Stress time series prediction for orthotropic-steel-bridge-decks welds using a TCN-LSTM neural network

The stress time-history at the rib-to-deck (RTD) connections of orthotropic steel bridge decks (OSBDs) is crucial for the near-real-time safety assessment and early warning of highway bridge structures. To improve the accuracy and efficiency of stress prediction, this study develops a hybrid Temporal Convolutional Network – Long Short-Term Memory (TCN-LSTM) model. The training dataset is constructed primarily from finite element simulations of stress responses, supplemented by real-world vehicle and temperature data collected via a Weigh-in-Motion (WIM) system and temperature sensors within the structural health monitoring framework of the Nanxi Yangtze River Bridge. Utilizing vehicle volume, speed, axle weight, and ambient temperature data collected and statistically processed by the WIM system within the structural health monitoring framework of the Nanxi Yangtze River Bridge, combined with finite element model simulations, a time-series training dataset was constructed. The TCN-LSTM hybrid architecture captures both local features and long-term dependencies within the time-series data. The model's performance was rigorously compared against five established machine learning models Backpropagation Neural Network (BP), Long Short-Term Memory (LSTM), Convolutional Neural Network (CNN), Radial Basis Function Neural Network (RBF), and Random Forest (RF) using four key statistical evaluation metrics: Mean Absolute Error (MAE), Mean Squared Error (MSE), Root Mean Squared Error (RMSE), and the Coefficient of Determination (R²). Results demonstrate that the TCN-LSTM model achieves the most accurate and stable predictions for stress time-history. Compared to the best-performing traditional neural network model, it significantly reduces MSE, RMSE, and MAE by 62.9%, 38.4%, and 62.1%, respectively, while improving R² by 6.4%. The stress time-series data used for model training and validation are predominantly generated through finite element simulations, incorporating realistic vehicle loading and temperature conditions derived from field measurements. The proposed TCN-LSTM model provides an efficient tool for stress prediction in OSBDs, significantly reducing computational and experimental costs. • First hybrid TCN-LSTM model for stress time-history prediction in orthotropic steel bridge decks. • Validated with field monitoring data (187k+ vehicles) and multi-scale FEM simulations • Enables 22% resource savings in bridge fatigue tests versus conventional methods.