The prediction of residual stress of welding process based on deep neural network

Abstract

The welding process has been an efficient method for producing essential and complex manufacturing parts in various industrial design fields. The post-weld residual stress can have detrimental effects on welded components. Therefore, systematic studies of residual stress are essential for evaluating welding behaviors and mechanisms in welded structures. They can provide a valuable reference and optimization for addressing residual stress relief. Numerical finite element analyses based on thermal-mechanical models offer a comprehensive approach to simulate real welding, providing a reliable means to determine and quantify the distribution of residual stress based on welding parameters and material properties. Furthermore, the finite element analysis is capable of generating adequate and dependable datasets in relation to the classical experiment. However, the finite element simulation is not considered an efficient method for predicting the magnitude and distortion of residual stress due to its high computational cost. A deep learning framework with powerful automatic learning abilities could potentially be used as an alternative method to efficiently predict residual stress. The purpose of the current study is to propose an innovative modeling approach for accurately and effectively predicting residual stress. A deep network model with Convolutional Neural Network using Adam optimization is integrated with numerical finite element analyses of a single-pass beam weld in SUS304 stainless steel. Finite element analysis is used to generate extensive residual stress datasets, which are partly used to train the deep network model and partly used for model validation. The deep network model aligns closely with the finite element analysis results, with a root-mean-square error (RMSE) of less than 12, an absolute fraction of variation (R2) of greater than 0.95, a mean absolute error (MAE) of less than 6.8 and a mean absolute percentage error (MAPE) of less than 1.1. Furthermore, this study highlights the potential advantage of using a deep network model with strong memory capabilities to directly predict residual stress for identical structural components and welding processes.