A deep learning-based method for identifying the dynamic stability of single-layer spherical reticulated shells

• A dataset was established for the stability and instability models of K6 and K8 single-layer spherical reticulated shells with different spans and rise-to-span ratios under the influence of three different seismic waves. • Using deep learning to identify the dynamic stability of single-layer spherical reticulated shells. • Using TSN as a baseline model, a TSN-GC deep learning model was proposed, which outperformed other common video understanding models across various evaluation metrics. Identifying the dynamic stability of single-layer spherical reticulated shells has been a persistent challenge in structural engineering. Leveraging the advantages of deep learning in computer vision, a deep learning-based method for dynamic stability identification was proposed. Kiewitt sunflower-type single-layer spherical reticulated shells were selected for study, and dynamic analysis was conducted under three different seismic waves. The Budiansky-Roth (B-R) criterion was used to determine the dynamic stability of the structure, and displacement cloud videos were generated to visualize stable and unstable states. A TSN-GC model was then developed by integrating a Global Context (GC) attention mechanism into the Temporal Segment Networks (TSN) model to address issues such as spatial and temporal dependencies, long-term dynamic change capture, and information loss. The model was evaluated using multiple metrics, including Accuracy, F1-score, Precision, Recall, and Specificity, on a dataset for identifying stability and instability in reticulated shell structures. Experimental results show that the TSN-GC model achieved an accuracy of 87.50 %, improving by 3.47 % over the original model. It outperformed TSM, TIN, I3D, Slowfast, and C2D by 2.08 %, 4.86 %, 4.25 %, 6.94 %, and 3.47 %, respectively, confirming its effectiveness and superiority in dynamic stability identification.