A Kolmogorov-Arnold network-based method for predicting underwater explosion shock spectrum considering cabin structural characteristics

Abstract

This study introduces an innovative approach for the rapid prediction of shock spectrum in ship structures subjected to underwater far-field explosion loads. A nine-compartment ship model is developed, with 31 test condition sets designed. For each set, sample measurement points are strategically placed at typical locations across various deck levels and on the inner and outer bottoms of the double bottom at different frame cross-sections. Shock spectrum (including spectral velocity, acceleration, and displacement) are extracted from these sample points to establish a comprehensive shock spectrum database. Using the Kolmogorov-Arnold Network, a rapid prediction model for ship structure shock spectrum is developed. The network's inputs include ship structural formal parameters (e.g., number and position of decks, transverse and longitudinal bulkheads), impact factors, feature transfer distances, and local shock factors. Spectral velocity, acceleration, and displacement are used as the network's outputs during training. A comparative analysis of prediction accuracy among the Kolmogorov-Arnold Network, Backpropagation Network, and Convolutional Neural Network is conducted. The results demonstrate that the Kolmogorov-Arnold Network exhibits superior predictive accuracy compared to the Backpropagation and Convolutional Neural Networks. In contrast to existing finite element simulation methods, the proposed approach is simpler and more practical. Furthermore, unlike current rapid prediction methods, the proposed method takes into account the influence of ship structural characteristics on underwater explosion shock responses, making it better suited to the requirements for rapid prediction of underwater far-field explosion shock spectrum in ship structures.