Abstract Scope |
The monitor and control of weld pool penetration state in gas tungsten arc welding (GTAW) are crucial for assuring the high-quality of the weld seam. However, the heat input of adjacent weld pool, variation of cooling conditions, and adjustment of welding parameters can influence the backside width of weld pool, so keeping the uniformity of backside width is a critical/key factor in the high-quality control. In the traditional welding process, the experienced welder observes the topside surface information of weld pool to justify the backside width. In order to realize the automation of welding process, detecting the weld backside width precisely by measuring the topside surface information of weld pool like a welder’s observation is a challenge. Hence, a simplified passive vision sensing system is developed with the intention that effective, enough, and clear information can be provided to predict the backside width. The experiments that are changing the weld current in a certain range to satisfy the full penetration are conducted to simulate the welding process accompanied by the variation of heat input. For the sensing system, 2-D information (topside images of weld pool) and 1-D information (current ,voltage and time) are used as the features, and the backside images that are calculated using the fitting curve are used as the labels for the deep learning model. The designed fusion model composed of convolutional neural network (CNN) and multilayer perceptron (MLP) is applied to the training, validation, and test dataset. The result of the test dataset shows that the error of fusion model is 0.525mm. The errors of CNN model and MLP model are 0.546mm and 0.553mm, respectively. Hence, the fusion model can effectively combine the characteristics of different information to produce optimal results.
Keywords:Intelligent Welding, Passive Vision, Data Fusion, Penetration Prediction, Deep Learning |