Multi-objective global sensitivity analysis of shipborne equipment based on AGPSO-CNN-BiLSTM-attention

To address the high computational cost in global sensitivity analysis of large shipborne equipment, a CNN-BiLSTM-Attention model optimized by a Genetic Algorithm-Particle Swarm Optimization with adaptive weight updating (AGPSO) is proposed. This model is combined with Sobol' global sensitivity analysis for efficient multi-input multi-output (MIMO) sensitivity evaluation of complex structural systems. Firstly, an AGPSO-CNN-BiLSTM-Attention model is developed to automatically optimize three key hyperparameters: learning rate, batch size, and L2 regularization, thereby obtaining the optimal network. Comparative results demonstrate that it achieves training accuracies of 99.49% for relative displacement and 99.12% for absolute acceleration, significantly improving prediction performance and computational efficiency. Secondly, integrated with the Sobol' method, a MIMO sensitivity framework is established to quantify parameter influence. Finally, applied to a gas turbine isolation system, the approach identifies critical isolator and limiter parameters affecting shock response. The method significantly improves efficiency and generalization, offering practical value for structural design and optimization of marine equipment.