Deep neural network architecture for box girder bridge weigh-in-motion system

Bridge Weigh-in-Motion (BWIM) systems estimate vehicle load based on the bridge’s structural response. This technology is crucial for road infrastructure protection. Recently, neural network-based approaches have been increasingly applied to such systems. While most previous studies have focused on improving the model’s accuracy and reliability, the required size of the training dataset has often remained unaddressed. The novelty of this work lies in the development of a neural network architecture for BWIM systems that provides high accuracy with a reduced training sample size. The proposed architecture integrates Recurrent Neural Networks (RNNs), Time2Vec embeddings, and fully connected layers, utilizing vehicle velocity as an input parameter. A box-girder bridge was modeled using the Finite Element Method (FEM), and its mechanical response caused by three types of trucks was determined. The resulting data were used to train the neural network. The model achieved a Mean Absolute Percentage Error (MAPE) ranging from 0.85% to 3.49%, depending on the scenario, with a training sample size of approximately 1800 elements per truck type. The results demonstrate that the proposed architecture provides competitive accuracy while requiring a smaller training dataset than existing solutions for box-girder bridges. This makes the model more suitable for practical implementation in situations when data collection is limited or involves significant financial costs.