An adaptive multitask deep learning model for reliability analysis of vehicle-bridge stochastic vibration systems

Y.S. Zhao, Xuhui He, K. Shi, Borong Peng, Zhengliang Xiang,
An adaptive multitask deep learning model for reliability analysis of vehicle-bridge stochastic vibration systems,
Mechanical Systems and Signal Processing,
Volume 238,
2025,
113190,
ISSN 0888-3270,
https://doi.org/10.1016/j.ymssp.2025.113190.
(https://www.sciencedirect.com/science/article/pii/S088832702500891X)
Abstract: The reliability assessment of vehicle-bridge interaction (VBI) system is crucial for ensuring the operational safety of trains on bridges. While deep learning models have been applied to predict VBI system dynamic responses, most models focus on either vehicle or bridge responses alone. Given the strong coupling and interaction between vehicles and bridges, single-response prediction fails to fully capture the dynamic behavior of bridges under moving vehicles. For this, an adaptive multi-task learning model (AMLM) is proposed, which integrates an optimization algorithm, a shared convolutional neural network (CNN), and two improved long short-term memory networks (LSTM), to simultaneously forecast the responses of both vehicles and bridges. An adaptive loss modulation strategy based on homoscedastic uncertainty is proposed to dynamically balances the loss contributions of different tasks. Furthermore, the probability density evolution method (PDEM) is employed for VBI system reliability analysis. The performance of the AMLM − PDEM framework is compared with the conventional PDEM, and the impacts of vehicle speed and noise levels on prediction accuracy are investigated. The results demonstrate that the devised framework can precisely and effectively predict vehicle and bridge dynamic responses simultaneously. Under the principle of three times standard deviation, the vertical acceleration reliability for the train and bridge are 0.9923 and 0.9977, respectively.
Keywords: Vehicle-bridge interaction; Reliability assessment; Multi-task learning; Deep learning; Probability density evolution method