TM-EDNet: cable force reconstruction method for cable-stayed bridges based on encode-decode structure

Abstract Cable-stayed bridges play a vital role in modern transportation systems, and the accurate evaluation of cable forces is essential for ensuring their structural safety. Traditional estimation methods often struggle under strong temperature effects and complex thermo–mechanical coupling. To address this problem, a Thermo–Mechanical Encoder–Decoder Network (TM-EDNet) is proposed for high-precision cable force reconstruction. The model establishes an end-to-end mapping from temperature and mechanical signals to cable force responses. The encoder integrates a Recurrent Neural Network (RNN) and a Temporal Convolutional Network (TCN). The RNN captures long-term temporal dependencies, while the TCN extracts multi-scale local dynamics through dilated convolutions. The decoder combines a Long Short-Term Memory (LSTM) network and a multi-head attention mechanism. The LSTM maintains long-term memory, and the attention module adaptively emphasizes key time steps related to abrupt structural responses. This hybrid architecture enables TM-EDNet to represent both slow thermal trends and rapid load fluctuations with high fidelity. The model was trained and validated using one year of monitoring data from the Shengli Yellow River Bridge in Shandong, China. TM-EDNet achieved the lowest Mean Squared Error(MSE), Mean Absolute Error(MAE), and Root Mean Squared Error(RMSE) among seven benchmark models, with R 2 ≥ 0.98 at all measurement points. Ablation studies further confirmed that each module contributes to the overall performance. TM-EDNet provides an accurate, stable, and physically interpretable framework for cable force identification under complex thermo–mechanical conditions.