Large-Span Spatial Structural Damage Recognition Based on a Transfer-Learning Improved LSTM Network

Abstract

The complexity of addressing long-time series data and data noise and outliers, overcoming the difficulties of cross-domain knowledge migration, and adapting to dynamic environmental factors are significant challenges in monitoring and forecasting large spanning structures. In this paper, we propose a real-time damage identification method based on long short-term memory (LSTM) and transfer learning. This method autonomously learns the damage features directly from the original acceleration signals and identifies the location and degree of the damage, which can effectively address these challenges and improve the accuracy and reliability of monitoring and prediction. The method is validated through finite element simulation and the shaker test. The method performs well, and the accuracy of the LSTM model optimized by transfer learning reaches 97.49% and 97.83% in the finite element simulation and test, respectively. A graphical user interface (GUI) is developed for engineering applications. T-distributed stochastic neighbor embedding (t-SNE) downscaling visualization shows that the model classification boundary is clear. This proves that the method can effectively identify damage features, achieve highly accurate real-time monitoring and warning, reduce maintenance costs, and show strong generalization performance in complex environments.