To predict the overall deformation and damage characteristics of ring-stiffened cylindrical shells subjected to underwater explosion loading， this study proposes a point-cloud displacement field prediction and deformation reconstruction method integrating a PointNet-based conditional diffusion model， K-nearest neighbor algorithm， graph neural network residual correction， and spatial interpolation. The deformation response data of the cylindrical shell obtained from numerical simulations are used for model training， and a deformation response prediction model for cylindrical shells under underwater explosion loading is established. The proposed model enables the prediction of three-dimensional deformation displacements and the reconstruction of complete surface deformation profiles under different charge masses， standoff distances， and time conditions. Error evaluation results show that the mean squared error， root mean squared error， mean absolute error， and coefficient of determination on the validation set are 0.0077 mm²， 0.0877 mm， 0.0548 mm， and 0.9858， respectively， indicating high displacement prediction accuracy. The reconstruction results can effectively reflect the deformation history and final overall deformation morphology of the cylindrical shell.