|
6869 |
|
Tuesday, March 09, 2021, Hall 1 1:55 PM Education, Diagnostics and Control Technologies |
|
Electron beam probing and deep learning to improve weld quality |
|
Norbert Sieczkiewicz* / Lancaster University/NSIRC, UK Andrew Kennedy / Lancaster University, UK Yingtao Tian/ Lancaster University, UK Colin Ribton/ TWI Ltd, UK Darren Williams / Lancaster University, UK |
|
Electron beam welding (EBW) is used in industries requiring the most exacting standards and quality, e.g. aerospace and nuclear, as it can produce joints with excellent integrity. Consistent welding performance is highly dependent upon reproducing the beam current density distribution. However, this is not just defined by the electrical parameters, but is also dependent on cathode wear in the electron gun and other machine maintenance issues that are not readily measured. Beam characteristics measured by specialised probes have unexploited potential in machine health monitoring. Any variations in machine condition influence the current density distribution of the beam and have an impact on weld quality. This work has exploited the rapid development of computer vision programming to recognise quality indicators in BeamAssure data. The beam probe signal was captured as a digitised electrical pulse. Using time-series imaging encoding methods, the 1D signal was converted to 2D images as a preparation for processing. The experimental results indicate that a convolutional neural network can identify the existing correlation between the beam focus state and the beam probing signal. Using this novel representation of the beam probe signals allows applying an image classification model to recognise the focus state of the beam. |