Real-time determination of weld penetration status during A-TIG welding of stainless steel employing deep learning approach

Abstract Deep learning approach involving a convolutional neural network (CNN) has been developed to perform semantic segmentation in captured real-time images of the weld pool and determine the weld penetration status in real-time during activated TIG welding (A-TIG). A robotic welding machine with a CMOS camera attached to the welding torch was employed to capture real-time images of the weld pool. Welding experiments were conducted by varying the current from 90 to 300A in steps to achieve various levels of weld penetration depth in the range of 2–10 mm. The above weld penetration range has been categorised into four classes of weld penetration status. A CNN model with VGG 16 architecture has been applied as an encoder in the U-Net framework for weld pool image classification. The accuracy of classification was 99% and the model execution time was 90 ms in a computer for prediction in a single frame of the image. To reduce the model execution time further, a few lightweight architectures were chosen as encoders for the U-Net model and their performance was compared. Among them, the most accurate EfficientNet-B0 was chosen for real-time implementation. The developed model was executed in real-time to predict the weld penetration status in NVIDIA Jetson Nano embedded hardware. The classification accuracy determined for the four classes was found to be in the range of 94 to 98% for the validation experiments. The execution time was found to be reduced to 55 ms for prediction of the weld penetration status in a frame of weld pool image.