A real-time prediction method for weld porosity of 5A06 aluminum alloy based on arc spectra using elemental attention, multi-scale convolution and LSTM during pulsed GTAW process

The pulsed gas tungsten arc welding (GTAW) is widely applied in aluminum alloys for its stability and good weld formation. Porosity is a common internal defect during the process and is difficult to detect in real time. Aiming at real-time monitoring of internal porosity defects in aluminum alloy during GTAW process, this research used arc spectra as non-contact sensing technology and proposed the elemental attention and multi-scale convolution based long short-term memory network (EAMC-LSTM), which is a real-time prediction method for porosity defects based on attention mechanism, multi-scale convolutional neural network (MCNN), and long short-term memory (LSTM). The final model contains a new feature extraction method and three model branches that are sensitive to different porosity formation conditions. The real-time performance was verified. Through the test experiment under normal welding conditions, the experimental results reached an accuracy of 96.68%, a recall rate of 83.34%, a precision rate of 81.82%, and an F1 score of 82.56%. It is expected to be applied to the monitoring process of intelligent welding system in the future.