Deep Learning Method for Vehicle Load Identification Using Digital Twin and Transfer Learning Strategies

In recent years, machine learning–based vehicle load identification methods have gained significant attention due to their real-time performance and accuracy. The prediction accuracy of machine-learning models is often dependent on the size of the data set, so the scarcity of data in engineering scenarios limits the precision of these methods. To address this issue, this paper proposes a vehicle load identification framework based on digital twin and transfer learning strategies. Considering uncertainties, a digital twin model of the bridge is constructed to generate simulation data sets for various operational conditions. These data sets are then used to train a Bidirectional Long Short-Term Memory vehicle weight prediction model in the digital domain. By fine-tuning the pretrained model with a small set of measured deflection data, an optimized virtual-to-real vehicle weight prediction model is obtained. The transfer learning strategy's effectiveness in small-sample scenarios is validated by creating virtual-to-virtual scenarios for two beam bridges in the digital domain. The method is further validated through an engineering application using monitoring data from a continuous beam bridge over 2 months. The results show that, after introducing the virtual-to-real transfer learning strategy, the average prediction error is 2.33%, with a standard deviation of 1.39% and a maximum error of 4.88%. Compared to the model trained solely on monitoring data, the error average is reduced by 3.94%, the standard deviation is reduced by 1.87%, and the maximum error is reduced by 8.2%. This approach enables high-precision vehicle load prediction in data-limited engineering scenarios and provides a valuable reference for other structural health monitoring tasks where data acquisition is challenging.