With the increasing requirement in efficiency, quality, individualization and flexibility for manufacturing, the current manufacturing style is transiting from traditional automatic style to smart manufacturing. The core idea of smart manufacturing is to combine the advanced sensing, communication and data processing technologies with the manufacturing processes such that the whole production process can be integrated, monitored, visualized, controlled and optimized through internet of things (IoT), Big data and artificial intelligence (AI). The “digital twin” is developed based on this background for manufacturing monitoring, visualization and control. The digital twin owns the same element and dynamics and is a digital replica of the physical process. In this study, a digital twin of the pulsed-GTAW is built for welding process monitoring, visualization and control. In this digital twin system, the direct information such as arc voltage, welding current, top-side images are sensed by sensors. The undirect information, back-side width is predicted from top-side images by deep learning. The experiments verify the effectiveness of the developed digital twin in welding process monitoring, visualization and control.
An automatic welding platform is built where the welding torch is fixed, and workpieces can be moved in one dimension derived by a step motor. The welding current is set pulsed with base current as 20A and peak current range from 100 A to 190 A. The pulsed frequency is 30 Hz and duty cycle is 50%. One camera is installed on the top-side to sense the weld pool images during base current period and arc images during peak current. To train a deep learning model, another camera is installed on the back side to capture the back-side images and compute the back-side bead width after the calibration experiments. We train three convolutional neural networks (CNNs) with the weld pool images, arc images and combined images as the input respectively to find the best model with back-side bead width prediction performance. Based on this model a digital twin is built with the welding current, arc voltage, welding pool images, arc images and back-side bead width visualized in a digital space developed in Unity, a popular engine for video game development. Furthermore, the online welding experiments are done with different welding current to achieve the target back-side bead width to verify the effectiveness of developed digital twin in monitoring, visualization and control welding process.
Results and Discussion:
The developed CNN model taking the combined images as the input can achieve the best validation performance with the root mean square error (RMSE) of the back-side bead width as 0.19 mm. Testing performance is also rather good with RMSE as 0.22 mm which verifies the developed CNN model. The developed digital twin based deep learning can facilitate welding process monitoring, visualization and control. This adaptive welding system can achieve the target back-side bead width in different welding current.
(1) The back-side bead width can be inferred better by combining the weld pool information and arc information.
(2) The developed CNN model can predict the back-side bead width well with RMSE as 0.22 mm.
(3) The developed digital twin based on deep learning can facilitate welding process monitoring, visualization and control in real time.