Research on the construction method of dynamic agent model for typical hull structural strength response based on adaptive collaborative point-adding strategy

High-precision structural strength response agent model can significantly improve the efficiency of multi-parameter, high-dimensional, and high-nonlinear solver, such as ship structural reliability analysis and optimization design. However, existing construction methods and point-adding strategies exhibit a deficiency in balancing accuracy and efficiency when addressing the demand for multi-output agent model to solve problems characterized by multiple working conditions, components, and strength response outputs. This study proposes a construction method of dynamic agent model for structural strength response based on Adaptive Collaborative Point-Adding (ACPA) strategy, combined with GA-PSO-BP neural network. A hybrid collaborative point-adding criterion is built by introducing a Key-domains Overlapping Penalty (KOP) term into the Expected Improvement (EI) criterion. Furthermore, combined with SMOTE oversampling algorithm and K-means clustering algorithm, the strategy dynamically adds points and iteratively updates to construct the optimal multi-output agent model by defining adaptive coefficients and sampling stopping criterion. Taking a typical hull stiffened plate structure and a high-dimensional complex ship hold structure as the object of study respectively, the dynamic multi-output agent models were constructed and analyzed. The established method can obtain high-precision agent models that meets the requirements with as few samples as possible, which provides a technical solution for further efficient safety and reliability assessment and optimization design of high-dimensional ship structures.