A Transformer-Based Bridge Structural Response Prediction Framework

Abstract

Structural response prediction with desirable accuracy is considerably essential for the health monitoring of bridges. However, it appears to be difficult in accurately extracting structural response features on account of complex on-site environment and noise disturbance, resulting in poor prediction accuracy of the response values. To address this issue, a Transformer-based bridge structural response prediction framework was proposed in this paper. The framework contains multi-layer encoder modules and attention modules that can precisely capture the history-dependent features in time-series data. The effectiveness of the proposed method was validated with the use of six-month strain response data of a concrete bridge, and the results are also compared with those of the most commonly used Long Short-Term Memory (LSTM)-based structural response prediction framework. The analysis indicated that the proposed method was effective in predicting structural response, with the prediction error less than 50% of the LSTM-based framework. The proposed method can be applied in damage diagnosis and disaster warning of bridges. The core code has been open-sourced at:

Deep Learning for Structural Health Monitoring: Data, Algorithms, Applications, Challenges, and Trends

“…Finally, using graph sensor data, the GNN was trained to successfully detect cable damage on cable-stayed bridges. Li et al [197] proposed a transformer-based time series prediction framework, verified by bridge strain data, and was found to have a more minor error than LSTM.…”