Hybrid-data-driven bridge weigh-in-motion technology using a two-level sequential artificial neural network

For existing bridge weigh-in-motion technologies, the main challenge in accurate weight estimation is to overcome the difficulty of identifying the closely spaced axles. To do so, many field test data are generally required for each bridge in application. To address such a challenge, a novel two-level sequential artificial neural network (ANN) model trained by the hybrid simulated-experimental data was proposed in this study to identify the gross weight and axle weight. For this, simulations and scaled experiments were conducted for the vehicle–bridge interaction system to develop the sequential ANN model. The sequential ANN model was realized by a special data looping strategy, in which the outputs of the global-level ANN served as partial inputs to the following local-level ANN to predict the axle weight. The optimized size of the training data and the appropriate hybrid ratio of the sequential ANN model were also explored. Finally, the proposed algorithm was applied to a real bridge application via transfer learning, as the optimized hybrid sequential ANN model serves as the pre-trained model. The results showed that for the small training datasets with only 5% experimental data, the proposed algorithm significantly improved the accuracy in weight estimation of moving vehicles with closely spaced axles. The field test demonstrated that the proposed algorithm also applies to different bridges within a gross weight identification error of 5%, showing the promise of the proposed algorithm in practical applications.