signal consisting of a range
of temporal-, spectral shape-, harmonic- and
perceptual features as well as statistical features
from a wavelet packet decomposition.
Classification data is produced in a robotic
GMAW cell by provoking the desired classi-
cation states. The acquired data is then preprocessed
and input to a function made to
train 110 ANN congurations for 15 combinations
of window size and overlap using both
gradient descent with adaptive learning rate,
GDA, and scaled conjugate gradient, SCG,
descent.
Based on the trained ANNs it was concluded
that classification of the three penetration
states was possible for ANNs trained using
SCG and partially possible if they are
trained using GDA. Furthermore, the results
for whether classification of metal transfer
mode is possible were inconclusive but showed
a tendency of correct prediction.