Reinforced concrete beam full response prediction with hybrid feature-orientation transformer-LSTM model

Abstract

Accurate and efficient prediction of load-displacement behavior in reinforced concrete (RC) beams is essential for data-driven interventions and predictive maintenance in structural engineering. Such predictions are critical for ensuring the safety, reliability, and longevity of structures. Develop a comprehensive approach that incorporates advanced deep learning (DL) techniques, leveraging historical data and time series modeling to achieve high-precision predictions. This study introduces a hybrid feature-orientation Transformer-LSTM (HFT-LSTM) model for accurately predicting the load-displacement response of RC beams, from initial loading to failure. It leverages a Transformer architecture, embedded with a multi-headed attention mechanism, to achieve a deeper representation of constant features and dynamic fusion of variable features, enabling a more comprehensive understanding of the RC beam's behavior. It employs a bidirectional LSTM to capture dynamic and temporal relationships within the displacement series. To gain a deeper understanding of the history-dependent mechanical behavior of RC beams, we conducted a detailed analysis of the influence of input features on the load-displacement curve. The proposed model significantly outperforms existing DL models, achieving a 46 % reduction in Mean Absolute Error (MAE), a 58 % decrease in Root Mean Squared Error (RMSE), a 5 % increase in R-squared (R²) and Explained Variance (EV), and a 36 % reduction in Median Absolute Error (MedAE).