Establishing inherent deformation database for Q355 steel T-joints in stiffened panels by numerical approach and artificial neural network

The typical welding joints in stiffened panels are generally subjected to different structural restraint conditions, and the welding processes applied to individual joints may also vary to some extent. Consequently, inherent deformations may differ among joints within a given structure, and it is difficult to acquire all the inherent deformation data for every joint. Therefore, the conventional inherent strain method faces a variation issue. This study focuses on the stiffened panel structures and takes low-alloy steel T-joints as the research object. By employing experiments and numerical simulations, the influence of structural restraint on the inherent deformations was investigated. Moreover, artificial neural network was utilized to construct a dynamic inherent deformation database. The results demonstrate that structural restraint reduces the difference in transverse shrinkage between the upper and lower surfaces of the flange plate, and exerts a strong suppressing effect on angular distortion. It also alters the evolution process and spatial distribution of longitudinal inherent strain, but does not change the magnitude of the final Tendon force. Compared with a fully connected neural network, the multi-task learning neural network yields better performance in deducing inherent deformation, achieving Pearson correlation coefficients above 0.97. By combining the network and the inherent strain method, an ANN-ISM approach was proposed to estimate welding deformations of stiffened panels efficiently. This approach reduces the computation time by over 99% while still meeting engineering accuracy requirements, and it was successfully applied to predict the welding deformations of large-scale ship blocks rapidly.