Damage Detection and Health Status Assessment of Bridge Structures Based on LTTB-ADSN Algorithm

Abstract In response to the problems of damage detection and health status assessment of bridge structures, an innovative two-stage framework is proposed, which combines time series simplified downsampling and adaptive dynamic sensor network technology. By simplifying the geometric feature selection strategy of the time series module, the dimensionality reduction efficiency has been significantly improved, compressing the 8192 point original sequence to 2048 points while preserving key features. The adaptive dynamic sensor module dynamically generates Shapelet features that conform to physical laws through adversarial training of convolutional generators and discriminators. The experimental results show that on the wooden truss bridge dataset, the model training loss rate reaches 0.52, and the testing loss rate is 0.41. In the dataset of Z24 Bridge in Switzerland, the training loss rate decreases to 0.13 and the testing loss rate is 0.12. In addition, the accuracy of the model proposed by the research in identifying minor injuries is as high as 86.98%, and the accuracy of the test set is 94.68%. The application of the damage sensitivity coefficient demonstrates excellent linear sensitivity and achieves precise localization of the damaged area. These results indicate that the proposed LTTB-ADSN framework has significant potential in improving the accuracy and efficiency of bridge health monitoring, and can effectively address the challenges faced by current bridge monitoring technologies.