Abstract Scope |
Welding is a basic technology in manufacturing industry, and Gas Tungsten Arc Welding (GTAW) is one of the most widely used welding processes at present. The key and main difficulty to realize the automation and intelligentization of GTAW is to sense the dynamic weld pool in real time, so as to provide sufficient information support for automatic control. In this research, the dot-matrix structured light based 3D vision method is applied to sense the 3D weld pool surface in stationary GTAW. A novel method of reconstructing the 3D weld pool surface from the 2D reflection image is proposed using the Long Short-Term Memory (LSTM) neural network in deep learning.
A simulation model of the dot-matrix structured light sensing system is established in advance to provide samples for training the reflection point mapping calculation model. Based on the LSTM neural network, the reflection point mapping calculation model is established, and the mapping calculation from the 2D imaging points on the imaging plane to the 3D reflection points on the weld pool surface is achieved. A pre-sorting step of the simulation sample points in training process is designed, which makes the reconstruction method skip the point recognition stage completely and greatly simplifies the calculation process. In view of the boundary error of the interpolation surface, two boundary extension methods are proposed. The GTAW test results show that after boundary optimization, the weld width error in the X direction is 2.43%, and the weld width error in the Y direction is 8.68%.
The output characteristics and robustness of the reflection point mapping calculation model are analyzed, and its ability to deal with the phenomena of "missing scattered points", "missing concentrated points" and "more points randomly" is studied. The results show that the established reflection point mapping calculation model has good prediction accuracy and robustness overall, but shows low prediction accuracy for initial output reflection points and poor robustness for initial input imaging points (on the upper left side of the reflection image). In view of this deficiency, a forward-reverse united reconstruction optimization method is proposed, which avoids the inaccurate initial output of the LSTM neural network, and greatly improves the robustness of the reflection point mapping calculation model. After optimization, the weld width error in the X direction is reduced to 1.62%, and the weld width error in the Y direction is reduced to 3.94%. |