This paper illustrates how a challenging problem be solved by assuring the adequacy of the raw information and using an appropriate deep learning network to extract the relevant information. The problem concerned is accurate monitoring of the weld joint penetration, in a fully penetrated weld pool, as measured by the back-side bead width. The challenge lies in that the penetration occurs underneath the workpiece and is not visible. A popular method is to use a weld pool image to derive it. Analysis of the physical process suggests that a single weld pool does not contain adequate information but most recent serial weld pools may. As such, although a deep learning model may extract information that is already there, the raw information may not be sufficient. Hence, a model that is capable of extracting information from dynamic serial weld pool images is needed. To this end, a CNN-LSTM (convolutional neural network combined with long-short term memory one) model is proposed. Dynamic weld pools are experimentally generated using randomly changing welding current and speed. The weld pools are imaged using an HDR camera during experiments. Images are also captured from the back-side surface of the workpiece to provide the ground truth for training, validation, and testing. It is found that the highly dynamically changing weld pool can be accurately predicted using serial weld pool images at 0.3 mm for its back-side bead width. Comparison has been made with results from comparative studies to verify the effectiveness of and the contribution from the information adequacy (by using serial images) and the feature extracting capability (by using deep learning).
Keywords: weld, weld penetration, deep earning, CNN, LSTM, CNN-LSTM, HDR image