Accuracy Control and Welding Distortion Prediction in a Deck Plate

Abstract

Many industrial processes make use of welding to assemble structural parts. While this is standard procedure, the high temperatures encountered during the welding process can generate distortions in the base metal. These distortions negatively impact the parts by generating reworks in order to overcome them. This detrimental effect cannot be avoided, but it can be controlled. One way of performing such control is to add an allowance, which is an accepted amount of distortion. However, a prediction tool is elementary to determine the necessary tolerances. Nowadays, the prediction methods can be grouped in three main approaches: experimental, computational (finite-element method) and machine learning. While the first two methods have been well studied, the machine learning approach is not as well understood. The aim of this study is to explore machine learning algorithms such as neural networks and polynomial regressions in order to come out with a prediction model. Along with this, a best fitting study took place so that a simple formula for predicting metal distortion could be outlined. In order to create some prediction models, the Knime software was used and main design parameters were gathered. The model’s workflow has been organized so that several cases could be tested. By using the workflow, results for eight variables were obtained. Nonetheless, the results were not satisfactory due to limitation of given data. Hence, the problem has its variables reduced to two, which increases the accuracy of the model. Finally, it was possible to generate some models for the welding distortion prediction and it has been proven that these methods can be applied to such problems. Yet, they were not asaccurate as the best fitting method, which means that more data is required so that the accuracy can be improved.