With a reliable and accurate method
for real-time feedback being a necessity
for robotic welding, various methods
have over time been proposed for
online weld monitoring.
This paper investigates the use of Convolutional
Neural Networks (CNNs) to
analyze the sound signature from the
welding arc, which contains information
about the penetration state of the
weld pool. The sound is transformed
into a spectrogram image using a series
of STFT transforms and the image
is input into a CNN which classifies
it as one of three possible penetration
states. Given a pre-recorded dataset,
the authors investigate how different
sampling strategies impact the ability
of the network to generalize to unseen
examples. Further it is investigated,
how various pre-processing parameters
used when creating the spectrogram
impact the prediction accuracy
of the network. Using the optimal
sampling strategy and pre processing
parameters a custom CNN is
built and tuned, achieving an average
testing accuracy of 69,94% in recognizing
the penetration state based on a
sound sample of 0,25s.