Identification Method for Vehicle Loads on Earth Dams Based on MPA-CNN Network

Accurate identification of moving vehicle loads on earth dam crests is critical to ensuring dam structural safety.To address weak anti-interference performance of traditional dynamic inversion methods and insufficient feature extraction of machine learning models,a vehicle load identification approach (MPA-CNN) is proposed by fusing marine predators algorithm (MPA) and convolutional neural network (CNN).MPA optimizes key hyperparameters of CNN adaptively to enhance time-series load feature extraction.Improved variational mode decomposition-singular value decomposition (IVMD-SVD) joint filtering is applied to denoise field acceleration signals and suppress environmental and sensor interference.A finite element model of an earth dam is built to simulate 16 cases with four vehicle speeds and four weights for training sample generation.Field tests on a homogeneous earth dam in Guangdong Province with 40 vehicles validate the method.Numerical results show that MPA-CNN achieves a minimum RMSE of 9.72N and a minimum MAPE of 0.027%,with accuracy improved by over 30% compared with conventional CNN.In field tests,RMSE ranges from 20.56N to 37.98N and MAPE is below 0.16%.The proposed method features high accuracy,strong robustness and good engineering applicability,supporting load monitoring and safety early warning for earth dam and similar geotechnical structures.