Analysis of Seismic Elastic Steel Structures Using a New Hybrid Optimization Deep Learning Model

The traditional seismic elastic analysis model for steel structures suffers from insufficient performance. This study proposes a deep feature learning network model. The model combines a cross-attention mechanism, a graph neural network, and a random sampling consistency algorithm to optimize the point set metric space. Experiments showed that the average F1 score of this model in cloud matching tasks was 0.919. This score was 16.5%-22.6% higher than traditional methods. The matching success rate reached 55.43% and remained more than 15% higher than comparison methods at a 50% occlusion rate. For semantic segmentation tasks, the intersection to union ratio (0.798) and Dice coefficient (0.879) of steel pipe components were higher than those of traditional methods. In actual scenario testing, the model achieved sub-millimeter level deformation detection. The minimum deformation amounts for steel pipes and welded balls were 0.44 mm and 0.31 mm, respectively. The generalization error (0.06 mm) and memory peak value (11.3 GB at 10K points) were reduced by 76.9% and 53.9% respectively, compared to the comparison method. This makes the method suitable for the real-time monitoring needs of edge devices. In summary, the proposed model addresses the shortcomings of traditional methods in dynamic interaction, noise suppression, and deformation sensitivity. It does so through multimodal feature fusion and hybrid optimization strategies. This provides a high-precision and low-energy solution for seismic analysis of complex steel structures.