A deep learning method for heavy vehicle load identification using structural dynamic response

Abstract

The identification of dynamic vehicle loads is crucial for bridge health monitoring. Currently, bridge weigh-in-motion (BWIM) system and image recognition methods are extensively used for vehicle load identification. However, BWIM systems are costly, while single image recognition methods struggle to accurately identify vehicle weight. Hence, an improved approach is proposed to address the above issues, which employs a Bidirectional Long Short-Term Memory (BiLSTM) network model to establish the mapping relationship between vertical deflection of bridges and vehicle loads. Firstly, based on practical engineering, the Grey Wolf Optimization-Variational Mode Decomposition (GWO-VMD) method is employed to address the influence of temperature effects on deflection response data. Then, through a time matching algorithm, the sliced deflection data and the corresponding BWIM system vehicle load information are paired to form the dataset for the deep learning models. The results of the models demonstrate that the optimal BiLSTM model exhibits better robustness and generalization performance. It achieves high accuracy of 97.9% in load identification. The method significantly reduces the economic cost and improves the accuracy of vehicle load identification.