GMAW Molten Pool Micrograph Image Recognition Based on Convolution Neural Network and Transfer Learning

  • Yanfeng Li
Keywords: Molten Pool, Micrograph Image Recognition, Convolution Neural Network, Transfer Learning

Abstract

The weld pool micrograph image carries a wealth of welding quality information, but it is very difficult to
recognize the welding quality directly with the weld pool micrograph image as the input. This paper, based on
convolution neural network and transfer learning, designs an end-to-end model to recognize the state of GMAW
molten pool. Firstly, a passive vision micrograph image acquisition system is designed experimentally with the
analysis of spectrum and the image data set is established so as to avoid the strong welding arc interference
involved in shooting molten pool image. Then, based on convolution neural network, a model of molten pool state
recognition is constructed with the number of filters optimized. The accuracy of the testing experiments is over
92.8%. Moreover, since the recognition accuracy is lower than the average value, we use the method of
instance-based transfer learning and parameter transfer learning by inputting the region of interesting(ROI) images
of welding pool micrograph images to further improve the recognition accuracy. Finally, the feature maps of the
model are analyzed to illustrate that this model can extract more abundant feature information of molten pool than
the shape and geometry size, and is more robust. The experiments show that this model can effectively identify the
welding pool state, and the approach of targeted transfer learning can further improve the recognition accuracy to
96.15%.

Published
2020-05-01