Overloaded vehicle identification for long-span bridges based on physics-informed multi-task deep learning leveraging influence line

Abstract

With the rapid development of the Structural Health Monitoring (SHM) system of bridges, data-driven methods for overloaded vehicle identification based on deep learning on large-scale monitoring data is playing an increasingly important role to ensure the long-span bridges safety. In recent years, physics-informed deep learning models incorporating domain knowledge have been demonstrated to have better performance, however, the state-of-the-art deep learning models for overloaded vehicle identification (OVI) have not yet well utilized structure knowledge of bridges. Such physics-informed models still remain to be developed and explored. In this paper, a novel multi-task deep learning model IL-MOVI is proposed for overloaded vehicle identification leveraging bridge influence line for identifying overloaded vehicles on long-span bridges. The proposed model IL-MOVI learns to mine the spatial features of the response data collected by the bridge SHM system by leveraging the bridge structure knowledge of the influence line, and maps the response data to the force distribution on the bridge, which significantly improves the spatial feature mining ability of the model. IL-MOVI uses temporal convolutional network to extract the temporal features of the sequence, and design attention mechanism on time scale to pay attention to important moments. In addition, the model is designed in a multi-task architecture to force the spatial features to align with the overall traffic flow state, improving the generalization of the shared spatial features and the identification performance. The experimental results on an OVI dataset, which is established by using the cellular automaton to model traffic flow and applying the modeled traffic flow to the finite element model of a long-span bridge, show that leveraging bridge structure knowledge and the multi-task architecture can effectively improve the capability of the deep learning model on the OVI task. The visualization of network parameters of the spatial feature mining module shows that the network parameters can fit well with the inverse matrix of the influence line, which demonstrates that the proposed method incorporating bridge structural knowledge such as influence line with deep learning model is feasible and interpretable.