Load identification and strain reconstruction of stiffened cylindrical shells subjected to underwater explosion based on deep neural network

This paper presents a deep neural network (DNN) framework for identifying an underwater explosion (UNDEX) load and reconstructing a strain field on a stiffened cylindrical shell using permanent deformation and maximum acceleration measurements at a few locations. Transfer learning is incorporated to improve the performance of strain reconstruction with limited training data. The training dataset is generated through acoustic-structural coupled finite element simulations performed using ABAQUS. Hyperparameters of the DNN, including the number of hidden neurons and dropout rate, are optimized via a grid search strategy to balance model efficiency and prediction accuracy. The analyzed results demonstrate that the proposed DNN model enhanced by transfer learning can accurately infer both the detonation location and charge mass while achieving high-fidelity strain reconstruction, particularly in regions undergoing large plastic deformations. It aims to develop a practical solution for rapidly obtaining the post-event structural information from stiffened cylindrical shells subjected to far-field UNDEX.