Decoupling of vehicle-induced structural responses of highway bridges based on bipartite matching U- Net

Abstract

Structural anomaly diagnosis based on data pattern exploration of vehicle-induced responses under specific vehicle loads can provide a basis for real-time structural safety assessment. However, in the operation of highway bridges, monitoring vehicle responses are typically caused by multiple unknown vehicle loads occurring simultaneously, making it difficult to obtain independent vehicle-induced responses. This creates obstacles to specific vehicle response identification and analysis. For effective condition assessment of bridges, a one-dimensional U-Net with a bipartite matching loss function is presented to decouple multiple vehicle-induced responses into single vehicle-induced responses. The U-Net structure is designed to capture spatiotemporal features and reconstruct single vehicle-induced responses using one-dimensional convolutional layers with symmetric skip connections. The channel attention mechanism and bipartite matching loss function based on the Hungarian algorithm are employed to get the most reasonable match between the decoupled and real single vehicle-induced responses. The established network learns adaptive mapping from randomly synthesised multiple vehicle-induced responses to single vehicle-induced responses in training stage, realising effective decoupling based on a small number of sample annotations. A case study on the vehicle-induced cable forces of a cable-stayed bridge demonstrates the effectiveness of the proposed method, providing a basis for high-precision structural health diagnosis based on vehicle-induced responses.