A self-supervised physics-informed residual temporal convolutional network with progressive sparse reconstruction for moving load identification

Existing model-driven moving load identification (MLI) methods are highly sensitive to the selection of regularization parameters, whereas purely data-driven approaches generally require extensive labeled datasets and often suffer from limited generalization under complex vehicle-bridge interaction conditions. To address these challenges, a self-supervised physics-informed residual temporal convolutional network (PI-ResTCN) framework is proposed for accurate and robust MLI. Firstly, an orthogonal matching pursuit (OMP)-based sparse inversion scheme is employed to obtain an initial load estimate from measured bridge responses. By integrating temporal gradients, statistical descriptors, and temporal encoding information, the identified initial loads are transformed into multi-scale feature representations that provide physical priors for network learning. Subsequently, a residual temporal convolutional network is developed to hierarchically extract temporal features and capture both local dynamic variations and long-range temporal dependencies of moving loads. To eliminate the reliance on labeled load data, a self-supervised progressive sparse reconstruction strategy is further proposed, in which a logarithmic regularization is used to iteratively compensate unexplained response residuals and update the target loads. Finally, the effectiveness of the proposed method is validated through comprehensive numerical simulations and vehicle-bridge model experiments. Results show that the proposed PI-ResTCN consistently outperforms the OMP method in terms of identification accuracy while exhibiting lower sensitivity to sparsity parameters. Moreover, the progressive sparse reconstruction strategy effectively recovers weak dynamic load components and further improves load identification performance.